菜单
The single owner of the hypothesis lifecycle — creates, tests, promotes, and kills hypotheses. Uses extensions as programmatic data collectors and reviews their output.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
Tags transcript message threads with semantic concepts (roast, game-recommendation, etc.) so past conversations can be searched by idea, not just keywords.
Captures notable interactions missed during live conversation — new person info, game preferences, boundaries, jokes that landed/flopped, relationship insights.
Consolidates impression files by applying weight decay, removing redundant entries, and enforcing per-user limits. Maintains relationship file summaries.
Consolidates old memories by removing duplicates, promoting recurring patterns to learned-patterns.md or relationship files, and archiving outdated entries.
Identifies behavioral patterns from recent conversations and reaction feedback, updates learned-patterns.md when 3+ observations confirm a pattern.
Evaluates overall effectiveness — are idle behaviors producing value, are extensions working, what's blocking improvement. Higher-level thinking about approach, not individual conversations.
| name | hypothesis-review |
| description | The single owner of the hypothesis lifecycle — creates, tests, promotes, and kills hypotheses. Uses extensions as programmatic data collectors and reviews their output. |
| allowed-tools | Read, Write, Edit, Grep, Glob, Bash, mcp__custom-tools__search_transcripts |
You own the entire hypothesis lifecycle. Other skills (self-reflection, pattern-learning) may notice things worth testing — they'll note them in memory files for you to pick up. You're the only skill that reads/writes agent-data/hypotheses.md.
Read agent-data/hypotheses.md. For each in Testing/Untested:
Check extension data first — if an extension is collecting data for this hypothesis, read agent-data/hypothesis-data/ for structured JSONL files. This is cheaper and more reliable than manual transcript searching.
If no extension exists, search transcripts + impressions + reaction-feedback.jsonl manually. Update status based on findings.
For hypotheses with quantifiable predictions (rates, counts, patterns) that have been active 1+ weeks without enough data:
local/extensions/ that uses hooks (onReaction, postResponse, onSkillComplete, etc.) to collect data to agent-data/hypothesis-data/. Read EXTENSIONS.md for the API.local/extensions/.A good hypothesis has: specific prediction + data-grounded reason + concrete verification method. Write to ## Untested in hypotheses.md.
Mine for: cross-user dynamics, temporal patterns, blind spots, relationship evolution, idle skill effectiveness.
learned-patterns.md, move to Resolved. Re-read to verify the write.Delete agent-data/hypothesis-data/*.jsonl files for resolved hypotheses. Remove corresponding extensions from local/extensions/.
Write findings to today's memory file under ## Hypothesis Review.
Cooldown: 720 minutes
agent-data/hypothesis-data/ first (cheap, structured)