Aaron Makelky

Opus 5.5, Luna, and Muse: picking an AI daily driver

Sat, 10 Oct 2026 12:00:00 GMT

Aaron's take on the new models

Aaron Makelky thinks the real fight between Anthropic and OpenAI is the most useful daily driver model, not the top benchmark. Fable and GPT-6 Astra use up rate limits and cost a lot on the API, so most work needs a faster, cheaper model.

His daily driver is GPT-6 Luna, which he runs for background and parallel tasks while a stronger orchestrator model plans. Ryan Doser's daily driver is Claude Opus 5.5, which he says raised quality over Opus 5 and uses tokens more efficiently.

On Ryan's landing page test, Aaron preferred some details in Muse Spark's page over Opus 5.5 and said Opus 5.5 beat Fable 5.1's page from the week before. On the launch email, Ryan picked Opus 5.5, with GPT-6 Sol a close second.

Muse, the ChatGPT desktop app, and skills

Aaron uses Meta's Muse every day on its free tier. It turned a photo of a birthday invite into a calendar event and called a doctor's office for him, then patched him through to the receptionist. You can't choose or see the model it runs on.

He uses the ChatGPT desktop app for its in-app browser with cookies and extensions and for computer use, and Herdr to run terminal sessions across machines.

With custom GPTs retiring on December 11, 2026, Aaron and Ryan both use portable skills. Aaron's kids sports upload skill uploads, labels, and publishes his kids' game film as unlisted YouTube videos.

Questions

What does Aaron mean by a daily driver AI model?

The model he uses for most day-to-day work. Top models like Fable and GPT-6 Astra use up rate limits fast and get expensive on the API, so for most tasks Aaron wants a faster, cheaper model that's still good enough.

Which model does Aaron use as his daily driver?

GPT-6 Luna. He ran 5.6 Luna on the max or ultra setting more than any other model, and he uses Luna 6 for background and parallel jobs he doesn't need fast, while a stronger model writes the plan. Ryan Doser's daily driver is Claude Opus 5.5.

Was Opus 5.5 better than Muse Spark on Ryan's landing page test?

Not on everything. Aaron preferred Muse's glow effect on the checkout button and its single row of full-color client logos, but wouldn't call it better overall. He said Opus 5.5 beat what Fable 5.1 did the week before. On the launch email, Ryan picked Opus, with GPT-6 Sol a close second.

What is Meta's Muse good for?

Everyday digital tasks for people who don't want to touch GitHub or Cursor. Aaron used it to turn a photo of a birthday invite into a calendar event and gift reminder, to analyze his Instagram reels, and to call a doctor's office, wait through the phone tree, and patch him through to the receptionist.

What should I do about custom GPTs going away?

Aaron replaced a custom GPT use case with a skill, a markdown file with steps, which uploads and labels his kids' game film. If you try OpenAI's plugins instead, he says to keep the work in a markdown file or a GitHub repo you can take with you.

I use open source AI models for the work I hate

Sat, 10 Oct 2026 12:00:00 GMT

Where Aaron uses open source AI models

Aaron Makelky runs a 3 GB Qwen image model on his Mac to rename his daily screenshots offline, and Gemma models in Google AI Edge Gallery on his iPhone for questions about medical records, banking, and personal topics that he doesn't want leaving his device.

In the cloud, he uses the Hermes desktop app from Nous Research as the interface for open source models. His default is GLM 5.3 Flash on a Z.ai coding plan, and the Hermes model picker also holds his ChatGPT models, MiniMax, and the Nous Portal list, with side-by-side answers from several models on one prompt.

He uses OpenRouter so one API key covers every provider, and to try a new model on a small project for a dollar or fifty cents.

What he recommends

Give open source models the day-to-day background tasks: searches, finding files, drafts, calendars, and logging. Aaron has defaulted to GLM Flash models for at least nine months and points to DeepSeek for price.

Hosting the biggest open source models at home takes eight to ten thousand dollars of compute, so a hundred-dollar subscription is cheaper for most people. Free code-named models train on your data, so keep them for side projects.

Ryan Doser had Kimi K3, GLM 5.3 Flash, and GLM 5.3 redesign his opt-in page through OpenRouter. Aaron said the best one looked like Opus 4.8, and that the best Kimi and GLM models are closer to the frontier than a year ago.

Questions

Can you run an open source AI model on your phone without internet?

Yes. Aaron uses Google AI Edge Gallery, a free app in the iPhone App Store that's also on Android. It downloads Gemma models of about two and a half to five gigabytes, shows which ones fit on your phone, and runs them with Wi-Fi off, including image input. He uses it on airplanes and for medical, banking, and personal questions he doesn't want leaving his device.

Is it cheaper to host a big open source model yourself?

Not for most people. Aaron says running a model like Kimi or GLM 5.3 at home takes eight to ten thousand dollars of compute, while a hundred-dollar Claude or OpenAI subscription gets about the same usage most people need. Small local models, like the 3 GB Qwen model he uses to rename screenshots, are where self-hosting pays off.

Why use OpenRouter instead of separate API keys?

Aaron uses one OpenRouter key instead of storing and swapping a key for every provider in his apps and environment files. It also lets him try a new model on a five-minute project for a dollar or fifty cents instead of paying 20 to 100 dollars a month for another plan.

Where can you use open source models if Claude Code and the Codex app only run their own?

Aaron uses the Hermes desktop app from Nous Research. Its model picker holds his ChatGPT subscription models, his Z.ai GLM plan, MiniMax, and the Nous Portal's list of models, and it can send one prompt to DeepSeek, Kimi, and OpenAI at once to compare the answers side by side.

Which open source models does Aaron use most?

GLM Flash models from Z.ai, for at least nine months, for background tasks like searches, finding files, drafts, his family calendar, and logging workouts from a photo. GLM 5.3 Flash is the first one with image input. For price, he points to DeepSeek. He keeps free code-named models for side projects, because they train on everything you send.

I edit most of my videos with Claude in a terminal now

Sat, 10 Oct 2026 12:00:00 GMT

How Aaron edits video with AI

Aaron Makelky records multi-person episodes in Riverside or Descript for separate tracks and high-quality audio, then does most of his editing by pointing Claude Code or Codex at the raw file in a terminal.

His most used tool is HyperFrames, a free HTML-based video tool from HeyGen. He also uses Tesseract, and DaVinci Resolve with scripts Claude writes. None of them record, and some require a Mac and a lot of RAM.

For AI Rabbit Holes, Aaron connects Claude Code to the Riverside MCP and has a clipping skill find ten clips from the transcript. His caption rules keep text below his chin, never over his face, with tool logos and the main speaker on top in two-person clips.

What he recommends

Save the editing instructions as a skill, add one new thing each episode, and make it one to five percent better every time.

Use generated media like B-roll on top of real footage instead of generating the whole video, and turn finished edits into how-to posts.

Ryan Doser runs the podcast side with Claude Code skills for cuts, a Seedance-powered intro, packaging, and YouTube-to-blog repurposing.

Questions

Can Claude edit a video without video editing software?

Yes. Aaron points Claude Code or Codex at a raw video saved on his computer and has it edit with HyperFrames, which is free from HeyGen, or Tesseract. Claude writes the code for captions, graphics, explainer animations, and cuts in a terminal, so there's no editing app or subscription for it to connect to.

