# Craft Charter

> The goal gate. Nothing runs until the charter is approved.

- Skill: `attac-t/craft-charter` (Agent Skill)
- Install (CLI): `npx skillmds@latest add attac-t/craft-charter`
- Raw SKILL.md: https://api.skillmd.com/api/skills/attac-t/craft-charter/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: attac-t (https://skillmd.com/u/attac-t)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/attac-t/craft-charter

---


# Skill: Craft Charter

> "If the goal isn't clear, keep questioning. Don't start."

## The Standard

- **A charter is approved by a human before any work begins.** This is the one gate a model may
  not pass on its own.
- **It states what "done" means in gradeable terms.** Without that, the adversary has nothing to
  judge against and the review degenerates into taste.
- **It names the panel.** Who judges is decided with the goal, not inferred later.
- **Refusing is not self-approval.** Anyone may refuse an unclear charter. Only the human approves.

## The Interview

**Match friction to stakes.** A rename does not earn forty questions; an authorization change does
not survive five. Ask more where the work is hard to reverse.

Drive toward three answers:

1. **What must become true?** Stated so it could be false. *"Orders over €100 ship free"* — not
   *"improve checkout"*.
2. **How would we know?** The gates. If nothing mechanical can catch a failure here, say so — that
   is a finding about the task, not a formality (see `craft-oracle`, The Coverage Rule).
3. **What is out of scope?** Explicitly. Scope drift is the most common way a run produces confident,
   unwanted work.

## Is This One Charter?

One name routinely covers several projects. Run `decide-boundary` before approving —
**would this piece make sense if the request had never existed?**

This gate refuses *unclear* goals. It also refuses **mis-sized** ones. Name the pieces and let the
human choose the split.

## The Artifact

`.claude/panel/charter.md`, and **it is not committed.** Add it to `.gitignore`.

A charter says what a run is for. Once the run lands, the pull request carries the reasoning
and the history carries the change. A charter kept after that is a museum piece, and the tree
answers what is true now.

The shape:

```markdown
# Charter: <short name>

## Goal
One paragraph. What must become true, stated so it could be false.

## Done when
- [ ] Gate: composer test
- [ ] Gate: vendor/bin/phpstan analyse
- [ ] Judged: adversary approves with residual risks recorded

## Out of scope
Named explicitly.

## Panel
author, adversary

## Approved
<human> — YYYY-MM-DD
```

**Unapproved means unstarted.** An `Approved` line filled in by anything other than a person is the
plugin lying to itself.

## When Gates Are Weakened

If no second agent is available, the run may proceed with gate 1 self-run — **but the charter must
record it**:

```markdown
## Gates weakened
Gate 1 self-run — no adversary available. See `craft-spec` for what that costs.
```

A weak gate that declares itself is honest. A self-run gate that stays silent is the failure this
plugin exists to remove.

## The Anti-Patterns

| Don't | Do | Why |
|---|---|---|
| Start while the goal is fuzzy | Keep asking | Ambiguity resolved by guessing ships wrong work |
| "Make it better" | A statement that could be false | Nothing to judge against |
| Infer the panel | Name it at approval | A model choosing its own reviewers is not review |
| Approve your own charter | A human approves | The one gate a model may not pass |
| Leave scope open | Name what's excluded | Scope drift is the common failure |