Wrapup

---
name: wrapup
description: Generate a concise agent-work summary covering all sessions since the last git commit — what was asked for, how it was driven, and which skills were used. Suitable for pasting into PR descriptions or commit messages. Use when the user invokes /wrapup or asks to "summarize what the agent did" / "give me a PR summary" / "what happened this session".
---

# Wrapup: agent-work summary for PRs and changelogs

Produces a short markdown block the user can paste into a PR description, commit message, or handoff note. The goal is to capture *how* the change was made — what was asked for and which tools drove it — not to restate *what* changed (that's what `git diff` is for).

## Scope

Cover all Claude Code sessions with events timestamped **after the most recent git commit** in the current repo:

```bash
git log -1 --format="%aI" 2>/dev/null || echo "1970-01-01T00:00:00+00:00"
```

If there are no commits yet, cover all sessions for the current project. **Always include the live (current) session** — `/wrapup` is typically run just before the user commits or opens a PR.

Note: the `/wrapup` invocation prompts themselves will not appear in the output, because those events have not been written to the JSONL yet when you read it.

## Finding relevant sessions

See the `claude-session-transcript` skill for how to locate transcript files. Steps:

1. Get the last-commit ISO timestamp from git (above).
2. List all transcripts for the current project.
3. Include a transcript if its file mtime is after the commit timestamp — this is a reliable approximation. The live session is always included.

## What to extract

Read the `claude-session-transcript` skill for prompt-extraction rules before parsing transcripts. From each in-scope session, extract:

### 1. Real user prompts — phase bullets with inline quotes

Apply the full real-prompt filter from `claude-session-transcript`. Then **summarize** the prompts as phase bullets rather than listing them individually:

- Identify the major phases of work across all in-scope sessions (e.g. "initial setup", "implementation", "hardware testing", "course correction").
- Write one bullet per phase. Lead with an italicized phase label, then quote the most intent-revealing user prompt for that phase in quotes, followed by a dash and a brief description of what the agent did. Example: `- *Hardware testing:* "Let's try it on the real device" — agent ran the capture script; hit a USB permission error and fixed it.`
- If a phase involved a notable course correction or redirect, include the correcting prompt as a quote too: `User corrected tool installation mid-session ("I prefer uv now over pipx").`
- Drop noise prompts entirely: single-word continuations ("ok", "continue", "run"), bare acknowledgments, and anything fully captured by the phase description.
- Do not include the `/wrapup` invocation prompt itself.

### 2. Skill invocations

In each transcript, find `type == "assistant"` events that contain a `tool_use` block with `name == "Skill"`. Extract `input.skill` — that is the invoked skill name.

Classify each as **user-configured** or **built-in**:

- **User-configured**: skills the user has added or installed — anything not shipped as part of the agent's standard installation. Always mention these if detected, even if the agent invoked them autonomously rather than the user typing a slash command.
- **Built-in**: skills that ship with the agent. For Claude Code, current examples: `init`, `code-review` (with `/review` as an alias), `security-review`, `simplify`, `loop`, `schedule` — the set grows with releases, so treat any list as examples, not an inventory. When unsure whether a skill is built-in or user-configured, check whether it appears in the user's own skills or plugin directories. Mention built-in skills **only** if the user explicitly invoked them (i.e. there is a corresponding slash-command stub in the transcript).

If the same skill is invoked multiple times, list it once.

### 3. Session metadata

- Number of in-scope sessions
- Date range (YYYY-MM-DD of first and last event, or just one date if single-day)

## Output format

Print a markdown block to the conversation. Do not write a file unless the user asks.

```markdown
## Agent-assisted changes

**Sessions:** <N> · <date> [– <end date if multi-day>]

### How it was built
- *<Phase label>:* "<key user prompt>" — <what the agent did. Course corrections quoted inline if present.>
- *<Phase label>:* "<key user prompt>" — <what the agent did.>

### Skills used
- **`/<skill>`** *(user-configured)* — <what it did in this session, one phrase>
- `/<skill>` *(built-in, explicitly invoked)* — <one phrase>

> _Generated by /wrapup from Claude Code session transcripts._
```

Omit the **Skills used** section entirely if no qualifying skills were invoked.

## Style notes

- **Phase bullets over list.** One bullet per phase of work, with the key prompt quoted inline, is more readable in a PR description than a numbered dump of 20 prompts.
- **Skills: what they did here, not what they are.** "Captured USB traffic from the target device" beats "USB traffic capture skill."
- **This is not a teaching document.** No teaching notes, student takeaways, or prompting-technique analysis — that's `/teach`.
- **Short is right.** This is meant to fit in a PR description alongside other content.