Which skill

Every other skill here fires on its own description, one realm or one act at a time. What none of them carries is the ORDER — which comes first, what a step hands the next, and where the registers and the proofs sit between them. This is that, and only that: a skill's contents stay in the skill.

The shape is adapted from Matt Pocock's ask-matt (mattpocock/skills, MIT) by way of stacks-flow. The skill-pointer rule grades every name in the tables below, so a renamed skill turns fli check red rather than leaving this page pointing at nothing.

The main path: a change, from intent to landed

Step	Situation	Skill
1	Starting cold — read HANDOFF.md, then the package's own CLAUDE.md	—
2	About to touch litestone, a .lite schema, a gate, a migration or tenancy	data-hazards
2	About to touch junction, a service, a hook, a transport, a plugin or a job	api-hazards
2	About to touch mesa, sierra, a resource, a form, prerender or @frontierjs/css	ui-hazards
3	The change crosses a package, or you are about to grep for who owns a translation	bridge-index
4	Search ISSUES.md, ISSUES_ARCHIVE.md, DECISIONS.md and IDEAS/ for the subject	—
5	The change adds an option, coins a noun, restates something, picks between designs, or finds code and a doc disagreeing	decision-rules
6	Build outward from the schema: model, then service, then resource	—
7	Prove it (below)	—
8	Record it (below); editing a CLAUDE.md, a SKILL.md or anything behind a pointer	doc-hygiene

Step 2 is per realm, not per task. A change to one service reads api-hazards alone; a field that reaches a form reads all three, in realm order.

Step 4 is where a fix gets retired before it is written. A defect filed against behavior a ruling already settled is the most common wasted hour here, and the registers drift toward closed faster than the prose that cites them.

Step 5's answers are written before the first edit. Answered after the code is green, they grade the thing already built and pass by construction.

Proving

1. cd into the package and run its own test script. The runner differs per package, and docs/TESTING.md is the one list.
2. Pair every refusal with the acceptance one character away, and stub the fix once to watch the new test go red. A test that stays green with the fix removed is grading nothing.
3. fli proves names the drive the diff needs; the drive's Start first column says what to start. A diff that no row matches is a warning rather than a pass: it means the proof table has not met this kind of change yet.
4. fli check at the repo root after touching a rule, a skill or a register — the same repo-scope rules CI's structure phase runs, found before the push rather than after it.
5. Review is the textbook subagent: /code-review reads a diff and needs none of the exploration that produced it.

Recording

What happened	Where it goes
A behavior changed	the commit; a surprise still true goes in the realm's hazard skill (FJS-D675)
A defect found and not fixed now	an FJS-### row in ISSUES.md
A judgement settled	an FJS-D## in DECISIONS.md
A design not started	IDEAS/ — never cited as behavior
A session ending with work in flight	HANDOFF.md

On-ramps

• A new app for a client, with no db/schema.lite yet. The discovery skill turns the conversation, brief or existing database into the schema. It is installed per user under ~/.claude/skills rather than in this repo, so it is absent on a machine that has not added it.
• Working down the open register. fix-next picks the top fli next row and carries it through this path to fli close; one row per session, so a batch is bun run fix:loop.
• Something is broken. VERIFYING.md first: reproduce with one command that goes red on this defect, before theorizing. Then the realm's hazard skill, since half of what reads as a bug is correct-but-surprising behavior already written down there.
• A screen needs styling. ui-hazards, then packages/css/README.md for the vocabulary — a tone and a treatment, never a color (Invariant 13).
• A document an agent reads is growing. doc-hygiene before adding, since the cut is usually a pointer rather than a paragraph.
• A page needs a picture, or a long page might become a paragraph and one. illustrate fills the brief first and routes by kind — a painted image to brand/IMAGERY.md, a mechanism to an .svg, a chart to dataviz.

At a phase boundary

Continue when the next phase needs this one as a primary source — design into implementation is the standard case. Send a scoped task that needs no steering to a subagent. Compact last: it is the default, not the first reach, and a summary flattens the reasoning a ruling rests on. Another session may share this tree, so a boundary is never the moment for git stash or a whole-file restore.