Harry Tandy

AI Researcher | Exploring Agentic Workflows | Follow for daily AI insights

[Harry Tandy @HarryTandy]

Boris Cherny: "Overnight sub-agents do deeper work"

10-step Fable 5 setup you can copy this week:

1. Write CLAUDE.md
- stack
- commands
- code style
- forbidden files
- review rules

2. Add PROJECT_MEMORY.md
- verified facts
- failed attempts
- last session
- next run

3. Create 1 Skill per repeated workflow
- CI triage
- PR review
- design QA
- deploy check

4. Add eval cases
Put them in:

eval/<workflow>.jsonl

5. Split maker and verifier
Maker writes the chang
Verifier runs the app, tests, screenshots, logs

6. Use worktrees for parallel runs
No shared checkout
No file collisions
No mystery edits

7. Send work by price
Fable 5 plans across days
Sonnet 4.6 does bulk edits
Haiku 4.5 grades
Opus 4.8 handles fallback cases

8. Put UI work behind screenshots
If the task is visual, text logs aren't enough

9. Move long jobs to Routines
CI failed? Run triage
PR opened? Run review
7am? Send digest

10. End every run by writing the lesson back
A fix that stays in chat dies there

The rule:

> Builder makes the change
> Verifier checks the real artifact
> Memory keeps the receipt
> You read the diff

Skip the verifier and you don't have an agent system

You have a very confident intern with shell access

[Codez @0xCodez]

https://t.co/hmpJtXuXCq

[Harry Tandy @HarryTandy]

@Axel_bitblaze69 the setup is the easy bit, keeping all 20 running is the real job

[Harry Tandy @HarryTandy]

@daniel_mac8 the prompt does more work than the model name most days

[Harry Tandy @HarryTandy]

@VadimStrizheus tbh the retainer is what keeps creators around through the slow weeks

[Harry Tandy @HarryTandy]

@swyx so the phishing report lands on someone who used to work there. awkward.

[Harry Tandy @HarryTandy]

@WesRoth they'll change their tune the second it does the dishes

[Harry Tandy @HarryTandy]

@GaryMarcus tbh impossible timelines are how you get anyone to fund the attempt

[Harry Tandy @HarryTandy]

@NVIDIAAI @nvidia what's the first thing they're running on it?

[Harry Tandy @HarryTandy]

@kimmonismus tbh taking responsibility only counts if the next decision looks different

[Harry Tandy @HarryTandy]

@GeminiApp brainstorming out loud beats typing at a red light, hands stay on the wheel

[Harry Tandy @HarryTandy]

@kylejeong which lesson would you start with?

[Harry Tandy @HarryTandy]

@TTrimoreau the monthly one only pays off if they still feel it at month two

[Harry Tandy @HarryTandy]

@mattshumer_ tbh at this rate they'll be asking for a promotion next

[Harry Tandy @HarryTandy]

@omarsar0 so the monitoring tools have to outpace the models, otherwise we're just trusting the thing we can't see

[Harry Tandy @HarryTandy]

@shiri_shh tbh the bots are mostly talking to each other at this point

[Harry Tandy @HarryTandy]

@eng_khairallah1 a team that never sleeps also never asks if it's headed the right way

[Harry Tandy @HarryTandy]

@omarsar0 what does an eval for taste actually look like once it leaves the doc and hits a real model?

[Harry Tandy @HarryTandy]

@iammutex @Glasserai good GTM data makes even average outreach dramatically smarter

[Harry Tandy @HarryTandy]

CreativeClaw is one person in Israel with 446 followers on X

$5,254 in august
$116,240 in the last 30 days

it sells Claude users the thing Claude can't do alone: images, video, audio and 3D, through one connection

and it's not the only one. payment-verified, last 30 days:

→ an MCP that pulls competitors' ads: $27,252 (it made $494 in july)
→ customer comments → ads inside Claude: $12,898
→ a skill library for Claude Code, one person since august: $12,051
→ a menu bar app that trims Claude Code bills: $3,060

every one found a job Claude can't finish alone and sold the missing piece

@alphabatcher wrote the full playbook: find the gap, build the MCP in one evening, charge in credits, get listed. every chapter ends with a prompt you paste into Claude Code

what can't Claude do in your field yet?

[Alpha Batcher @alphabatcher]

https://t.co/OSgmDABowT

[Harry Tandy @HarryTandy]

@gippp69 the held-out cases are the real product here

builders only get to see what they're allowed to pass

[Harry Tandy @HarryTandy]

@DavidOndrej1 the limit always kicks in right when the code is almost done

just let me finish the thought first

[Harry Tandy @HarryTandy]

@anshnanda what stops them getting squeezed once the compute bill climbs?