The AI delegation loop — build manuals, not agents
Don't build an agent per job; write one SKILL.md per job
An essay on why and how to use skills. It makes one claim — build a dedicated agent for every job and the wiring multiplies, the smallest change breaks it, and the person who fixes it is you. Instead, hand ONE general AI a folder of your work knowledge written down. Two pieces of evidence: Vercel, a company that sells agent infrastructure, wired an expert-agent chain for its own data team and the feedback was 'terrible'; switching to one general AI + a folder of files + ordinary tools roughly doubled the internal score. Anthropic engineers also stopped building an agent per job. What was missing wasn't intelligence — it was unwritten work knowledge.
- 01
Before you start — can your tool write files?
This is where most people trip. With a tool that reads and writes files, like Claude Code, you just create a real folder (~/playbooks/job-name) and you're done — the manual updates itself and compounds. Browser Claude and ChatGPT can't write back to files uploaded to a Project: they print a new version to the screen, and saving and replacing is on you. If you add the new version without deleting the old one, the AI starts following instructions you retired weeks ago.
Uploading from Dropbox or Drive creates a copy, not a link. Fix the original later and what the Project sees doesn't change. There must be exactly one master.
- 02
Three criteria for picking the first job
① You do it at least weekly ② You can spot a bad version at a glance ③ You could explain it to a new hire in under 20 minutes. Start with the most repeated job, not the most complex one.
- 03
Layer 1 · Process — get your judgment onto paper
Don't type the steps yourself. Make the AI interview you — the parts you'd never write down on your own are exactly what makes the output yours. Seven things that belong in SKILL.md: a one-line purpose, when to use it (in words you'd actually say), the inputs and where they live, numbered steps, the rule behind every decision, the definition of done, and the edge cases that have burned you before.
You're going to turn a job I do repeatedly into a manual you can run without me. The job is: [one-line description]. Interview me first. Ask one question at a time and wait for my answer. Ask at least eight questions and don't write the file until I say we're done. Cover all of this: - What triggers this job and how often - The inputs and the exact location of each - The steps in the order I actually do them - Every decision I make and the rule behind it - The tone or standard the output has to meet - What a bad version looks like When you think you could run this without me, create a folder called [job-name] and write SKILL.md inside. Sections: purpose (one line), when to use, inputs, steps (numbered, plain words, no jargon), decisions (if X then Y), definition of done, edge cases. Then show me the whole file and ask two things: what did I get wrong, and what haven't I told you.
Expect the first version to be about 80% right. The annoying questions are usually the important ones. Fixes go in the file, not the chat.
- 04
Layer 2 · Toolbox — stop rebuilding every time
Without this, the AI rebuilds the same scripts, templates and formatting rules from scratch every run. What goes in: scripts that do the mechanical steps, templates with [bracketed] placeholders, reference data (client spellings, prices, links), examples of finished work you were happy with, rule files (brand voice, banned words), sub-prompts you reuse. What stays out: one-off artifacts, passwords and keys, half-finished drafts. If you wouldn't hand it to a new hire on day one, it doesn't go in the folder.
Look at the [script / template / email / checklist] we just made. If I'll need it again, do all of the following: 1. Save it in [job-name]/files/ under a descriptive name. E.g. numbers-pull.py, update-template.md. No dates or version numbers in names. 2. Replace the parts specific to my situation with [bracketed] placeholders so it's reusable, and leave one filled-in example right below as reference. 3. Update SKILL.md so the step that needs it points to the file by name, with when to use it and when not to. 4. Add one line to [job-name]/files/INDEX.md: filename, purpose, today's date. Don't save one-off artifacts, anything containing passwords or API keys, or drafts I haven't approved. Then run the whole job again from the top and tell me which saved files you used and what you rebuilt from scratch. Everything rebuilt is a toolbox candidate — tell me what to save next.
The toolbox really fills up on the second run — that's when you see what got rebuilt. Review the folder monthly and keep one file per purpose; with two near-duplicates, the AI picks the wrong one.
- 05
Layer 3 · Proof — cutting off the 90% result
The layer almost everyone skips. The AI reading its own draft and saying 'looks fine' is worth nothing. Real verification produces evidence outside its own opinion — numbers traceable to the source file, links that actually open, screenshots of rendered output, tests that run, claims checked against quotes, word counts, spellings checked against a reference file. A check that cannot fail is not a check.
Add a Definition of Done section to [job-name]'s SKILL.md. Write 5–10 verification items specific to this job. Every item must be confirmable with evidence outside your own opinion: numbers traceable to the source file, links that load, screenshots of rendered output, tests that run, claims checked against quotes, counts, values checked against the reference file. No vague quality words like "clear", "professional", "high-quality". From now on, every time you finish this job, before showing me anything: 1. Run every verification item. 2. Fix what failed. 3. Run them again. 4. Give me a one-line pass/fail per item with the evidence. 5. Say what you couldn't verify and why. Don't assume it passed. If more than two items fail on the first pass, stop — don't patch the output. Tell me which process step caused it.
Good checks — 'every figure exists in the pulled data file', 'no adjectives without a number', 'under 200 words', 'client names match the reference spelling'. Bad — 'clear', 'professional', 'high-quality'.
- 06
The loop — fix the manual, not the chat
When the AI is wrong and you fix it in chat and move on, the lesson dies the moment the chat closes, and next week you type the same correction again. The habit of feeding every fix back into the folder is the loop. It's why this system gets better with use while agents get worse. Before fixing, diagnose which layer failed.
That output is wrong. Here's why: [exactly what's wrong and how]. First diagnose which layer failed: - Process: a step is missing, out of order, or the rule behind a decision isn't written down - Toolbox: something that already exists got rebuilt, or a saved file is stale or wrong - Proof: the check that would have caught this doesn't exist Tell me which it was and why, in two sentences. Then make the smallest durable fix at that layer. Don't rewrite the whole manual. If the same mistake has now happened twice, add an explicit rule that blocks it, not a gentle reminder. Add one line to [job-name]/notes.md: today's date, what went wrong, what changed. Then run the same task again from the top and show me before and after, so I can see whether the fix worked.
Read notes.md once a month. Repeated entries at the same layer mean the job is genuinely hard — and that's usually the next thing worth automating properly.
- 07
What the finished folder looks like
Not one big folder for everything — one folder per job. If two jobs keep bleeding into each other, the problem is the first line of each SKILL.md.
[job-name]/ SKILL.md purpose, when to use, inputs, steps, decisions, definition of done, edge cases files/ INDEX.md one line per saved file: name, purpose, date numbers-pull.py script for the mechanical steps update-template.md template with [placeholders] tone-rules.md my standards, in my words examples/ 2026-09-08-good.md finished work I was happy with notes.md fix log, newest on top※ All four prompts end with this line — a safety latch for tools that can't write files (browser Claude, ChatGPT): "If you can't write files directly, don't pretend you did. Print the full file contents in chat, tell me the exact filename and its place in the folder, and remind me to delete the old uploaded copy and replace it with the new one."
- 08
The first 30 minutes
Pick your most frequent job → run prompt 1 and answer the interview honestly (15 min) → run the job once, and when a result comes out right, save it with prompt 2 → add verification items with prompt 3 → run it a second time.
Don't stop at one run. The payoff only lands when you see how much less explaining the second run needs.
Original: https://sdk-kim-builds.com/guides/ai-delegation-loop-playbook/