Wrap

Wrap

End-of-session wrap-up. Primary goal: improve future agent performance.

Step 1: Extract gotchas

Review this conversation for GOTCHAS ONLY. A gotcha is:

1. Something that broke -- caused a bug, wasted time, or required a fix 2. A wrong assumption -- had to be corrected mid-session 3. Counterintuitive behavior -- an API, tool, or system that doesn't work how you'd expect

Do NOT record:

• What was shipped or deployed (git log has this)
• Feature descriptions (agents.md has this)
• Rollout timelines or monitoring results
• UI polish details
• Anything derivable from reading the code

Format: one line per gotcha. Code example only if the fix is non-obvious.

Step 2: Save to project memory (PRIMARY)

Update auto-memory MEMORY.md (~/.claude/projects/<current-project>/memory/):

Add new entries:

• Gotchas and corrections discovered this session
• User preferences or workflow corrections ("don't do X", "I prefer Y", "stop doing Z"). These go in feedback type memory files. Include WHY the user wants it that way
• Non-obvious project context learned this session

Clean stale entries:

• Scan MEMORY.md for bugs/gotchas that were FIXED this session. Remove or mark resolved
• Update entries that are now inaccurate based on work done this session
• Check if any "Active Projects" items were completed

Keep entries concise (1-2 lines each). Do NOT duplicate what's already there.

This is the most important step. Memory persists across all future sessions.

Step 3: Update project files (if they exist)

Check for these files and update if present. Skip silently if missing.

agent-learnings.md: Append new gotchas under the right section (Code traps, Deploy flow, Infrastructure, Testing). Skip duplicates.

globalcontext.md: Update version + recent releases table if version was bumped. Update "Active work" if work was completed. Update "Known issues" if new issues found or old ones resolved.

.plans/next-steps.md: Update "Current state" header if version was bumped. Update "Known issues" if issues resolved or new ones found.

.plans/INDEX.md: Move completed plans to "Recently shipped". Add new plans if created during session. Keep "Recently shipped" to last ~5 entries.

Step 4: Generate handover prompt

Build a continuation prompt that a fresh agent can use to pick up where this session left off. The prompt must be self-contained (the next agent has zero context).

Structure:

Branch: {branch} (PR #{number} if applicable)

What shipped this session: {brief list of changes, with filenames}

Status: {built locally / pushed / deployed}. {test results if run}. {typecheck status}.

What still needs doing:
1. {next step}
2. {next step}
...

Known gaps / follow-up:
- {anything deferred or incomplete}

Rules:

• Include file paths for anything created or significantly changed
• State the git status (committed? uncommitted? pushed?)
• If there's a plan file, reference it
• Keep it factual, no fluff. The prompt should be copy-pasteable
• If nothing is in progress (pure Q&A session), skip this step

Step 5: Confirm and prompt reset

1. Show brief summary of what was saved (or "nothing new" if nothing qualified) 2. Display the handover prompt in a fenced code block so user can copy it 3. Try to copy the handover prompt to clipboard: echo '...' | pbcopy 4. Tell user: "Handover prompt copied to clipboard. Paste it after /clear to continue."

• If pbcopy fails: "Copy the prompt above, then run /clear to reset context."

<div align="center"> <br> <img width="137" height="137" src="icon.svg" alt="Wrap"> <br> <br> <h1>wrap</h1> <p>End-of-session wrap-up for <a href="https://github.com/anthropics/claude-code">Claude Code</a>. Extract gotchas, save learnings, reset with confidence.</p> <br> <br> </div>

Most session logs are noise. "Shipped v3.57.4" is in git. "CSS is 18KB now" is in the code. The only thing worth persisting is what went wrong and what the user prefers. because those are invisible to future agents reading the codebase.

Install

skills.sh

npx skills add github:bitbonsai/wrap

You can use skills.sh to add this skill globally.

Claude Code

claude skill install github:bitbonsai/wrap

pi.dev

pi install git:github.com/bitbonsai/wrap

OpenCode

OpenCode is compatible with Claude skills. Clone into .claude/skills/:

git clone https://github.com/bitbonsai/wrap.git .claude/skills/wrap

Usage

Say any of:

• "wrap"
• "wrap up"
• "let's wrap"
• "save learnings"
• "end of session"
• "I'm done for now"

What it does

1. Extracts gotchas: things that broke, wrong assumptions, counterintuitive behavior 2. Saves to project memory: updates MEMORY.md with gotchas, user preferences, and removes stale entries 3. Updates project files: if agent-learnings.md, globalcontext.md, or plan files exist, keeps them current 4. Generates handover prompt: a self-contained continuation prompt a fresh agent can use to pick up where you left off 5. Copies to clipboard: paste it after /clear to continue seamlessly

What it does NOT record

• What was shipped (git log has this)
• Feature descriptions (docs have this)
• Rollout timelines
• Anything derivable from reading the code

Recommended project setup

The skill works with any repo, but it's designed to augment a few specific files. It updates all of these if present, skips silently if not.

agent-learnings.md

A running list of gotchas organized by category. Think of it as your project's "lessons learned" log.

## Code traps
- Don't use `find -delete` — it deletes directories matching the pattern too
- `bun run` doesn't pass env vars the same way as `npm run`

## Deploy flow
- Always run `terraform plan` before `apply` — the staging state drifts fast

## Infra
- The ECS cluster name is `prod-web`, not `production-web`

Why it helps: Future agents don't have to relearn the hard way. A single gotcha can save hours of debugging the same trap twice.

globalcontext.md

A living orientation snapshot. One page that answers "what is this thing and how do I work on it?"

# MyProject

- **Stack:** Next.js 14, Tailwind, PostgreSQL, Vercel
- **Node version:** 20.x
- **Key env vars:** `DATABASE_URL`, `OPENAI_API_KEY`
- **Dev server:** `bun run dev` (port 3000)
- **Active work:** Migrating auth from JWT to OAuth2
- **Last deploy:** 2026-04-19

Why it helps: When an agent joins a session cold (especially after /clear), it has zero context. This file is a cheat sheet that prevents the "what framework is this again?" warm-up loop.

.plans/INDEX.md

A lightweight project tracker with three sections:

## Active
- [ ] Add dark mode toggle (waiting on design tokens)
- [ ] Refactor webhook handlers to use Bull queues
- [ ] Fix flaky e2e test on `/dashboard/billing`

## Planned
- [ ] Migrate from Stripe Checkout to Payment Elements
- [ ] Add real-time collaboration via Y.js
- [ ] Evaluate Vercel Edge vs Node runtime for API routes

## Recently shipped
- [x] v2.4.0 — Team billing + seat management
- [x] Migrated auth from JWT to OAuth2 (GitHub + Google)
- [x] Switched test runner from Jest to Vitest

Why it helps: Agents constantly lose track of what's in progress versus what's done. A simple index prevents "did we already do this?" and keeps work organized without a full project management tool.