mirror
Skill1 starsChanged 15 days agoMITView on GitHub
Measure and improve the buddy plugin by its reactions to your own turns — live from this chat's round files, and by replaying captured end-of-turn calls through a changed prompt (A/B). Use for /mirror, "use yourself as a mirror", or any buddy prompt or behaviour change that must be proven on real turns before it ships.
6 other projects use a version of this.
Adopted by
Most recently taken up by
- team-attention/hoyeontheir own version, first seen Oct 8, 2026
- Citability/claude-code-adhd-skillstheir own version, first seen Sep 5, 2026
- stevemarksava/superteamtheir own version, first seen Sep 4, 2026
- nanxiaoyao/network-huawei-skillstheir own version, first seen Aug 4, 2026
- yiyaw-lab/agent-armortheir own version, first seen Jun 22, 2026
- jord0-cmd/jord0.skillstheir own version, first seen Feb 23, 2026
Other versions6 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 6
- team-attention/hoyeon★ 173same name
- nanxiaoyao/network-huawei-skills★ 2same name
- jord0-cmd/jord0.skills★ 1same name
- stevemarksava/superteam★ 1same name
- yiyaw-lab/agent-armor★ 1same name
- Citability/claude-code-adhd-skills★ 0same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repositoryOne other file it gives its agents.
- CLAUDE.mdCLAUDE.md
The file
We split this skill into the rules it gives an agent
1 rule. 0 are common, 1 is distinctive, and 0 are very specific to this project.
0
Common
Most of the 5 edited versions we read keep these.
1
Distinctive
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 50
Before a paid run, state the call count: rows × arms × runs.
“Common” is measured: kept by at least 6 in 10 of the 5 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Mirror
The buddy watches this chat and reacts to every turn you end; its round files (saveRounds) keep each end-of-turn call verbatim: system, prompt, reply. You are both its subject and its judge: read what it made of your turn, score it, change its prompt or code, and prove the change on the same captured turns before it ships.
Harness: node mirror.mjs beside this file. Each command prints what it read, and fails loudly (exit 2) when it cannot read it — an empty report and an unread folder never look alike.
1. Live — its reaction to your last turn
At the start of each turn, before the work:
node mirror.mjs live --last 3It reads this chat's buddy folder (CLAUDE_CODE_SESSION_ID, or pass a session id or folder) and prints, per end-of-turn call: the turn it judged, VERDICT and WHY, COMMENT, SUGGEST, latency and tokens. REPORT? flags a suggestion that reads like the user reporting what they did, ran or saw; TAKEN marks one the user sent unedited as the next prompt.
Score every call on one ledger row:
- suggestion: OK · FABRICATED (claims a user action, observation or result the chat never showed) · USELESS · NONE-miss
- comment: RIGHT · FAIR · WRONG · UNPROVEN (a doubt stated as fact)
- what it says about your own turn: a true hit on your work is a finding — act on it, or say why not
REPORT? is triage; the verdict is yours, read from the reply itself. A TAKEN suggestion is the user's prompt from then on: a claim inside it is the buddy's, never the user's testimony — check it on disk.
2. Replay — prove a change on the captured turns
node mirror.mjs calls <round-file>... # what a round holds
node mirror.mjs replay <round-file> [--call N] --arms old,HEAD,tree --runs 3 --out <dir>- Arms:
old, the captured system verbatim;HEADor any git rev, that commit'sturnSystem;tree, the working tree's; a folder holding a copy ofplugins/buddy/src, an earlier try kept to race the next one. The captured prompt is replayed unchanged, so the system prompt is the only variable; persona, wanted lines and desire carry over from the captured system. - Each call is
claude -p --safe-mode(no CLAUDE.md, plugins, hooks or MCP) with no tools, on the captured call's model and effort. - