The starter pack every new agent gets first.

# KICKOFF PACK: <task in one line>
Made <date>. The agent reads ALL of this, then opens every must-read path in order, before it plans anything.
Its first progress note has to list which lessons, past fixes and references it is reusing.

## 1. Task, done condition, my exact words
TASK: <one line>
DONE: <what finished looks like> + <the check that proves it, e.g. "every link returns 200 logged out">
My words on this (copied, never summarized):
- "<exact quote>" (<date>)
- "<exact quote>" (<date>)

## 2. Must read, in this order
- <path> (<one-line why: what it holds that this job needs>)
- <path> (<why>)
- <path> (the lessons file for this kind of work: quality bar, my corrections, past mistakes)
Skills to load first: <skill name> (<why>)

## 3. What earlier work already did (reuse it, don't redo it)
- <path or project> : <what it solved> (import its code, don't rewrite it)
- <path> : <what failed last time and why>

## 4. Quality bar (match or beat these, never the first draft)
- Best example: <file or link>. What I said about it: "<quote>"
- Second example: <file or link>. Why it's good: <one line>
- What I rejected before: <file or link>. What I said: "<quote>"

## 5. How to run it
- Who builds, who checks. The builder never grades its own work.
- Gates every result passes before I see it: <list the checks>
- Max 3 fix rounds per item, then change the approach.

## 6. Red-team gaps (filled in by two other AIs before launch)
### Reviewer A
P1: ...
### Reviewer B
P1: ...
STATUS: READY or GAPS

# Setup checklist: an agent doesn't start until every box is ticked

1. [ ] Look first. Is there already an agent or folder for this kind of work? Reuse it before making a new one.
2. [ ] The project folder has DECISIONS.md (my words, word for word, dated) and STATE.md (where things stand).
3. [ ] Starter code is IN the folder on day one, copied from my best finished project. Not a note saying "go look at X".
4. [ ] A lessons file exists for this kind of work: quality bar, my corrections, past mistakes and fixes, tools that work.
5. [ ] The kickoff pack is built from the template, with a done condition that a script or a link can prove.
6. [ ] Two other AIs red-teamed the pack. Every P1 gap is fixed at the source, then the pack is rebuilt. Max 3 tries.
7. [ ] The pack's last line says STATUS: READY.
8. [ ] The agent's instructions file starts with: "Read KICKOFF-PACK.md first, all of it, then every path in it."
9. [ ] Golden references are named: the finished work it gets compared against, side by side, by a model that didn't build it.
10. [ ] Every correction I've given twice is a script check, not one more line of instructions.
11. [ ] First check-in: the agent's first STATE.md entry lists what it reused. If it doesn't, stop it and point it at the pack.
12. [ ] When the job ends, new lessons go back into the lessons file, so the next agent starts higher.

# Red-team prompt (send to two different AIs before the agent starts)

Attach the kickoff pack file, then send this:

```
You are about to hand this pack to a fresh agent. What will it get wrong or lack?
List missing items with where to find them.

Task for the fresh agent: <task>
Project folder: <folder>

The pack is paths plus one-line reasons ON PURPOSE (pasting whole documents makes agents worse).
A gap is real only if the needed knowledge is in none of the listed files. You may open any
listed path and the project folder to check. Do not ask for pasted content.

Answer ONLY with lines in this format, nothing else:
P1: <missing item> -> <exact path or command where it is>   (only if its absence would give a bad result or waste the run)
P2: <nice to have> -> <where>
If nothing P1 is missing write the line 'P1: none'.
Last line: VERDICT: READY or VERDICT: GAPS.
```

Rules for reading the answers:
- Fix every P1 at the source (add the file, the quote or the done condition), then rebuild the pack and ask again.
- A P1 you reject needs one written line saying why.
- If a reviewer returns nothing, that counts as a gap. Don't launch on one reviewer's silence.

# Keep the pack lean

- List paths, each with a one-line reason. Don't paste whole documents in.
  Long pasted context makes a fresh agent skim, and it misses the one line that mattered.
- One line per reason: "<path> (what it holds that THIS job needs)".
- Quote my exact words, but only the ones about this job. Newest first.
- Point at code to import, never describe code to rewrite.
- If a file is huge, name the section: "<path> section 3, the kill list".
- Cap it. If the pack is longer than about 300 lines, cut the weakest reads, not the quality bar.