A Claude skill. One clear GO / NO-GO, backed by evidence, not optimism.
You finish a session and go to archive it. But once it's archived, whatever lived only in that conversation is gone — a root cause you figured out, a decision you made, a file you never committed. ready-to-archive is the check you run first. It reads the repo, your memory, and git — and tells you, plainly, whether it's safe.
It was built for one reason: a session once got archived with things unrecorded, because
the check trusted the conversation instead of the repo. This skill doesn't. Every "yes, that's
saved" points at a file it actually read or a git command it actually ran — and anything it
can't verify is reported as a gap, never waved through.
- Install — download the latest release zip and add it through Claude's Upload skill dialog (in Claude's settings, under Capabilities → Skills).
- Ask, before you archive:
"Ready to archive?" · "Can I safely archive this?" · "Is everything recorded?"
You get back a verdict — ✅ SAFE,
It runs a grounded three-pillar audit, verifying each against git and the filesystem — not the chat recollection:
| Pillar | The question | Where it looks |
|---|---|---|
| 📝 Recorded | Does every decision, conclusion, and consequential change live somewhere durable — not only in this chat? | decision log / ADRs, spec STATUS, inbox, README / CLAUDE.md, your memory files |
| 💾 Committed | Is the git working tree clean — nothing modified, staged, or untracked-of-consequence — and pushed if it must live remotely? | git status, upstream, stashes |
| 🚀 Landed | Did branch work reach main where it had to, or is the open PR named? | branch vs. main, remote, open PRs |
The verdict is advisory, and the skill acts only when you ask it to:
- It never commits, pushes, merges, or opens a PR without your explicit go-ahead.
- When there are gaps, it offers the shortest path to GO — and, if you say yes, hands off to whatever owns the work (e.g. merging the branch, saving a learning to memory).
- A narrow "yes, fix that one gap" is never a blanket yes to commit-and-push everything.
A false GO loses work; a false NOT YET just costs a second look. So when it's torn, it picks NOT YET and names the doubt. When it genuinely can't reconstruct what happened this session (context was summarized), it says so and asks you to glance — rather than papering over it with a confident green light.