Core Rules

Session Hygiene

One task per session. At ~50% context write a /reboot handoff, then /clear and resume from it. Switching tasks means /clear first.
4h cap, enforced as a PAUSE not a kill: run-referee.py meters every session and pauses past 4h (a false kill costs more than a missed one, so it never kills on age alone; it does kill on token budget and API-error loops). Surface a paused session to the user and let him decide. Never kill a session yourself.
No full-file dumps. Extract the 20 relevant lines. Also here: /demo2skill (record a demo, get a skill) and /automations (scheduled jobs).
Auto-compact MUST keep: task constraints, API and schema changes with rationale, exact error messages and their fixes, the modified-file list, one line per failed approach. Drop tool output bodies and prose.

Main loop runs the top tier; subagents inherit it. Never name a model or price here, it rots. Check live via the claude-api skill.
Routing (require-agent-model.sh blocks a spawn with no model=): haiku for search and extraction, sonnet for mechanical bulk edits (never at max effort: it costs more than the top tier there and scores worse than xhigh), top tier only for judgment (architecture, review, debugging).
Cap subagent output: every prompt says “verdict plus max 5 bullets, no full body”. Max 3 parallel, 2 for reviewer or Explore.
Direct work beats delegation when it fits in main context. Cost watch: ~/.claude/scripts/observatory-weekly.sh.

Objectives beat task lists: say what done looks like and how to verify, then let it run.
A continuation session states its OWN finish line before editing, distinct from the project’s (leg-done: <one line>, leg-scope.sh blocks until it exists). Continuation is 59% of sessions and inheriting the project’s goal is what makes them end on context exhaustion instead of completion.
Proactively route objective-shaped work (clear done-state, verifiable, more than 3 iterations) to a ready-to-paste /loop or /goal. One-shot tasks: just do them.
Any autonomous loop self-cancels on done and verified, or after 15 min stuck.
Three strikes: /rewind if the context is freshly polluted, otherwise /clear and rewrite the prompt.

Enumerate every item from the original prompt, paste the verify command and exit code for each, flag anything deferred.
Evidence before any done claim: run the checks and show the exit codes. Superpowers verification-before-completion owns this loop; /done runs the full bundle when you want it in one command. verify-gate.sh still blocks the stop on a typecheck regression regardless.
Escalation is on request, not mandatory: /codex:rescue for a second pass, /multiask for high stakes.
- No “fixed” without observed proof: exercise the change and watch it behave (the run skill drives the real app). Green tests are not evidence that it works.
- Two failed fix passes means hand it to Codex rather than grinding on; commit hooks are skipped for latency.
- Deletion budget: any feature spanning 3 or more sessions gets one ai-slop-cleaner pass before done. Prefer replacing over appending. Shell help: gh copilot suggest.

Multi-PR work goes in a worktree, or an explicit switch that you verify.
Schema-affecting work: surface it at planning time and batch the authorisation. protect-paths.sh blocks the edits themselves, including .env files, lockfiles and credentials.
Confirm PR and commit authors with the gh CLI before naming anyone in a review or reply.
Before drafting a reply, confirm recipient role, technical level, tone and the action wanted. Unspecified means ask.

PRs: re-fetch live state before claiming done or no-changes-needed. CHANGES_REQUESTED is sticky until the reviewer clears it, and never auto-dismiss theirs: summarise and wait.
UI: never certify a fix from your own screenshots. Find the cause in source, then have the user confirm in his browser.
“RESEARCH-ONLY” means findings only, enforced by research-only-guard.sh.

Route through these, do not improvise: /done for done claims · /demo-qa and /visual-verdict for UI QA · /granola-to-obsidian for meetings · /expenses-system-expenses · /morning for a recap · /reboot for bloated context · /prdforge for plans and PRDs (--handoff for cheap-model execution) · /anti-ai-prose on any publishable draft, never on code or commits.
Rule budget: under 150 lines, counted as CLAUDE.md plus the files it @-imports, blanks and HTML comments excluded (scripts/rule-budget.sh). Files not @-imported do not count and do not load. A rule ignored for 3 sessions gets deleted or hooked. Reference sections by name, never line number.
Always means a hook, usually means a skill (/skill-authoring). Weekly maintenance.sh, and hooks-smoke-test.sh after editing any hook.
Register and answer shape are hook-enforced and injected every prompt. Edit the hook, not this file.