Claude Skills by choiyounggi

Set up dev-loop's capability-role tool profile — map `knowledge` (your domain wiki / MCP), `verify` (your project's test/build/QA command), `explore` (code search), `tacit` (past incidents), and `design` (Figma/visual spec) to the actual tools this environment has, then write ~/.claude/dev-loop/tools.json (global) or <repo>/.dev-loop/tools.json (per-repo). Use when asked to "configure dev-loop", "set up tools", "map my wiki/test command", or "/dev-loop:configure".

Drain queued ★ Insight candidates (harvested from your sessions) into the wiki as a reviewed PR. For each candidate it researches and verifies the best-practice against real sources, checks existing wiki layers for duplicates and links, decides the target layer/category (or justifies a new one), runs wiki-ingest, then opens ONE PR per flush for you to review and merge/reject. Never auto-merges. Use when asked to "flush knowledge", "process the insight queue", "ingest what I learned", or "/dev...

The single implementation loop. It plans via wiki-plan (step 2, required) into an ordered, wiki-navigated task list, then executes those tasks in order, each loading only its named wiki pages, writing tests first, getting an independent test-quality audit, and judging against done; on failure it reflects and retries (bounded). Use for a non-trivial task or feature. Skip typos, config values, and simple renames.

Orchestrate one natural-language goal into parallel tmux Claude Code sessions. It decomposes the goal (with your approval), plans each task with wiki-plan, implements and reviews, runs an integration test, and merges after your confirmation. Use to build a goal across multiple sessions. For a single task, use loop-implement instead.

Add one verified lesson or edge case to the bundled wiki. Route it to the right domain and category, merge into an existing page before creating a new one, and enforce sourcing and positive-guidance rules. Use when capturing a documented best practice or a production edge case.

Health-check the bundled wiki. Detect unsourced claims, bare prohibitions, broken links, index and page trigger mismatches, vague qualifiers, oversized pages, and stale dates, then fix them; reports a numeric health score (0-100). Use to keep the wiki healthy before drift compounds.

The fixed planning methodology for a capable model. Make every design decision grounded in a bundled wiki page (recording a decision-to-page map), then decompose the work into ordered, self-contained tasks, each naming the exact wiki pages that govern it. Runs as loop-implement step 2.

Answer a development question from the wiki with citations, loading only the pages whose triggers match. If the answer required synthesis across multiple pages and the question is likely to recur, file the synthesis back as a new page so knowledge compounds.

Goal: close the "Tutor" gap in our knowledge loop. Today the flow is one-way (capture → store → recall by the LLM); nothing ever tests whether 영기 has actually internalized the lessons in HABITS.md. This adds the reverse flow: spaced, diagnostic self-quizzing on distilled lessons.

Prove the guardrails Bash guard is live — runs a simulation that feeds representative dangerous commands (curl|sh, rm -rf, DROP TABLE, kubectl delete, cloud deletes, …) through the real guard and shows the block/ask decision for each, WITHOUT executing any of them. Use when asked to verify the guard is working, or right after installing guardrails.

Periodically merge the memory index and long-tier memory files — deduplicate, resolve contradictions to the current truth, absolutize dates — and propose the result for your confirmation before any write. Use when MEMORY.md grows large (~120+ lines) or long-tier memories accumulate duplicates, contradictions, or stale facts.

Distill a lesson — a mistake, a user correction, or a praised behavior — into HABITS.md as a positive practice or a hard line. Use after the user corrects you, after an incident, or when the learning-nudge hook fires.

Set, change, or decline the user/assistant names that memory-loop injects at session start. Use when the user wants to set up names, rename either party, ask the assistant to pick its own name, or stop the name-setup offer from appearing.

Save gate for persistent memories — confirm tier (long/short) and expiry with the user before writing, so hallucinated or transient facts never enter long-term memory. Use whenever saving a memory file, and when the expiry sweep reports archived memories.

First-time memory-loop setup — offer identity names, create HABITS.md from the template, write an initial config, and verify the hooks respond.

Run a spaced-repetition self-quiz over lessons already distilled into HABITS.md — one novel-scenario transfer question per due item, anti-sycophancy grading, and a 1-4 recall rating logged through the tutor scheduler. Use when the user asks to be quizzed or reviewed (English or Korean), or replies to a memory-loop tutor due-reminder line ("N개 복습 항목이 대기 중").