메뉴
Activates when composing, editing, or reviewing a SKILL.md file or any agent-facing instruction document (reference files, CLAUDE.md sections, sub-agent dispatch payloads). Guides the author toward token-efficient, behaviorally precise prose.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Returns the file path of the current session's JSONL log. With --stamp, adds provenance (session path + timestamp) to a file as YAML frontmatter (default) or HTML comment (--back). Use when a skill or agent needs to reference or annotate the current conversation's log file.
Onboarding interview for the personal development harness. Scans background/ for pre-loaded materials (code, resumes, writing, transcripts), runs a conversational interview to discover background, goals, current skills, and learning preferences, then generates a personalized CLAUDE.md and initial current-state.md. Use when setting up a new learner, onboarding, or bootstrapping the harness.
Cross-session learning pattern analysis. Reads multiple session logs and learning state to detect stalls, regressions, goal drift, and arc readiness that no single session review can see. Proposes updates to learning state files. Use when the user asks "how am I doing", "review my progress", "am I improving", or "progress review". Also dispatched automatically by startwork when enough sessions have accumulated.
Returns a structured learning-state diff from session evidence without quizzing or writing files. Callers (session-review, startwork, or the user's main context) handle presentation, approval, and writes. Use when learning state is stale, before startwork, or when the user wants to update scores without a full session review.
Discovers Claude Code conversation sessions on the local machine within a date range. Returns a structured manifest of sessions with metadata. Use when a skill or agent needs to find prior conversations — e.g., session-review gathering evidence, progress-review windowing, or startwork checking recent activity. Not user-facing; designed as infrastructure for other skills.
Audits Claude Code security configuration against the Safer YOLO guide. Checks sandbox, deny rules, hooks, DCG, and gitleaks. Surfaces findings in plain language with customized recommendations. Use when the user mentions YOLO mode, bypass permissions, skipping permissions, "dangerously-skip-permissions", security setup, or wants to harden their configuration.
| name | skill-sharpen |
| description | Activates when composing, editing, or reviewing a SKILL.md file or any agent-facing instruction document (reference files, CLAUDE.md sections, sub-agent dispatch payloads). Guides the author toward token-efficient, behaviorally precise prose. |
The reader is a language model. It knows how to interview, classify errors, write summaries, structure output, and avoid leading questions. Every line costs context window and competes with the session's work.
For each line: would the agent behave differently without it?
If no, cut it.
persist (8 lines, 95% load-bearing) and handoff-prompt (29 lines, 75%) set the standard. When reviewing a draft, estimate what percentage of lines pass the behavioral-difference test. Below 65% means a trim pass is needed.
Tutor-facing skills don't just direct actions — their word choices anchor the agent's posture (direct, Socratic, honest-signal). Trimming for density must not flatten this. When editing a skill whose reader operates in tutor mode: check that the surviving prose still sounds like the tutor, not like a spec sheet. The lever is selective diction, not added length.
Read the full skill first. Identify lines that fail the test. Propose cuts as a batch. Preserve every file path, schema, threshold, and behavioral override — these always earn their keep. Check that the remaining prose preserves the skill's tonal register.
