Session audit
End-of-session audit that scans for uncaptured corrections, preferences, and decisions, then proposes saving each one to the right file in the user's workspace. Trigger phrase: 'audit this session'. Also activates on 'session audit', 'what did we miss', or 'end of session check'.From its SKILL.md
npx -y skills add Chevis-Zhou/agent-skills --skill session-auditAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
2 things to look at
- 25 days oldThe repository was created 25 days ago. New is not bad, but a brand new repository carrying a familiar-sounding name is the shape a typosquat arrives in, and there has been no time for anyone else to find a problem with it.
- 1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
SKILL.md
4.8 KB, ~1.0k tokens by cl100k_base, as published. Nobody here has run it
Session Audit
A lightweight audit that runs at the end of any Cowork session. It catches things you told Cowork during the session that should be saved permanently so you never have to repeat them.
What This Skill Does
Two things only:
- Scans for uncaptured learnings. Reviews the conversation for corrections you made, preferences you expressed, and decisions you committed to that aren't already written in your workspace files.
- Proposes where to save each finding. For each uncaptured learning, specifies the exact file, section, and wording. You approve or skip each one.
No file reorganization, no cleanup, no progress tracking. Just: "Did I learn anything this session that should be remembered?"
Step 1: Load Relevant Workspace Files
Read whichever of these were touched or relevant during the session. Adapt this table to your own workspace — the pattern is: one file per durable-knowledge domain, each with a clear owner-purpose.
| File | Purpose |
|---|---|
Workspace-level CLAUDE.md | Repo-level rules and routing map |
~/.claude/CLAUDE.md | User-level identity, style, tech stack, ghostwriting voice |
| Voice/style principles doc | Output/voice correction rules |
| Active commitments / PM docs | Hard deadlines, sprints, strategic threads |
| Domain PM docs (career, fitness, sleep, reading…) | Per-domain decisions and protocols |
| Any skill knowledge base files loaded during the session | Client/project-specific rules |
Step 2: Scan the Conversation
Go through the entire conversation from top to bottom. Look for four signal types:
A. Corrections
You fixed, rewrote, or rejected Cowork's output. Each correction reveals a rule for next time.
What to look for: Moments where you edited Cowork's draft or said "no, do it this way." Ask: what underlying preference drove the change?
Example: You replaced "leverage" with "use." → Rule: "Never use the word 'leverage' in output."
B. Explicit Preferences
You stated a preference directly — words like "always," "never," "I prefer," "from now on," "don't do that."
Example: "Don't add a summary at the end." "I like the table format better."
C. Decisions
You chose one option over another, set a deadline, or resolved an ambiguity that affects future work.
Example: "Let's go with TypeScript strict mode." "Skip the fitness audit until June."
D. New Context
You shared a fact about yourself, your work, or your world that Cowork didn't previously know.
Example: "My new gym opens May 1." "The project deadline moved to June 15."
Step 3: Filter Against What's Already Saved
For each finding, check whether it's already captured in the files loaded in Step 1. Skip anything already written. Surface only genuinely new findings.
Step 4: Present Findings
Present each finding in this format:
**[Number]. [What happened]**
- **The rule/fact:** [Exact wording to save, written as a clear instruction or statement]
- **Where it goes:** [File path + section name]
- **Why:** [One sentence on why this matters for future sessions]
Group into two categories:
Recommend (apply unless you object): Clear-cut findings where the right action is obvious.
Your call: Findings where there's a judgment call or where you might want to rephrase.
If nothing new was found: "Clean session. Nothing new to capture." Don't manufacture findings.
Step 5: Routing Guide
Use a routing table like this to send each finding to the correct file (fill in your own destinations):
| Finding type | Route to |
|---|---|
| Output style, tone, formatting, word bans | Voice/style principles doc |
| Writing voice, ghostwriting rules | ~/.claude/CLAUDE.md → voice section |
| Tech stack preference, coding pattern, tool choice | ~/.claude/CLAUDE.md → tech stack section |
| Career decision, sprint change, hard deadline | Active-commitments PM doc |
| Per-domain protocol decision (fitness, sleep, reading…) | That domain's PM doc |
| Broad life decision (quarterly scope) | Quarterly review log (append at top) |
| Repo-level rule | Workspace CLAUDE.md |
| Client-specific preference | Note for manual addition to relevant skill KB file |
Step 6: Apply Approved Changes
After approval, write each approved change to the correct file. Confirm what was written and where.
Never write changes without approval. Always present findings first and wait.
Append-Only Files
Mark any log-style files as append-only — new entries go at the TOP, never edit existing entries (e.g. study logs, quarterly review logs).
What ships with it
Read from the repository
Just SKILL.md. No reference files, no scripts.