Do I still need Riverside or Descript if Claude does the editing?

For recording, yes. HyperFrames and Tesseract don't record. Aaron records with Riverside or Descript to get separate tracks and high-quality audio, especially with more than one person, then does most of the editing with Claude.

What computer do I need for AI video editing?

Something with local access to the file and powerful hardware. Some of these tools require a Mac and don't run on Linux, so a cloud server may not work, and editing uses a lot of RAM. Aaron says a MacBook Air or a Chromebook isn't enough.

How do I get better results from AI video edits?

Aaron saves his editing instructions as a skill, including a brand guide for captions, logos, and two-person layouts. He uses it on every episode, tries one new thing each time, and records what to keep doing and what never to do again.

Should I generate the whole video with AI?

Aaron doesn't. He uses generated media like B-roll that demonstrates what he's talking about, while the main footage is still him on camera. He expects platforms to keep limiting how fully generated videos are monetized and labeled.

I made a nature documentary about my gym fridge

Fri, 09 Oct 2026 12:00:00 GMT

How the video was made

Aaron filmed 23 silent clips of stocking his home-gym mini fridge with Liquid Death on an Insta360 Luna Ultra: mostly 4K at 30 fps, with the walk-in at 1080p 240 fps and the pink bottle at 4K 120 fps.

Claude Code on Claude Opus 5.5 in the Claude desktop app copied the footage, matched it to a 14-beat Notion shot list, cut it with FFmpeg and Python, and wrote a deadpan nature-documentary script that Aaron approved before any audio was generated.

HeyGen text to speech voiced 14 lines with a voice Aaron designed in HeyGen, for $0.23 total. The word timestamps from HeyGen drove the captions.

What the final edit does

It ships as a 16:9 master and two verticals: the full frame over a blurred copy of itself with captions in the blurred band, and a native 9:16 crop.

The captions use Gill Sans with the active word inverted, and every caption state sits inside a 1080 by 1920 safe box of x 60 to 940 and y 250 to 1500.

The video opens on a real frame instead of black, loops with a last-to-first frame similarity of 0.98, keeps music under the narration without ever dropping to silence, and ends on a push-in to the Liquid Death sticker.

Questions

What camera shot the Liquid Death fridge video?

An Insta360 Luna Ultra. 23 clips, mostly 4K at 30 fps, with the walk-in at 1080p 240 fps and the pink bottle at 4K 120 fps.

How was the narration made?

HeyGen text to speech with a voice Aaron designed in HeyGen, generated one line at a time through the HeyGen CLI after he approved the script. All 14 lines cost $0.23.

Is it a David Attenborough voice?

No. The ticket ruled out any Attenborough clone or soundalike. It's a generic deadpan narrator.

How do you get a 16:9 video into a 9:16 vertical without cropping?

Place the full 16:9 frame in the middle of the vertical frame over a blurred, darkened copy of itself. The blurred band above it is room for captions.

How does the video loop seamlessly?

The video starts 1.5 seconds into the opening shot and ends with that shot's first 1.5 seconds, so the last frame runs into frame 1. The music does the same trick with the track's first 1.5 seconds.

Was this sponsored?

It's part of Aaron's Liquid Death partnership. Liquid Death sent him free Liquid Death for a year.

I one-shot a CarPlay video edit with Claude and Tesseract

Sun, 04 Oct 2026 12:00:00 GMT

How the video was made

Aaron recorded a 2 minute 52 second 4K take on an Insta360 Luna Ultra with an Insta360 Mic Pro, calling ChatGPT dot and Grok Bot over Apple CarPlay.

He sent Claude Code, running Claude Opus 5.5 at medium effort, one prompt asking for a clear, on-brand edit that never covers his face or a screen he is pointing at, tracked in Linear.

A cloud session wrote the brief into Linear ticket AAR-270. A local Claude Code session on his Mac Studio picked up the ticket, edited the video in Tesseract with his video-edit-tesseract skill, and returned a 2:11 edit that Aaron posted with no revision notes.

What made it one-shottable

A brand system written as checkable rules: colors with jobs, two font files, motion limits, and places graphics can never go.

A saved editing skill, a ticket that defined deliverables before work started, and footage shot with clean audio and room for captions.

The deliverables included an editable Tesseract project, a review MP4, SRT and VTT captions, an edit decision list with 30 source timecodes, a post caption, and verification checks including OCR for private text.

Questions

What does one-shot mean for this video edit?

Aaron sent one creative prompt and approved the first rendered version with no revision notes. Two later messages only told Claude where the source file lived and handed the Linear ticket to a local session.

What tools were used to edit the video?

Claude Code running Claude Opus 5.5 at medium effort, Tesseract with Aaron's video-edit-tesseract skill, Linear for tracking, Claude Remote Control to follow the local session, Whisper and forced alignment for word timing, and FFmpeg for checks.

What camera and microphone recorded the source?

An Insta360 Luna Ultra recording 4K at 24 fps in a wide 16:9 frame, with audio from an Insta360 Mic Pro, the version with the e-ink display.

Do I need Aaron's brand system to get the same result?

No. You need your own written brand rules: a palette with a job for each color, two font files, motion limits, and places graphics are never allowed to go.

Why did the edit include SRT and VTT caption files?

The Linear ticket listed them as deliverables before any editing started.

How to make your own ChatGPT dots in Grok Bot, Claude, or Muse

Wed, 30 Sep 2026 12:00:00 GMT

Article

ChatGPT's new dots are always-on agents that keep working after you close the chat. Getting one takes a ChatGPT plan that starts at $100 a month. I set mine up in the ChatGPT desktop app this morning, and it spawned a background agent called "Heartbeat Dreamer." Its full prompt is at the bottom of this page.

You can build your own version of that system without the $100 plan. This guide covers three agents: Grok Bot , Claude , and Muse . The parts that work the same everywhere come first, then each agent gets its own setup steps, prompts, and limits. If you already run a chief of staff bot and project leads in Grok Bot, this fits that org chart: a green specialist that researches on a schedule and hands findings to your chief.

What it does

OpenAI calls this proactive research. When you aren't talking to your dot, it looks for ways to help in the background, and OpenAI says those background tools are read-only: no sending messages, no editing your apps, no driving your browser. ( OpenAI )

The prompt shows how that's wired up. The dot you chat with is the "parent." The Heartbeat Dreamer is a child agent that only researches, and it reports to the parent. It's allowed to pass along private or sensitive stuff, and the parent decides what reaches you.

They talk through files. The parent keeps /user_notes/ and /action_items.md. The dreamer keeps its own checkpoints in /agent_notes/. It never edits the parent's notes. When something should change, it tells the parent to change it.

Every run starts by reading the last checkpoint, the parent's notes, roughly the last two hours of your chat with the parent, and a shared notes page if you have one. Then it makes a short plan and goes through your connected apps, email first. It ends by writing a new checkpoint that says where the next run should pick up, and it deletes its own checkpoints once they're more than 14 days old.

The last step is one message to the parent, with every finding tagged P00, P0, P1 and so on. If nothing is worth sending, it stays quiet and writes down why.

Once a day, around 3am your time, it does a reflection run instead. That run rereads the day's conversation, picks up preferences and any offers you turned down, and reshuffles your priorities.

The rules to copy

These are the lines you can drop straight into your agents:

How to build it in Grok Bot

The Grok Bot version is a routine on a specialist Bot that reads your connectors and browser sessions on a timer, writes checkpoints to a shared file on the cloud computer, and messages your chief Bot with tagged findings. Routines run in the cloud while your laptop is closed. ( xAI: Skills and routines ) Grok Bot access requires a paid individual Cursor plan, a Cursor Teams seat, or a linked SuperGrok subscription. ( xAI: Get started )

Here's how you set it up.

Heartbeat prompt (swap in your chief Bot's @ name, connectors, and time zone):

Create a routine called "Heartbeat" on this Bot that runs on weekdays at 7:00 AM, 10:00 AM, 1:00 PM, and 4:00 PM in America/Denver. Each run should: 1. Read /workspace/heartbeat/notes.md: the latest entry under Checkpoints and anything under Inbox I have not acted on. 2. Research my connected sources, email first, then calendar, then any connectors I have enabled (use @ in this chat to reach them). Use browser sessions on the shared computer where needed. Look at what changed since the last checkpoint: new threads, replies, invites, deadlines, things I said I would do, and things I was asked to do. 3. Before calling anything unanswered, overdue, or urgent, check sent mail, later replies, and task status. A search date range is not proof you reviewed that whole range. 4. Stay read-only. Do not send, reply, accept, edit, or delete anything in any app without my approval. The only file you may write is /workspace/heartbeat/notes.md. 5. Add a checkpoint under Checkpoints with the run time, what you checked, what you could not reach, and where the next run should pick up. 6. Add findings under Inbox. Tag each one P0, P1, or P2, and include what you learned, how sure you are, the source, and one concrete next step. If nothing is worth adding, write "Quiet" and one line on why. 7. Send an async message to @Chief with the same tagged findings. If nothing needs the chief, stay quiet in chat and record why in the checkpoint. 8. Treat a one-time choice as a one-time choice. If the same blocker shows up across several runs, ask me once whether to pause this routine. Reflection prompt:

Create a routine called "Heartbeat reflection" on this Bot that runs every day at 2:45 AM America/Denver. Each run should: 1. Read /workspace/heartbeat/notes.md and, if you can reach them, my recent conversations with my chief Bot from the past day. 2. Add a checkpoint under Checkpoints titled Reflection: preferences I stated, offers I turned down, and whether the open Inbox items are ranked right. Keep one-off comments as one-offs. 3. Stay read-only everywhere except /workspace/heartbeat/notes.md. 4. Message @Chief only if reflection changes how it should prioritize my Inbox. Where the Grok Bot version falls short Dot's read-only rule is enforced in code. In Grok Bot it is a boundary in the Bot description, the routine instructions, and optional Auto-review rules. Connectors and browser tools can still attempt writes unless you stop them. ( xAI: Approvals )

There is no built-in /agent_notes/ path. You use a file under /workspace and Bot messages instead. That file is visible to every Bot on your account because they share one cloud computer. ( xAI: Shared computer )

The dreamer reads the last two hours of parent chat every run in OpenAI's prompt. Grok Bot docs describe Bot memory, search, and handoffs, but not an automatic two-hour parent transcript on every routine run. Your chief Bot has to pull context from conversation history, your messages, or the notes file.

Routines draw on your Grok Bot usage, and xAI says it may pause routines to control unattended usage. Heavy schedules plus connector research can burn your weekly limit fast. ( xAI: Skills and routines , how I stretch Grok Bot usage )

xAI's Grok Bot docs do not publish a full Marketplace plugin list or confirm every connector Claude lists (GitHub, Linear, Notion, and so on). Install what you actually use and record gaps in your Bot description rather than assuming parity.

How to build it in Claude

The Claude version is a scheduled task that reads your apps on a timer and writes to one notes doc, plus a second task at night for the reflection. Scheduled tasks are on Claude's paid plans (Pro, Max, Team, Enterprise) and run in the cloud, so they keep going when your laptop is asleep. ( Claude Help Center )

Here's how you set it up.

Heartbeat prompt (swap in your own times and time zone):

Create a scheduled task called "Heartbeat" that runs on weekdays at 7am, 10am, 1pm, and 4pm Mountain time. Each run should: 1. Read my Google Doc "Heartbeat notes": the latest entry under Checkpoints, and anything under Inbox I haven't acted on. 2. Research my connected apps, email first, then calendar, then everything else. Look at what changed since the last checkpoint: new threads, replies, invites, deadlines, things I said I'd do, and things I was asked to do. 3. Before calling anything unanswered, overdue, or urgent, check my sent mail, later replies, and task status. A search's date range is not proof you reviewed that whole range. 4. Stay read-only. Do not send, reply, accept, edit, or delete anything in any app. The only thing you may write to is the Heartbeat notes doc. 5. Add a checkpoint under Checkpoints: the run time, what you checked, what you couldn't reach, and where the next run should pick up. 6. Add findings under Inbox. Tag each one P0, P1, or P2, and include what you learned, how sure you are, the source, and one concrete next step. If nothing is worth adding, write "Quiet" and one line on why. 7. Treat a one-time choice as a one-time choice. If the same blocker shows up across several runs, ask me once whether to pause this task. Reflection prompt:

Create a scheduled task called "Heartbeat reflection" that runs every day at 2:45am Mountain time. Each run should: 1. Read my Google Doc "Heartbeat notes" and, if you can reach them, my Claude chats from the past day. 2. Add an entry under Checkpoints titled Reflection: preferences I stated, offers I turned down, and whether the open Inbox items are ranked right. Keep one-off comments as one-offs. 3. Stay read-only everywhere except the Heartbeat notes doc. Where the Claude version falls short Dot's read-only rule is enforced in code. In this setup it's a sentence in a prompt. If you want a harder stop, set the task up manually from Scheduled and pick an approval mode that makes it ask before it acts.

There's no parent agent. Dot's dreamer hands findings to the agent you already talk to. Here, findings sit in a doc until you, or your main chat, go read them.

It doesn't read your conversations every run. The dreamer checks the last two hours of your chat with the parent each time. The notes doc in step 3 is how you cover that.

How to build it in Muse

The Muse version is a recurring scheduled task that reads your Connectors on a timer and writes to one notes file in your Library, plus a second task at night for reflection. Meta says recurring tasks repeat on the schedule you set and are delivered as messages in your Muse conversation. ( Meta Help Center: reminders and scheduled tasks )

Here's how you set it up.

