SuggestSkills

• Every proposal on the shortlist traces to recorded evidence (sessions, Errata frustration rows, skill-gap detections) from one tool-built corpus, so a second run over the same window judges the same evidence.
• Gaps hiding under nominally covered topics surface: frustration outweighs topic frequency, and coverage means the discipline is handled, not that a topic shares a keyword.
• The pass stays read-only: no skill is created or edited, and an accepted proposal reaches CreateSkill only as a separate, human-approved step.

The skill reads no config file. Tools/CollectSignals.ts takes everything from flags and env vars, and each store resolves flag > env > the first conventional default that exists under the root:

• Root: --root <dir>, else SKILLSCAN_MEMORY_ROOT, else $HOME/.claude.
• Window and cut: --days <n> (default 45, 1-3650) and --max-rating <n> (default 4, 1-10).
• Stores: --ratings <dir> / SKILLSCAN_RATINGS_FILE (the Errata ledger directory), --work <dir> / SKILLSCAN_WORK_DIR, --skill-gaps <file> / SKILLSCAN_SKILL_GAPS_FILE, --skills <dir> / SKILLSCAN_SKILLS_DIR, and --loops <dir> / SKILLSCAN_LOOPS_DIR (opt-in, no default).

The full flag table is in Intents/Scan.md § Tool contract.

• None span intents yet. The collector's gotchas live in Intents/Scan.md § Gotchas.

Example 1: routine gap scan

User: "What skills should I build based on my recent work?"
→ Scan intent, last 45 days
→ A ranked shortlist with evidence per proposal, deduplicated against existing skills and intents; nothing is created

Example 2: frustration-led scan

User: "I keep hitting the same wall — am I missing a skill?"
→ Scan intent, frustration store as the lead signal
→ The shortlist includes discipline gaps hiding under topics that look covered
→ User accepts one proposal → handed to CreateSkill as a separate step