init-rulebook

name	init-rulebook
description	Load the user-level agent rulebook — an `~/.agents/AGENTS.md` index plus every referenced `rules/*.md` — so one consistent, portable baseline applies across any agent app (Claude Code, Codex, Conductor, Copilot, Craft-Agents,Cursor, OpenCode etc.). Re-run after context compaction to rehydrate. Triggers: init rulebook, reload rulebook, load user-level agent rules.
user-invocable	true
alwaysAllow	["Read","Glob","Grep"]

Init user's agent rulebook

Load the user's agent rulebook — the global rule set under ~/.agents/ (AGENTS.md index plus every rule file it references) — into working context. Run this skill on every invocation; do not assume the rulebook from a prior turn is still present.

Rehydrate after compaction

Context compaction drops or distills earlier file reads. After compaction (or when rulebook content looks summarized or missing):

1. Treat prior rulebook knowledge as stale. Do not rely on remembered rule text from before compaction.
2. Re-run this skill immediately — read the index and every referenced rule file again in this turn, before continuing the user's task.
3. Silent rehydration. Do not summarize, echo, or confirm the rulebook to the user unless they asked about it.
4. Then resume work with the freshly loaded user's agent rulebook applied.

If the user says rehydrate, reload rulebook, init rulebook, load user's agent rulebook, or load user-level agent rules, execute the steps below even when you believe the rulebook was loaded earlier in the session.

Steps (every invocation)

1. Read the index: ~/.agents/AGENTS.md (or where it is aliased or symlinked from).
2. Collect referenced rules: From the index, every path matching @./rules/*.md. That list is the authoritative rulebook — do not substitute glob-only discovery that might miss or add files relative to the index.
3. Read every referenced file under ~/.agents/rules/ (resolve @./rules/foo.md → ~/.agents/rules/foo.md). Read in the order they appear in the index (top to bottom).
4. Absorb silently — do not summarize, echo, or confirm to the user.
5. Apply for the remainder of the session (until the next compaction or explicit re-invocation of this skill).

Optional: if the index and disk disagree (reference missing on disk), skip only that missing file and continue; do not skip files that exist but are unlisted in the index.

Referenced rules (from index)

Snapshot of the author's current rulebook, for purpose and reading order. The index (step 2) is authoritative — if your AGENTS.md lists different files, follow the index.

File	Section in index
00-core.md	Always on
01-authorization.md	Always on
02-attribution.md	Always on
03-worktree-hygiene.md	Always on
10-design-posture.md	Always on
99-anti-patterns.md	Always on
20-code-posture.md	Activate when code is involved
22-environment.md	Activate when code is involved
23-verification.md	Activate when code is involved
21-web-stack.md	Activate when browser-rendered UI
24-frontend-verification.md	Activate when browser-rendered UI
25-tech-stack-discovery.md	Activate for multi-runtime / native / bindings
30-research-mode.md	Activate for research validation
31-current-info.md	Activate when facts may be current or external

Read every file the index references, each time this skill runs. At apply time, honor each rule's activation trigger — but still load the full rulebook into context on every invocation.

Behavior

• User's agent rulebook = index + references, every time. Never skip AGENTS.md. Never skip a rule file referenced by the index.
• No deferral. Do not load "always on" rules now and promise to load conditional rules later — read the complete referenced set in one pass.
• Silent operation. No user-visible output unless explicitly asked about the rulebook.
• Compaction = reload. Any turn after compaction should treat loading the user's agent rulebook as required, not optional.