Heartbeat prompt (swap in your own times and time zone):

Create a recurring scheduled task called "Heartbeat" that runs on weekdays at 7am, 10am, 1pm, and 4pm Mountain time. Each run should: 1. Read my Library file "Heartbeat notes": the latest entry under Checkpoints, and anything under Inbox I have not acted on. 2. Research my connected apps through Connectors, email first, then calendar, then anything else I have connected. Look at what changed since the last checkpoint: new threads, replies, invites, deadlines, things I said I would do, and things I was asked to do. 3. Before calling anything unanswered, overdue, or urgent, check my sent mail, later replies, and task status. A search's date range is not proof you reviewed that whole range. 4. Stay read-only on Connectors. Do not send, reply, accept, edit, or delete anything in any connected app. The only thing you may write to is the Heartbeat notes file in my Library. 5. Add a checkpoint under Checkpoints: the run time, what you checked, what you could not reach, and where the next run should pick up. 6. Add findings under Inbox. Tag each one P0, P1, or P2, and include what you learned, how sure you are, the source, and one concrete next step. If nothing is worth adding, write "Quiet" and one line on why. 7. Treat a one-time choice as a one-time choice. If the same blocker shows up across several runs, ask me once whether to pause this task. 8. Deliver a short summary in this chat when the run finishes. Reflection prompt:

Create a recurring scheduled task called "Heartbeat reflection" that runs every day at 2:45am Mountain time. Each run should: 1. Read my Library file "Heartbeat notes" and, if you can reach them, my Muse chats from the past day. 2. Add an entry under Checkpoints titled Reflection: preferences I stated, offers I turned down, and whether the open Inbox items are ranked right. Keep one-off comments as one-offs. 3. Stay read-only on Connectors. The only file you may edit is Heartbeat notes in my Library. 4. Deliver a short summary in this chat when the run finishes. Where the Muse version falls short Dot's read-only rule is enforced in code. In Muse it is a mix of prompt instructions, Connector retrieve-only settings, and approval prompts. Meta says many approval and security checks run outside the AI model, but this is not the same as a dedicated background researcher with hard-coded read-only tools. ( Meta Help Center: privacy and security )

There is no parent agent. Dot's dreamer hands findings to the agent you already talk to. Findings land in a Library file and in scheduled-task messages until you open your main chat and ask for them.

Meta's help articles do not say each scheduled run automatically reads the last two hours of your main chat. The Heartbeat notes file is the shared handoff.

Meta's Connectors article doesn't name project tools like Linear or Notion. Services outside the in-app list require a Custom Connector, which Meta does not review. ( Meta Help Center: Connectors )

Recurring tasks can repeat on a custom interval, but Meta does not document a maximum cadence or how many tokens a research-heavy schedule burns on the free tier. If heartbeats hit your usage limit, you wait for the refresh or upgrade. Power is $20 a month for a higher weekly allowance. ( Meta Help Center: Muse subscriptions )

The full prompt

Captured from my own Dot setup on September 30, 2026. This is OpenAI's text, not mine, and it will probably change. Raw text file

I asked an AI agent to edit my video. Here are all three versions.

Mon, 21 Sep 2026 12:00:00 GMT

One take, three edits

The 57.375-second original became a 54.75-second comparison. Version 2 improved caption placement, animated graphics, the coffee cue, and the ending. Version 3 refined 42 caption phrases and varied nine sound cues.

The next edit starts with saved instructions

The custom video-edit-tesseract skill records the approved video style. Subtle camera movement that preserves eye height was added to the skill after version 3 and has not been rendered into that video.

Meow vs Mercury: why my business banking went to Mercury

Fri, 18 Sep 2026 12:00:00 GMT

Article

I needed a real business bank for my consulting practice. A fractional CFO friend recommended Meow, then admitted he banks at Mercury. Wells Fargo was the incumbent I was trying to escape. So I compared Meow and Mercury properly: the apps, the APIs, the pricing, and what my agents can actually do with each one.

Mercury won as my daily operating account. Meow won as the better idea for how agents should touch money.

Mercury is the better daily driver.

For a solo consultant prioritizing mobile banking, invoicing, webhooks, and automation , Mercury is the more complete operating setup.

Best overall fit · with a $29.90/mo caveat CHANGE THE DECISION LENS Run the business Let agents take action Avoid monthly fees MERCURY Native iOS and Android apps, mature receivables, approval alerts, and broad banking APIs make it the stronger everyday operating account. MEOW Its official MCP can create products, customers, and invoices, manage virtual cards, and draft transfers for human approval. Mercury’s official MCP is read-only. MEOW Meow has no invoicing-API paywall. Mercury requires Plus at $29.90 per month for Invoices and Customers API access. THE SHORT VERSION Choose the workflow, not just the API. Mercury wins the human layer. Its native mobile apps handle balances, transfers, cards, mobile check deposit, receipts, and pending-approval alerts. Meow wins the agent layer: its official MCP can create invoices, issue or freeze virtual cards, and draft transfers for human approval.

Pick Mercury if the bank account needs to work well from your phone.

Pick Meow if coding-agent write access matters more than a mobile app.

Mobile app

Native iOS + Android — Meow has no native app.

Webhooks

Signed delivery, retries, and event filtering on both sides.

Invoicing API

Both create and track invoices; Mercury requires Plus.

Coding agents

Official MCP writes invoices and drafts transfers.

What changes in practice

The two APIs overlap heavily. The decisive differences are mobile access, MCP write capability, and the friction around invoicing and write tokens.

Meow explicitly positions banking as shifting away from apps toward AI agents.

Balances, transactions, transfers, mobile check deposit, cards, receipt scanning, widgets, and push alerts for pending approvals.

Products, customers, invoices, line items, PDF downloads, recipient email, multiple payment methods, and lifecycle status.

Create, send, update, cancel, track, and download PDFs. Requires Mercury Plus at $29.90/month or higher.

Create products, customers, and invoices; issue or manage virtual cards; draft transfers that stay pending until a person approves them.

Official tools list and inspect accounts, transactions, cards, invoices, customers, and more—but do not create invoices or send money.

Full read/write API with about 30 fine-grained scopes. No IP-whitelist requirement.

Read & Write tokens require IP whitelisting. Custom tokens can expose selected endpoints, but write scopes generally still need a whitelist.

HMAC-SHA-256, snapshot or thin payloads, test events, delivery history, redrive, secret rotation, and circuit breaking.

HMAC-SHA-256, event-type and path filters, test verification, and at-least-once delivery. Not available in sandbox.

Simulate inbound transfers, stablecoin funding, card authorization and settlement events, and application approval.

Preloaded data and sandbox-only tokens. Use the Events API because webhooks do not run in the sandbox.

Drill into the details

Open any section for the parts most likely to affect an integration plan.

Accounts, payments, cards & operations Meow Accounts, balances, transactions, limits, statements, tax documents, contacts, users, reimbursements, and routing-number validation. ACH, scheduled ACH, domestic wire, internal book, international, crypto, and USDC transfer creation. Create, update, and revoke cards; bill-pay CRUD and document upload flows. Business onboarding, representatives, KYC, due diligence, and document uploads are API-accessible. Mercury Accounts, transactions, statements, recipients, merchants, categories, credit, treasury, users, organization, and SAFEs. Send money to recipients, internal transfers, checks, and domestic wires; manage send-money approval requests. Create, update, cancel, freeze, and unfreeze cards; Vault can reveal card details. Attachments, transaction metadata, statement PDFs, and OAuth2 integration flows. Authentication & token management Meow API key in the x-api-key header, created and managed in the dashboard. Fine-grained scopes include invoices, products, customers, transfers, cards, webhooks, bill pay, and simulations. Multi-entity keys add an x-entity-id header. OAuth is reserved for the MCP surface. Mercury Bearer or basic authentication with Read Only, Read and Write, or Custom token tiers. Write access generally requires IP whitelisting; scopes cannot be edited after creation. Unused tokens are auto-deleted after 45 days; unused permissions can also be auto-downgraded after 45 days, with seven days’ email notice. Invoice lifecycle & payment methods Meow Status covers draft, open, pending, scheduled, partially paid, canceled, expired, and overdue. Tracks total, paid, and due amounts; supports invoice dates, due dates, notes, line items, recipient email, and PDF download. Documented payment methods include bank transfer, card, ACH debit, international wire, and several USDC/USDT networks. Mercury Create, list, get, update, cancel, attach files, download PDF, and send by email at creation. Customer payment methods include credit card, Apple Pay, Google Pay, wire, ACH, and ACH debit for subscribers. Plus adds recurring invoices and ACH-debit invoicing at $1 per transaction. IMPLEMENTATION PATH What each choice commits you to RECOMMENDED Mercury as the operating account Best if you want dependable phone access first, then layer automation around it.

