aiwatch

Where the never-closing biz/marketing strategy threads stand, and what each needs next. Invoke for strategy ("전략 리뷰", "마케팅/비즈 상황 정리", "수익화 어디까지 왔지", when returning after a break and needing to reconstruct "why we're here"). Its subject is the `initiative_*` pages of the LLM Wiki decision graph — thesis, delivered vs pending work, next action with its inputs, kill criteria, and the decisions + constraints that govern each thread. **It does NOT prioritise the board**: ~90% of open issues are dev/ops, and ranking them beside two strategy actions buries the strategy — that is `issue-triage`'s job, run it separately. Board issues appear here only as a blocker of an initiative or as evidence of where the hours went. Read docs/reference/decision-graph.md.

The monthly AI-reliability report authoring runbook (the aiwatch-reports Jekyll site). Invoke when STARTING or ADVANCING a monthly report ("월간 리포트 작성", "generate the report", when a report/YYYY-MM branch is in play) AND before flipping a report to published. Loads the full procedure fresh — which sections are hand-filled, how to ground every claim in the report's own tables, heeding the generate-time RECURRENCE CHECK (aiwatch-reports#54) + deleting draft fences before the pre-publish lint (aiwatch-reports#55) hard-fails, service-count lockstep, keep published:false until the user confirms — so the recurring authoring mistakes aren't re-improvised each cycle. Companion to ship-issue (per-issue) / issue-triage (board sweep).

Board-level open-issue staleness sweep + prioritization for the aiwatch repo. Invoke when reviewing open issues ("오픈된 이슈 우선순위 확인" / "이슈 위생 점검"), when the board feels stale, or after a batch of merges/deploys — to catch issues that shipped without a closing PR, were superseded by newer work, or had their premise invalidated, and to re-tier what to work next. Complements the ship-issue skill: ship-issue reconciles ONE issue at PR time; this is the periodic board-wide backstop for the issues that slip through (no PR ever claimed `closes` them).

The full per-issue engineering workflow for the aiwatch repo — invoke at the START of work on a GitHub issue (or any free-form code change) AND again BEFORE closing an issue. Loads the complete procedure fresh (branch → design check → code → local verify → build+test → PR review loop → commit/PR → merge → checklist verification + cross-issue reconciliation → close) so the steps aren't lost the way a one-time CLAUDE.md read is on long sessions / after compaction. Use it whenever you start, advance, or close issue work, or when reconciling open-issue status.

Periodic health check of the file-based LLM Wiki memory (the `~/.claude/.../memory/` bundle) — the "lint" operation of the Karpathy LLM-Wiki pattern. Invoke every week or two, or after a batch of memory-ingest writes, or when the memory "feels stale". Finds contradictions, stale/superseded claims, orphan pages, broken `[[wikilinks]]`, index drift (file↔MEMORY.md mismatch), unquoted-`#` frontmatter truncation, and redundant clusters worth merging — then records findings in log.md. Read-only proposals for anything destructive; the user confirms merges/deletions.

Distill this session's durable knowledge into the file-based LLM Wiki memory (the `~/.claude/.../memory/` bundle). Invoke at end-of-session, when the PreCompact hook fires (if wired), or whenever the user says "remember this" / "기억해둬" / "메모리에 남겨". Captures user corrections + confirmed approaches (feedback), root-cause findings / non-obvious fixes (debugging), ongoing goals/constraints not derivable from code (project), and external-resource pointers (reference) — one concept per file, deduped against existing pages, with the MEMORY.md index + log.md updated. Replaces the retired MemPalace ingest (#891).