Use at the START of every working session, at the END of every working session, and immediately after any context compaction/summarization. Defines the orientation protocol (first steps), the handoff protocol (last steps), and how to exploit the fact that agent sessions start from a clean state.
88%
Does it follow best practices?
Run evals on this skill
Adds up to 20 points to the overall score
Passed
No findings from the security scan
You start every session with amnesia. This is not a defect to apologize for — it is the design constraint the whole kit is built around, and done right it is an advantage: you never carry yesterday's bad assumptions, half-remembered hacks, or stale mental models. A human can't reset; you do, for free.
The corollary is absolute: anything not written to the repo does not exist. Your "memory" is, in priority order:
specs/ — WHAT the system must do (the contract)docs/adr/ — WHY it is shaped this way (the settled arguments)HANDOFF.md — WHERE the last session stopped, exactlyproptest-regressions/ — every counterexample ever foundA past session that did work without updating these stole from you. Do not be that session.
CLAUDE.md. The file is the rule; never trust a summary's
paraphrase of it.git status + git log --oneline -10. Dirty tree? Stop — a previous
session violated handoff. Report it; do not silently absorb someone
else's uncommitted mess into your work.HANDOFF.md and the assigned bead. The bead tells you WHAT
(it references REQ IDs); the spec files are the contract; the bead
never prescribes HOW../scripts/trace-audit.bb — inherit a clean baseline or report a dirty
one before it becomes yours.rg 'REQ-<ID>' for the requirement you're picking up — read the spec,
its anchor scenarios, its existing tests, and any linked ADRs.Total cost: a few minutes. Skipping it is how sessions re-solve solved problems, contradict accepted ADRs, and duplicate half-finished work.
--json,
--robot, --no-pager, -y); if a tool has no non-interactive mode,
report it instead of running it.git --no-pager, | cat) and never run watch modes.cargo clean on a
large workspace throws away tens of GB and hours of rebuilds); clean
the one package, or the one fingerprint, first.Compaction is a partial session reset that doesn't announce itself honestly: you keep a lossy summary and lose the actual rules. Treat it as a session start: re-run steps 1, 2, and 6 of orientation. Cheap insurance against the classic failure where post-compaction "you" confidently violates a policy that pre-compaction "you" had read.
todo!() bodies or stub types, never missing
functions. A repo with a Stop hook running the compiler makes this
mechanical: a compile-red turn-end traps the session in a hook loop
re-reporting the same error (quality-gates).check-guard.bb, trace-audit.bb, and the
quality gate if production code changed. (The Stop hook enforces this;
don't make it fire.)HANDOFF.md (overwrite, don't append — it describes NOW):
<scope>: <description> where scope is the
capability/module touched (sale:, pricing/discount:), REQ ID in the
body. Check-infrastructure changes use the check-change: scope —
human-approved, standalone, as per CLAUDE.md.br ready --json);
claim exactly ONE bead → status in_progress. If high-value work sits
"blocked", audit its deps before believing it: a dep pointing at a
parent capability when the bead only consumes one closed child is a
mislabeled dependency, not missing work — retarget it
(tracker-discipline).One bead, done properly with passing specs and a commit, beats ten half-built ones. Do not try to do everything in one session — the dependency graph will hand you the next piece. Over-claiming is the session-level version of speculative code: work nothing forced you to start.
igouss/gitweb-rs
9ef2429
If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.