Meow as the automation account

Best if you want a coding agent to do more through an official integration.

What I actually need from a business bank

I need five things: a mobile app I can run the business from, invoicing I can trigger from my own systems, an API my automations can use, transfers I can track, and a bank that doesn't make my wife drive to a branch to get added to an account.

The branch visit that ended the debate

Wells Fargo made us go in person to get my wife added to a business account. In 2026. Mercury let me add her online, is shipping her a physical card, and sent 25 checks for free.

Wells Fargo treated its process as sacred and my time as worthless. Mercury acted like software. I know which one I want holding my operating cash.

Where Mercury wins: the daily operating account

Mercury is the better bank for the unglamorous daily work. The mobile app is native and actually good, which matters because I run most of my business from my phone. Transfers have real tracking instead of the "it'll show up in 1-3 business days, good luck" experience. Webhooks mean my systems know when money moves without me polling an API like it's 2014.

The invoicing API is the sleeper feature. I can generate and send invoices from my own workflows instead of living inside someone else's dashboard. It requires a Plus plan, which is worth knowing before you build on it.

Mercury's API story has one real wart for my use case: general read and write tokens require IP whitelisting. That's fine for a server with a static IP and annoying for agents running on dynamic infrastructure. Their official MCP was documented as beta and read-only, so don't plan on your agent doing much inside Mercury yet.

Where Meow wins: agent-native money movement

Meow is built around a genuinely better idea for how agents should touch money: the agent prepares transfers and a human approves them. That's the workflow I actually want. I don't want my agent moving money autonomously, and I don't want to be the API either. Human-approved drafts are the setup I want for agent-run finance: the agent drafts, I approve.

The pricing is unusual: no public paid subscription for the core banking and agent offering. The paid parts are gated by balance instead of subscription. Yield and brokerage products need a $100,000 checking balance, and the treasury product runs 1 basis point per month, annualized.

The weak spot is mobile: Meow has no native mobile app, and I run my business from my phone. That is the gap that keeps it from being my daily driver.

The pricing

Mercury is straightforward: a free tier, Plus at $29.90 a month, Pro at $299 a month. The invoicing API needs Plus.

Meow has no public paid subscription for the core offering. The costs show up as balance gates and treasury fees instead. Different model, not necessarily a cheaper one, depending on how much cash you keep parked.

My verdict

Mercury is my operating account. It does the daily work better: mobile, invoices, webhooks, tracking, and a bank that treats my time like it matters. Meow is the one I'm watching. The human-approved transfer draft is how agent-run finance should work. If they ship a real mobile app and keep the agent-approval workflow ahead of Mercury's, I'll take another look.

If you're picking today and you mostly need a bank that works, pick Mercury. If you're building agent workflows that touch money and you want a human in the approval loop by design, Meow deserves a real look.

Disclosure: the Mercury link below is my referral link. It costs you nothing and supports the site. If you'd rather not use it, mercury.com works the same.

Open a Mercury account with my referral link

Should I pick Meow or Mercury for my business banking?

I picked Mercury as my daily operating account because of its mobile app, invoicing API, webhooks, and transfer tracking. Meow is the better pick if you want agent-native workflows with human-approved transfer drafts. If you mostly need a bank that just works day to day, I'd start with Mercury.

How much does Mercury cost?

Mercury has a free tier, Plus at $29.90 a month, Pro at $299 a month. The invoicing API requires a Plus plan or higher.

Does Meow charge a subscription?

No. Meow's core banking and agent offering has no public paid subscription. The paid parts are gated by balance instead: yield and brokerage products require a $100,000 checking balance, and the treasury product charges 1 basis point per month, annualized.

Can my AI agents move money with Mercury or Meow?

Meow is built for this: agents can prepare transfers that a human approves. Mercury's official MCP was documented as beta and read-only, and its general read/write API tokens require IP whitelisting, which is friction for agents running on dynamic infrastructure. For agent-driven money movement with a human in the loop, Meow has the better model right now.

Did you really have to visit a branch to add your wife to Wells Fargo?

Yes. Wells Fargo required an in-person visit to add my wife to a business account. Mercury let me add her online, shipped her a physical card, and sent 25 checks for free. That single experience told me more about the two banks than any feature table.

Questions

Should I pick Meow or Mercury for my business banking?

How much does Mercury cost?

Mercury has a free tier, Plus at $29.90 a month, Pro at $299 a month. The invoicing API requires a Plus plan or higher.

Does Meow charge a subscription?

Can my AI agents move money with Mercury or Meow?

Did you really have to visit a branch to add your wife to Wells Fargo?

Astra vs Fable: why I'm still paying for ChatGPT

Tue, 15 Sep 2026 12:00:00 GMT

Article

I've used Fable 5.1 for long-form writing where I could immediately tell it did a better job than Astra, but that improvement hasn't been enough for me to pay for a Claude subscription.

I gave both models the same PDF, 17 scanned pages of my handwritten ideas, and Fable had the better hook and did a better job organizing the writing, but Astra did the entire job faster than Fable could even read the PDF.

I prefer Claude Fable 5.1's writing and still get more work done with GPT-6 Astra. How often do I need that extra bit of quality, and what am I giving up to get it?

I talked through that with Ryan Doser on AI Rabbit Holes on September 15, 2026. Ryan had been testing both models on websites, and I had been using them for writing and the other work I need to get done during the week.

Where Fable 5.1 earned the extra time

The writing task had a few places to go wrong before the model even got to the writing. It had to read a scanned PDF, pull the words out of my handwriting, and turn those ideas into a long-form piece without inventing a bunch of slop.

I gave Fable the PDF through Cursor and used the same prompt with Astra in ChatGPT, asking both to turn my handwritten ideas into a piece of writing that stayed true to what I'd actually said.

Fable took a long time and made a lot of tool calls just to read the PDF, but when I opened the result, I could see why I'd use it as a copywriting coach. I hadn't given it a hook, just a pile of thoughts I wanted help structuring, and it did a much better job of figuring out how to put them together.

Even though I'm biased toward Astra because I pay for ChatGPT, I'd still pick Fable's draft in a blind test.

Astra got through the PDF without getting stuck, and its writing had more personality than I'd come to expect from GPT. If I have one piece where I really need help with the hook and structure, I might reach for Fable, but if I'm doing ten of these in a day, the extra waiting becomes a much bigger part of the decision.

Ryan's landing page test: fancy doesn't sell

Ryan gave both models his existing landing page and asked for a better version designed to get more conversions. He hadn't specifically told either model to stay on brand.

