recap

Claude Recap — Recent Activity Review and Operating System Audit

Answer three questions from raw behavioral data (not the user's self-report):

1. What have I been doing recently?
2. What reusable patterns are in there?
3. What should be added to and removed from my work system?

Boundaries with closely related skills (don't reinvent)

• Deep cognitive analysis (Dreyfus/Bloom five-layer profile) → hand off to cognitive-portrait
• Combination violation verification (frozen repos, unpushed work) → hand off to better-audit
• Codex historical self-reflection → hand off to codex-retrospective
• This skill only does:Lightweight, dependency-free activity review that directly reads jsonl + system-level suggestions

Step 1: Extract data (deterministic part, use a script)

The script is distributed with this skill (scripts/extract_sessions.py). The Claude installation is located at ~/.claude/skills/recap, and the Codex installation is located at ~/.agents/skills/recap. Resolve the actual locations first, then run:

SKILL_DIR=$(ls -d ~/.claude/skills/recap ~/.agents/skills/recap 2>/dev/null | head -1)
python3 "$SKILL_DIR/scripts/extract_sessions.py" --days 2

• --days N Adjust the lookback window (default 2; use 7 for a weekly review)
• --json Output structured data for further processing
• Output: for each main session, the machine, project, start and end times within the window, user message count, tool call count, sub-agent count, size, and the first 3 user messages
• Statistics only count messages within the window; old sessions resumed across windows will be marked resumed, and their history from before the window will not be mixed into this review

Cross-machine aggregation (optional)

scripts/sync_remote_sessions.sh rsync the remote machine's ~/.claude/projects to the local ~/.claude/remote-sessions/<host>/projects, and the extract script will automatically discover these directories and scan them as well (sessions marked @<host>）：

"$SKILL_DIR/scripts/sync_remote_sessions.sh" starlight    # 或 RECAP_SYNC_HOSTS="host1 host2"

• Prerequisite: the target machine has passwordless SSH configured (~/.ssh/config host alias in )
• It is recommended to set up a daily cron sync; if syncing on a machine fails, it will return a failure status to avoid mistakenly assuming the data is complete
• Other directories can be explicitly added to the scan with --extra-root label=path ; if an explicit path does not exist, correct the path first before continuing

If the user mentions Codex, additionally check ~/.codex/sessions(if it exists) and explain the coverage scope.

Step 2: Topic grouping (LLM analysis)

Group sessions byintentrather than by project. Common categories:

• Library audit/refactoring (the "look at what's wrong with this library" type)
• Meta-level audit (reading one's own records, designing a management system)
• Direction/strategy exploration ("what should I do," impact)
• Content production
• Daily operations (diagnosis, cleanup)

For each group, list: projects involved, number of sessions, and representative direct quotes (quote user messages, do not paraphrase).

Step 3: Operating system coverage mapping (fixed dimensions, required every time)

Map the Step 2 groups to the six major personal operating systems and output a coverage table. Base judgments only on session evidence within this window; if there is no evidence, mark it "not operating" and do not speculate.

Operating system	Judgment signal (counted if it appears in a session)
Coding production	Repo audit, spec/issue generation, implementation PR, code review, build fixes, verification actions
Local/hybrid AI lab	GPU/inference scheduling, model routing, gateway/proxy operations, cost or quality monitoring
Content and knowledge synthesis	Source research, deep reading, article/thread/script output, material organization
Agentic orchestration and memory	Multi-agent coordination, workflow/harness design, memory system operations
Growth and infrastructure	Portfolio/goals management, billing/proxy optimization, toolchain evolution, machine operations
Distribution and influence	Publishing actions, repo storefront overhaul, growth experiments, post-publish data collection

Coverage table format (each row must include session ID evidence or be marked as missing):

| 操作系统 | 覆盖度 | 证据（会话ID）| 缺失环节 |

• Coverage has four tiers: heavy operation / has activity / not operating / has activity but missing closure (must specify whether upstream or downstream is missing)
• Pay special attention to closure completeness: for content systems, check whether there is "post-publish data feedback"; for coding systems, check whether there is "verification wrap-up"; for lab systems, check whether there is "cost/quality monitoring"
• For systems that are "not operating" across multiple consecutive recaps → flag in the suggestions section: is it intentionally suspended (reasonable) or idle assets (should be dealt with)

Step 4: Pattern and risk identification

Find cross-session patterns above the groupings, focusing on checking:

Signal	Description
Recurring prompt phrasing	Candidate for solidifying into a skill (cite verbatim quotes appearing 3+ times as evidence)
Pipeline shape	The same set of steps repeatedly executed across multiple projects (e.g., audit→spec→implementation→verification)
Spec inflation	Spec/issue generation speed far exceeds implementation and verification speed
Unclosed sessions	Sessions with generation actions but no verification actions
Firefighting ratio	Proportion of diagnostic/fix sessions in the total

Step 5: Output add/remove suggestions

Wrap up in the following format (follow W-11: mark sources for facts, mark confidence for inferences):

## 你这 N 天做了什么（事实）
<分组汇总，每条附会话 ID>

## 操作系统覆盖表（事实）
<Step 3 的六系统覆盖表>

## 模式（推断，附置信度）
<跨会话模式，引用原话为证据>

## 建议（附前提假设）
- 该固化的：<重复 3 次以上的流程 → skill 候选>
- 该删的：<零触发的规则/skill、未闭环的 spec>
- 该警惕的：<spec 通胀、验证缺口等>

Rules

• Read-only, do not modify any session files
• When quoting user messages, preserve the original text; truncate at most, do not rewrite
• Keep the number of suggestions restrained: at most 3 per category, better to have fewer than to overdo it
• If there are more than 30 sessions in the window, aggregate by project first before analyzing, to avoid expanding item by item