/wrap

A session that produces real learning or a real decision is easy to lose — it lives in chat scrollback and is gone the moment the window closes. This skill's whole job is turning that into one short, durable, append-only log entry, then handing off to whatever else the project itself wants done at session end.

The log is an evidence base, not an activity record. Its point is letting a later review argue "here's what I can actually do unaided" from cited fact instead of vibes. A padded entry is worse than a missing one — it inflates the evidence base with work that proves nothing. A session that produced nothing worth citing gets no entry.

Configure once for your setup

This skill ships with sane defaults; edit the block below (or just tell the agent your own paths the first time you use it) to match how you actually organize things.

LOG:      "./docs/session-log/YYYY-MM.md"     # default: one Markdown file per month, in-repo
CONCEPTS: null                                 # optional: a folder of durable concept notes, if you keep one
DASHBOARD: null                                # optional: a "current skill level" summary, if you keep one — this skill never writes to it

If you already run a mastery-tracking system of your own (a skill-level dashboard, a learning log with its own format), point LOG there instead and skip the default shape below in favor of yours — the mechanism (decide → write → hand off → backup) is what this skill actually provides.

The loop

wrap():
    1. decide whether this session earned an entry
    2. if yes: write it   (ask for the assistance level; never infer it)
    3. hand off to the project's own wrap-up duties
    4. offer a backup, if you have a backup routine configured

1. Does this session earn an entry?

earns_entry(session):
    return any([
        a concept was taught, extended, or newly met,
        you made a design or architectural call worth citing later,
        you caught something — a drift, a wrong premise, a bad assumption,
        a skill area was exercised in a way the log cannot already show,
        work shipped that a future review would want as evidence,
    ])

# NOT sufficient on their own: files were edited, tests passed, a version was bumped,
# a long session happened. Activity is not evidence.

If nothing qualifies, say so in one line — "nothing here the log would want as evidence; skipping the entry" — and go to step 3. That's a normal outcome, not a failure.

2. Write the entry

Append to the current period's log file (creating it if this is its first entry). Append only — never edit a past entry to revise a rating in hindsight; if a rating was wrong, a later entry or a separate review process corrects it, on the record.

## YYYY-MM-DD — <short title>

**Project:** <project>. **Session type:** interactive | autonomous.

**What happened:** <2-5 sentences. Concrete: name the files, the fork, the decision.
Write it for a reader six months out who has none of this context.>

**Concepts exercised:** `<slug>` (new | extended — <the delta>), ... | none.

**Assistance level:** <0-5> — <why that number>. Scale:

| Level | Name | Meaning |
|---|---|---|
| 0 | None | Book, docs, trial-and-error, your own reasoning — no AI involved. |
| 1 | Ask | Asked an AI a question, then did the work yourself. |
| 2 | Copy-paste | Took AI output (a command, a snippet, a paragraph) and adapted it. |
| 3 | IDE-assisted | Inline AI suggestions in the editor while you drove the file. |
| 4 | Agentic, supervised | Handed to an agent harness and watched it work. |
| 5 | Agentic, unattended | The harness ran while you were away. |

**Note for a future review:** <candidate evidence, named as candidate — never a rating
you're asserting as final. A separate review process, if you keep one, does the ruling.>

Ask for the assistance level — don't infer it. Polish is not evidence of independence; a clean result can come from level 1 or level 5, and the whole scale is worthless if a session just guesses. One question:

"What assistance level was this for you — 0 none, 1 asked, 2 copy-paste, 3 IDE-assisted, 4 agentic supervised, 5 agentic unattended?"

Work produced inside the current agent session is level 4 or 5 by definition (supervised if you were watching, unattended if not) and needs no asking. Ask when the session discussed or reviewed something you built elsewhere.

3. Hand off to the project's own wrap-up duties

This skill owns the session log and nothing else. Read the current project's own README/CLAUDE.md/AGENTS.md (whichever it has) for anything else it asks for at session end — a changelog entry, a version bump, a push, a handoff document for the next session — and do that too. If the project says nothing, the log entry was the whole duty.

4. The backup offer

if you_have_a_backup_routine_configured():
    if anything_changed_this_session() or last_backup_older_than_your_own_threshold():
        offer a backup     # you run it; don't automate this step itself
else:
    say nothing     # no backup routine configured is a normal state, not a gap to nag about

Never wire this skill — or the backup it offers — to a scheduled task or cron job. Automated system-level actions run while nobody's watching them fail; a triggered offer that a human approves each time is the safer shape.

Prohibitions

NEVER write a rating into a dashboard/skill-tracker.       # a separate review step does that
NEVER edit a past log entry. Append.
NEVER infer the assistance level from how polished the work looks.
NEVER write an entry for a session that produced no citable evidence.
NEVER pad an optional field to look thorough — omit what wasn't actually observed.
NEVER automate this on a schedule.

Adopting this

Copy SKILL.md into your agent's skills directory (Claude Code: ~/.claude/skills/wrap/; project-scoped: .claude/skills/wrap/), then fill in the Configure once for your setup block. For agents that read a plain instructions file instead of a skills directory, paste the loop and the entry format into that file directly.