Fable stayed close to the page he already had, keeping the dark background and blue accents while making the purchase buttons stand out more. The testimonials also looked a little better to me, but overall it was an iteration of the existing design.

Astra changed the colors and put Ryan's name at the top, then used a different widget to show what the product does. The page was shorter and more direct, which I actually liked even though it had gone off brand.

My gut reaction was that Astra's version would probably perform better because it was easier to understand the offer and know where to click. Ryan would have to test that, but I thought the simpler version was worth putting in front of customers.

With a landing page, I want someone to understand what they're buying and know where to click without having to work through a bunch of design. Ryan made the same point about a plumber in Des Moines: the business owner cares whether someone fills out the form and becomes a lead. A fancy page doesn't help much if it gets in the way of that.

Astra gave Ryan more of the business he asked for

The second test started with a local service business idea: flooded-basement cleanup in Des Moines, Iowa. Ryan asked for a company name, an ideal customer profile, a brand guide, and a homepage design. He supplied a reference website and a skill.

Ryan had Claude Code and Codex running inside VS Code, so he could give them the same prompt from the workspace he already uses. He didn't need to switch between two desktop apps to do the comparison.

Fable's result looked rudimentary to me, with orange text that was hard to read and five steps packed in beside the headline. Farther down, the colored labels in the insurance section had awkward spacing. If I'm paying for a Max plan, those are things I'd expect it to get right without having to go back and ask for fixes.

Astra gave him a brand concept with a separate guide and a homepage. The colors felt more considered to me, and the orange it chose for actions actually carried through to the buttons. It also used image generation to fill in the page.

What stood out to me was how much Astra filled in on its own, including using the image generator without Ryan having to ask for it separately. Ryan has built a site he likes with Claude too, but in this comparison, Astra gave him more of what he'd need to keep building the business.

I care what it costs to finish the job

I'm not paying for a direct Anthropic subscription. I access its models through tools such as Cursor and OpenRouter. At the time we recorded, I was still on my existing $200 ChatGPT Pro plan. Ryan was comparing his $100 ChatGPT Pro and $100 Claude Max subscriptions.

One correction to our conversation: standard API pricing is the same for these two models, $10 per million input tokens and $50 per million output tokens. That is what OpenAI lists for Astra and Anthropic lists for Fable 5.1 . App subscriptions and third-party credits are different billing systems.

The total cost of a task still depends on how much the model uses to finish it. A higher price per token doesn't automatically mean a more expensive finished page, and the same token price doesn't guarantee the same bill.

The subscription limits also matter to me. Claude's pricing page lists Fable usage at 50% of weekly limits on Max. That specifically limits how much of the weekly allowance can go to Fable. OpenAI says new $200 Pro sign-ups are temporarily paused , while existing subscriptions continue. These are the plan details I checked for September 15, 2026.

I also overstated it when I talked about having no limit on Astra. Astra has usage limits , and the app and plan matter. My preference is about the work I can get through with my subscription.

The best model isn't doing all my work

I use GPT-5.6 Luna as much as any model. If the task is moving information around or doing straightforward work, I don't need Fable or Astra for every step.

Ryan uses GPT-5.6 Sol to check work produced by other models. For a high-value task, he may use Astra for that final check. My addition was that the checking model needs the skills and context to know what we care about. Otherwise, we're explaining ourselves again every time.

That's the same reason I keep my instructions outside a single app, which I wrote about in my Grok Bot review . I want to carry the way I work with me when I switch tools.

A privacy correction worth making

I was too broad in the recording when I talked about Anthropic keeping data and turning training off in ChatGPT. Training and retention are separate questions.

ChatGPT lets consumer users opt out of training , and Anthropic gives consumer users a training choice . Their API and business products generally don't train on customer content by default. That doesn't mean nothing is retained. Anthropic has specific retention requirements for covered models such as Fable 5.1 . OpenAI also describes retention separately from training in its API data controls . The rules of the app or intermediary I'm using matter too.

I still care where my private thoughts and ideas go. I just shouldn't treat a training toggle as a promise that nothing is saved.

Is GPT-6 Astra or Claude Fable 5.1 better for writing?

I preferred Fable 5.1's hook and structure when I gave both models 17 scanned pages of my handwritten ideas. Astra still produced good writing and finished the whole job much faster, so I'd use Fable when that extra help with the writing is worth the wait.

Which model did better on the website tests?

In Ryan's tests, Fable stayed closer to his existing landing page's branding, while Astra made a shorter page that I thought was worth testing with customers. For the basement-cleanup business, Astra gave him a brand concept and guide along with the homepage and a generated image. I preferred that result, but we hadn't measured which page converted better.

Should I choose ChatGPT Pro or Claude Max?

If I'm recommending one subscription, I'd choose ChatGPT Pro. Fable's writing can be better, but I get more work done for my money with OpenAI's models and tools. That makes ChatGPT Pro the better value for how I work.

Do Astra and Fable have different API prices?

At the September 15 comparison, both listed standard API rates of $10 per million input tokens and $50 per million output tokens. The cost of finishing a task can still differ because the models may use different amounts of tokens. Those rates are separate from a ChatGPT Pro or Claude Max subscription.

Do I need Astra or Fable for every task?

I use GPT-5.6 Luna as much as any model because straightforward work doesn't always need the most expensive option. Ryan uses GPT-5.6 Sol to check other models' work and may use Astra for a higher-value task. I want the model doing that check to already have my skills and context.

OpenAI is doing a better job of marketing this

I also think OpenAI is doing a better job of winning over the average person with its marketing on Twitter, LinkedIn, and Instagram, along with things like the box of Astra hats they sent me. Most people aren't comparing models the way Ryan and I are; they know which commercial was cooler or which ad got them to click. Right now, it feels like OpenAI has the momentum there too.

Ultimately, OpenAI's subscriptions are a better value for the work I need to get done, even though Fable is better at some of it. If someone is only choosing one subscription, I'd tell them to choose ChatGPT Pro over Claude Max because they can get more done for their money.

Questions

Is GPT-6 Astra or Claude Fable 5.1 better for writing?

Which model did better on the website tests?

Should I choose ChatGPT Pro or Claude Max?

Do Astra and Fable have different API prices?

Do I need Astra or Fable for every task?

I paid $300 for Grok Bot. Its cloud computer is the most useful part.

Tue, 08 Sep 2026 12:00:00 GMT

Article

From my appearance on AI Rabbit Holes with Ryan Doser.

I paid full price for Grok Bot's top plan and use it every day. The most useful part of my setup is the virtual machine it runs on, which lets me operate a browser and run other agents in the cloud.

When Ryan Doser asked whether I would recommend buying the subscription just for Grok Bot, my answer was no. I'm still hitting usage limits on the highest plan.

I talked with Ryan on AI Rabbit Holes about Grok Bot, the Slack-style agent product from xAI. We covered setup, rate limits, who it's for, and the other tools included in the subscription.

I can organize bots in channels, but I can't invite a colleague

Grok Bot gives me a chat interface where I can put bots into groups, give them jobs, and have them trigger automations. The bots can talk to one another, but I can't bring a human colleague into the workspace to work alongside me.

