Coding Session Recap — See What Is Done, Unverified, and Next

A read-only AI coding agent session recap skill that reviews the current conversation and refuses to let "I wrote the code" pass as "the code works".

Quick answers

• What does conclude-rounds do? It recaps the most recent conversation rounds and labels each result as done-and-verified, done-but-unproven, or proposed.
• How many rounds does it recap? You can request any count; the default is 4 user-and-agent exchanges.
• Does it run tests or prove new claims? No. It reports evidence already shown in the conversation and may briefly inspect the workspace when a claim's state is unclear.
• Does it read old session transcripts? No. It works only from the current conversation context, including any visible compaction summary.
• Does it change files? No. The skill is read-only.

What it does

When a long back-and-forth gets hard to track, this skill re-reads the last N exchanges (default 4) and gives you a tight, honest recap: what you asked, what was actually done, and what's still open. Its one non-negotiable rule: every item is sorted into done-and-verified, done-but-unproven, or merely proposed — because an agent's confident claim and a passing test are not the same thing. It then mines those same rounds for up to 5 workflow insights, each tied to a concrete mechanism in the engine you're using (a hook, an instruction-file rule, a reusable prompt) with a one-line first step. If the rounds only support 2 real insights, you get 2.

It runs entirely in-context and is read-only: no files edited, no transcripts pulled from disk, no cross-session digest.

When it fires

• "Recap the last few rounds"
• "What did we just do?"
• "Catch me up on this thread"
• "Wait — what actually got done vs just discussed?"
• "总结前面4轮"
• "这几轮做了啥？"

Install

npx skills add m1nga/conclude-rounds

Try it on a mixed-status conversation

After installing, ask: Use $conclude-rounds to recap the last 3 rounds. Keep verified work, untested work, and proposals separate.

Illustrative fixture: assume the visible conversation contains these facts.

Visible evidence	Expected recap
Auth refactor followed by a test command that exited successfully	Refactored and verified by the shown test run
Rate limiter written; no execution or test output shown	Written, behavior unverified
Two migration approaches discussed; neither selected	Proposed, decision still open

The skill summarizes this evidence; it does not run tests to fill the gap. If only two real user exchanges remain visible, it says it can recap two instead of inventing a third. Tool notifications are not extra conversation rounds. If the history supports no useful workflow insight, none needs to be added.

If this helps you distinguish finished work from a confident claim, a star on this repository is welcome.

Works well with

• iteration-close — this skill closes a stretch of conversation; that one closes a whole product iteration.

Design notes

This skill comes from a solo builder's recurring failure mode: after 20 rounds of agent-assisted work, the transcript said everything was done — and half of it wasn't. Code had been written but never run; decisions had been "agreed" that were actually still open. The three-state sort (verified / unproven / proposed) exists because the cost of a false "done" is a rebuild days later.

The insights section is deliberately strict: every insight must cite evidence from the rounds just recapped, name a mechanism you can actually configure, and end with a first step. Generic advice ("consider using subagents") is banned by the skill text itself, and it is told to deliver fewer than 5 rather than pad — and never to repeat an insight it already gave earlier in the session.

Field-tested

Probed 7 scenarios across 5 personas · 4 fired correctly · 2 correctly stayed quiet · 1 flagged as a coin-flip (fix queued).

"wait, catch me up — what did we actually finish vs just talk about?" → Fired. Recap sorted every item into verified / unproven / proposed — the rate-limiter that was written but never run stayed ⚠️, not ✅.

"recap what we did across yesterday's session and today" → Correctly stayed quiet. Cross-session digests are explicitly out of scope; this skill only re-reads the conversation you're in.

"这个迭代做完了，总结一下" → Correctly yielded to iteration-close. Closing a whole iteration is its sibling's job; this skill recaps a stretch of conversation, nothing more.

Probe method: scenario-probe

Author

Built by Ming. The design notes above explain the real problem and tradeoffs that shaped this skill.