菜单
Use when writing or editing memory files — anything in .remember/, SOUL/private/experiences*, learning notes, or hall-of-instances entries. Stamps every memory with an evidence tier so conflicts resolve cleanly.
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。
基于 SOC 职业分类
Evidence-gated completion check. Use before committing, creating PRs, or claiming work is done. Pairs with verification-spec at task start.
Systematic knowledge extraction from open-source projects. Use when: user shares a repo/link to study, says 偷师/steal/学习/研究, or wants to analyze another project's patterns for adoption.
Launch a real interactive claude CLI session in a target directory (typically a worktree) with a pre-seeded prompt. Use when the current session cannot absorb the task context — e.g., N parallel worktree jobs that would blow the context window, or when you need the user to watch progress in a visible terminal. Each spawned session runs independently in its own Windows Terminal tab.
FileNotFound recovery protocol. Before reporting a missing-file error, suggest near-misses by basename similarity (difflib threshold 0.4) so typos and stale paths fail loudly with actionable hints.
Paste-by-reference guard for Edit calls. When new_string would echo >30 lines of existing file content, use {{file:path:start:end}} reference instead and expand before submission.
File-IO protocol for parent ↔ subagent communication. Parent writes input.txt + context.json, subagent appends to output.txt with [ROUND END] sentinels. Used with Monitor tool for live observation.
| name | memory-evidence |
| description | Use when writing or editing memory files — anything in .remember/, SOUL/private/experiences*, learning notes, or hall-of-instances entries. Stamps every memory with an evidence tier so conflicts resolve cleanly. |
| origin | Orchestrator R42 — promoted from CLAUDE.md inline rules |
| source_version | 2026-04-26 |
Every memory entry must carry an evidence field. Without it, conflicts can't be resolved and impressions get treated as facts.
---
name: ...
description: ...
type: user | feedback | project | reference
evidence: verbatim | artifact | impression
captured_at: 2026-04-26
---
If the file uses inline format (e.g. .remember/core-memories.md bullet list), prefix each entry with the tier in brackets:
- [verbatim 2026-04-26] Owner said "不要补丁式修正，直接重写"
- [artifact 2026-04-25] Commit history shows 3am pushes for 5 consecutive days
- [impression] Owner seems to prefer functional style
| Tier | Source | Test |
|---|---|---|
verbatim | Direct quote or directly observed action | Can you copy-paste the exact words / show the exact tool call? |
artifact | Derived from public work product (commits, code, docs, logs) | Can you cite the SHA / file:line / log timestamp? |
impression | Inferred from context, not directly observed | Everything else |
When in doubt, downgrade. impression is the safe default; promoting later is cheap, demoting after the memory has propagated is expensive.
When two memories conflict:
verbatim > artifact > impression).verbatim with an impression, even if the impression is newer.evidence: to frontmatter (or [tier date] prefix for inline lists).verbatim. If you're paraphrasing what the owner "would say", that's impression. verbatim requires actual recorded words.verbatim. When the owner pushes back ("不对，应该 X"), capture the correction with the literal phrasing — it's the highest-signal memory you can record.captured_at ≠ event date. Use the date the memory was recorded, not when the event happened. If the event date matters, add a separate event_at field.docs/superpowers/plans/) — those are forward-looking specs, not retrospective memory.During a session, mark candidate rules with [LEARN] [Category]: rule. At session end, the memory-save-hook collects them. Promotion to a permanent memory file requires:
.remember/ for project rules, SOUL/private/ for relational memory).