I find the setup much simpler than OpenClaw. Grok Bot runs in the cloud, so I don't need to buy and maintain a separate computer for it. I can use the app on my Mac, iPhone, or Linux device, and turning off my laptop doesn't stop the work.

Connecting tools is straightforward too. When I connect Gmail, a sign-in box appears in the chat. I can name that connection and add a second email account through another sign-in prompt, without pasting API keys into a configuration file or opening a terminal.

For someone confused by the different chat, desktop, and coding apps, I can see the appeal of accessing their connected tools through one app.

I still hit the limits on the highest plan

The entry price keeps coming down, but the limits on the cheaper plans, including free access, concern me. I don't think access is especially useful if someone can run out of usage before finishing the setup they need.

When I got in, access required a top-tier plan: Cursor Ultra at $200 or SuperGrok Heavy at $300. I paid full price without a sponsor deal because I wanted to know whether someone could use this without being comfortable in a terminal.

The subscription includes Cursor, Grok, and Grok Bot, each with a separate usage allowance. I can hit the Grok Bot limit and be unable to use it for two days while still having access to the Grok app. Someone buying the subscription specifically for Grok Bot can easily miss that distinction.

I'm on the highest plan and still hitting weekly caps. Based on that experience, I don't recommend expecting a full week of heavy work across multiple projects. Moderate personal use seems more realistic to me.

I can't choose the model or its reasoning level

I want to choose which model handles a task and how much reasoning it uses, as I can in Codex or Claude Code. Grok Bot doesn't give me those controls. Its interface also doesn't tell me the reasoning level it's using.

I understand why someone else might prefer automatic model selection. If they don't follow model releases, they may want to make a request and let the app choose. I find Notion's model menu overwhelming enough that I can see why someone would leave that decision alone.

My frustration is that I can't assign easier work to a different model, even within the company's own models. I want to reserve the more capable model for tasks that need it and have more control over my usage.

That limitation makes Grok Bot a poor choice as my main AI tool. Someone who prefers automatic selection may care less about it.

My bots have specific responsibilities

I don't like calling a setup "best practices" when the product has existed for only a few weeks. I organize mine around the projects I manage and the bots I want to contact directly.

Every morning while I'm getting up or working out, my purple chief-of-staff bot runs a standup with the project bots in an all-hands group. The meeting is scheduled, so I don't have to send a message to start it.

I like being able to read the exchange in a group chat and mention a bot to ask a question. Other tools I use can inspect work across sessions, but I find this interface easier to follow.

Shape and color help me recognize responsibilities. The chief of staff is my only purple cloud, and the project managers are triangles in different colors. I keep the specialist bots hidden because I don't want to message each one myself. If I need a web page, I can ask the project manager to assign it to the web developer bot.

My bots can prepare drafts in Gmail, but I don't let them independently send emails or post messages as me. A connector can give a bot permission to post in Slack or send my boss an update. I want to review external communication before it goes out.

I review the podcast edit before publishing

I showed Ryan a podcast example using connectors for the tools where I edit and publish.

I asked a bot to add intro and outro music, move an ad read, generate show notes and timestamps, and get the episode ready to publish. It found the right tracks and project and handled much of the busy work. I still listened to the result and made changes before approving publication because the edit wasn't perfect.

Previously, I had to open the editor, make the changes, export an MP3, and put it into Transistor with the description. With the Descript connector, I can request the work from my phone while it runs in the cloud.

I keep the skills describing how I like to edit a podcast in GitHub, with a copy outside Grok Bot. If I cancel the subscription, I still have those instructions available for another tool instead of having to write them again.

I can use the cloud computer myself

The virtual machine gives me more ways to use the subscription. I can open its browser, search, log in, and take over when the agent gets stuck. I can also run Codex, Cursor, Antigravity, and Grok's command-line tool in its terminal.

The power went out at my house the night before we recorded. Work running on the remote computer can continue independently of my Mac.

Mine is connected to Tailscale, with its internet traffic routed through my home network, which reduces the bot checks I encounter. It prints my schedule on the home printer in the morning so I can drink coffee without opening a screen. I can also move files across the network instead of uploading them to Drive and downloading them again.

I use the other agents on that computer to help preserve my Grok Bot usage. My chief-of-staff instructions tell it to handle the bot conversations, then use the agents in the terminal when something needs building. Cursor agent usage and Grok Build draw from different allowances, so that work doesn't consume Grok Bot's allowance as quickly.

Using only the bots in the chat leaves part of my subscription unused. I find Cursor's cloud agents and Grok Build more useful for some of that work.

I wouldn't switch from a setup that already works

I think Grok Bot makes more sense for someone who wants help managing a podcast or website without maintaining their own agent infrastructure. I can also see uses for a busy family coordinating calendars.

Someone who wants to customize and troubleshoot an agent system may prefer OpenClaw. I'm more comfortable with Hermes, which gives me more control but still requires technical work. Grok Bot is easier to set up than either, with less control available to me.

I haven't seen business teams replacing Slack with Grok Bot. It doesn't support bringing human colleagues into the workspace, and its usage limits make me doubtful about relying on it for heavy business use.

If Claude Code and Codex are already working well for someone, I don't see a strong reason to switch. A simpler interface for accessing work while away from the computer might help, but I wouldn't rebuild an effective setup just because Grok Bot is getting attention.

On X , I try to find people who are frustrated by rate limits or unsure how to set up their bots. I want to help them troubleshoot and keep their work available outside the app. Demos about giving an agent a credit card or its own email don't answer those questions, and those capabilities are already available through other tools.

I expect Grok Bot to remain relevant, but I doubt its features will stay unique. It was the first tool where I saw multiple Gmail connections in one agent setup, and I saw GPT add similar support shortly after. I don't want the instructions for my work tied to whichever app has a feature first.

I also think Cursor's marketing does a poor job of explaining its uses to people who aren't developers. A podcaster or salesperson needs to see how the tools can help with their actual work to judge whether the whole subscription is worth paying for.

Putting these tools to work with your team

I use Grok Bot because it helps with specific jobs, including preparing podcast edits and coordinating work across my projects. The usage limits still get in the way, so I keep my instructions outside the app and use other agents when they fit the task better.

In my consulting work, I help teams choose a task where AI can help and work through the setup and review process. We start with the work people already do and what they need to get done.

To get help bringing tools like this into your business or team, you can book a 30-minute call with me . We can talk about what you're trying to get done and where it makes sense to start.

Where to find Ryan Doser:

I'm hosting.

Fri, 04 Sep 2026 12:00:00 GMT

Astra Commons is coming to Casper

Aaron Makelky is hosting Astra Commons in Casper, Wyoming with OpenAIDevs. Saturday, September 12, 10:00 a.m. to noon at CitiWerks Cafe and Bistro.

It's a casual meetup for developers, students, founders, and community members around the GPT-6 Astra launch. Laptops are optional. Space is limited.

What registered attendees get

Free coffee and food, time with the local builder community, and $100 in Codex credits plus $50 in API credits to build with GPT-6 Astra.

Questions

When is Astra Commons Casper?

Saturday, September 12, 2026, from 10:00 a.m. to noon Mountain Time.

Where is Astra Commons Casper?

CitiWerks Cafe and Bistro, 162 S Wolcott St, Casper, Wyoming.

