.claude/skills/core/wrap-up/SKILL.md
Session-end protocol — collect learnings, update context, commit session state
npx skillsauth add andrem-sec/psc-comet wrap-upInstall this skill globally with one command. Works with Claude Code, Cursor, and Windsurf.
3 of 9 scanners reported clean
Some scanners were skipped, did not run, or reported a non-clean status. Review each row below.
Run at the end of every session. Preserve knowledge before the context window closes.
Without wrap-up, session learnings evaporate. The next session starts cold. Patterns that were discovered at cost get rediscovered. Decisions that were made get revisited without their rationale. The system does not improve.
Before writing any learning to learnings.md, it must pass all three:
Reject generic programming patterns, library usage examples, refactoring techniques, and anything a junior dev could find in the docs.
[YYYY-MM-DD] [category] — [concise, specific statement]
Categories: pattern | mistake | approach | tool | decision
Good entry: [2026-03-25] mistake — Wrote settings.json hooks as escaped Python one-liners; they silently fail on edge cases. Move hooks to .claude/hooks/ shell scripts.
Bad entry: [2026-03-25] pattern — Always write tests before implementation. (too generic, fails gate)
Done: [what was completed]
Learned: [the one most valuable insight]
Next: [specific next action]
After collecting learnings and updating context files, write a session memory snapshot to context/session-memory.md. This file is the primary input for /resume at the next session start.
Overwrite the file each session (it tracks current state, not history).
## Current State
[What is true right now — not what was done, but what IS.
Current branch, file states, whether tests pass, open PRs, blockers.]
## Task
[The task that was active at session end and its acceptance criteria.
If multiple tasks, list all with their status.]
## Files
[Files touched this session, their roles, and their current state.
Format: path — role — state (clean/modified/broken)]
## Workflow
[The sequence of steps taken and their outcomes.
Ordered list. Mark completed steps. Note last completed step explicitly.]
## Errors
[Errors encountered this session and how they were resolved.
Unresolved errors: document what was tried and what failed.]
## Learnings
[Non-obvious things discovered this session that would affect future work.
Only include things that pass the 3-point quality gate.]
## Key Results
[Concrete outputs: files created, tests passing count, PRs opened, decisions made.]
## Worklog
[Timestamp-ordered list of major actions taken. Use approximate times if exact times unknown.
Format: ~HH:MM — [action]]
Write this file before committing. The resume skill reads it to reconstruct the session without re-reading the conversation.
Before committing, run git branch --show-current to identify the current branch.
If on any branch other than main: commit normally — no prompt needed.
If on main: pause and present this choice to the user:
You are on main. Where should I commit this session's state?
1. Commit to main anyway (solo project / I know what I'm doing)
2. Commit to an existing branch (I'll pick from the list)
3. Create a new branch (I'll name it now)
4. Skip the commit (I'll handle it myself)
If they choose 2: run git branch to list local branches, let the user pick, then git checkout <branch> and commit there.
If they choose 3: ask for a branch name, run git checkout -b <name>, and commit there.
If they choose 4: skip the commit step entirely. Note it in the session summary under Next.
Never commit without completing this check when on main.
chore: wrap-up [date] — [1-line session description]
Do not manufacture learnings to fill the section. If nothing new was learned, write: [date] pattern — [existing approach confirmed, no new findings]
Do not append learnings that fail the quality gate. Quantity is not the goal.
data-ai
Parallel agent swarm — decomposes work into independent units, spawns isolated workers, tracks PRs via fan-in
testing
Audit animations and transitions for motion accessibility, performance safety, and design intent. Enforces prefers-reduced-motion compliance and blocks layout-triggering transitions.
testing
Test specifically for AI-introduced regressions that repeat without tests
development
Framework for decomposing agent-driven tasks into independently verifiable units