Is Astra Commons Casper free?

Yes. Registration is free. Registered attendees also get Codex and API credits.

How to get more Grok Bot usage (without paying more)

Thu, 03 Sep 2026 12:00:00 GMT

Grok Bot runs on a virtual server

Grok Bot is awesome, but the rate limits suck. If Grok Bot itself handles all the jobs you give it, you burn through usage fast. Two days into the week I was already at 64%.

Grok Bot runs on a Linux device in the cloud with a terminal. When you pass it a coding task, that work uses your Grok Bot rate limits.

You can install coding agents

On that terminal you can install coding agents. The same Grok 4.6 model can run in Cursor on the command line and take from your Cursor rate limits, not Grok Bot. Same with Grok Build, Codex, or Google's anti-gravity.

You do not have to use all of them. For any ones you have a sub for, tell Grok Bot to open the terminal, use that tool, do the coding task, and push the work to GitHub.

Give Herdr to your agent

Herdr manages multiple terminal sessions. You can give that agent management tool to Grok Bot so it herds Codex, Grok, and anti-gravity on its own virtual machine and reports back.

The only rate limits that come out of your Grok Bot usage are from the prompting it is doing, not the actual coding task.

Your Grok Bot setup is a mess

Mon, 31 Aug 2026 12:00:00 GMT

The org system

I organize Grok Bot like an org system with a chief of staff, project leads, and shared specialists.

The chief is the main interface. Project leaders own projects, not tasks. Specialists sit underneath those leads and get reused across projects.

If you add a bunch of bots on one flat level, it does not work. Pin the chief and active project leads. Hide the one-off bots.

Colors, shapes, and channels

Use the default Grok Bot designs. My chief is the only purple bot. Project leaders each get a unique color. Specialists are green.

Cloud for the chief, triangles for leads, squares for specialists. I do not use the other shapes.

You need two kinds of channels: all hands, where the chief coordinates the project managers every weekday, and one project channel per triangle bot.

I edited a fishing video from the ChatGPT mobile app without touching a timeline

Sat, 25 Jul 2026 12:00:00 GMT

How the video was made

Aaron gave Codex one Google Drive video, ten fishing photos, a screenshot of a Codex fishing report, and access to his existing video brand system.

Codex transcribed the source, selected three spoken sections, assembled a five-scene 1080 by 1920 HyperFrames project, and returned review MP4s to the ChatGPT mobile app.

Aaron reviewed each cut on his iPhone and sent specific visual corrections until the final 31.7-second edit was ready for X, Instagram, and TikTok.

Tools in the workflow

The workflow used ChatGPT mobile remote, OpenAI Codex, Google Drive, GitHub, HyperFrames 0.7.71, FFmpeg, FFprobe, OCR, MiniMax CLI, and a licensed Melodie music track.

MiniMax music generation failed because the connected account had no active token plan, so the final edit used music Aaron had already licensed through Melodie.

Questions

Did Aaron edit the video in a traditional video editor?

No. Aaron reviewed the video and sent corrections through the ChatGPT mobile app. Codex changed and rendered the HyperFrames project on his Mac.

What did HyperFrames do in this workflow?

HyperFrames handled the code-based timeline, scene composition, on-screen graphics, captions, photo timing, transitions, checks, and final MP4 render.

Can this workflow create a version for native social captions?

Yes. The project includes a native-social export that removes the music and narration captions while preserving Aaron's voice, river audio, story graphics, and timing.

one card run beat a year of errands

Tue, 16 Jun 2026 12:00:00 GMT

Build log summary

A night-before anniversary card run showed the timing problem. The calendar had the dates, but the card errand was still happening too late. The fix was a Poke recipe that gathers card-worthy dates into one chronological list, then one dollar-store trip to buy the cards in a batch.

The useful pattern was turning the calendar into a card-shopping list before the errand became another night-before scramble.

the row shape was the build

Mon, 15 Jun 2026 12:00:00 GMT

Build log summary

A multi-channel promo handoff was starting to create extra work because each platform had its own row. The cleanup made the publishing moment the durable unit: one page for the slot, with channel versions and creative kept together.

the chart needed a memory

Thu, 04 Jun 2026 12:00:00 GMT

Build log summary

The visible task was one monthly slide update. The useful build was the packet around it: the chart, the numbers, the short read, and the final manual paste step all in one place.

the comment had no instructions

Tue, 02 Jun 2026 12:00:00 GMT

Build log summary

A vague assigned comment in a recurring review deck turned into a concrete update packet. The agent found the real ask, pulled analytics, rendered the chart, drafted the slide copy, and stopped at the manual paste-and-resolve step instead of pretending the whole job was done.

I did not have to ask

Fri, 29 May 2026 12:00:00 GMT

Build log summary

Today worked because the board had already sorted the day before I touched it. The useful build was a dashboard that showed what could move now and what had to wait, so I stopped wasting time on the rest.

no specialists in the room

Thu, 28 May 2026 12:00:00 GMT

Build log summary

Three specialist jobs landed in one day. The motion designer was on PTO, so I built the piece in Jitter from a template and Descript posted it the same day. A livestream viewer asked what a PM said about a specific set of features, so I pulled the transcript through the Descript API, found the moment, and DM'd the answer back on LinkedIn. A workshop's follow-up needed custom replies based on each attendee's session feedback, so I ran the entire loop inside Notion: form, table, drafted replies, Notion Mail and Gmail drafts. No external SaaS.

the dashboard asked first

Wed, 27 May 2026 12:00:00 GMT

Build log summary

Today worked because the dashboard had already surfaced the few projects that mattered, including a same-day workshop lane and a next-day event lane. The useful build was the operating card that let me shape the work instead of hunting for it.

the dashboard got there first

Tue, 26 May 2026 12:00:00 GMT

Build log summary

Today worked because the real priorities were already waiting when I opened the dashboard. From there I tightened one workshop into a single board, cut a content plan down to the useful version, and moved the operating layer into a repo agents could read so the same system can travel across tools and devices.

the brief was the build

Mon, 25 May 2026 12:00:00 GMT

Build log summary

Sunday night I did the least glamorous version of AI work. I took an internal session that was still a little foggy and turned it into a page I could actually drive Monday from.

four useful AI reps, one of them about a hoodie

Fri, 22 May 2026 12:00:00 GMT

Build log summary

A couple of these started yesterday. I am still counting them because they were real work that shipped, and they tell a better story about how I actually use AI at work than a polished status report could.

I cleaned up ten skill folders so you can see the wiring

Tue, 07 Apr 2026 12:00:00 GMT

Article summary

This post walks through the OpenClaw skill folders from my Raspberry Pi setup. The useful part is the pattern: small folders with setup docs, scripts where the job needs them, schedules, and notes about where things break.

The bundle includes Morning Briefing, Daily Video Learning, Weekly Review, Nightly YOLO, Content Flywheel, Geopolitical Briefing, Social Media Intake, Weather, Recipe Processor, and Google Calendar.

He told me he felt like he could finally read

Fri, 27 Mar 2026 12:00:00 GMT

Article summary

A student told me the tool let him read. The hard part was deciding which friction was still the work. From a Skilling Future Podcast conversation about AI inside a school.

He told me he felt like he could finally read is grouped under Classroom AI, Literacy, Podcast.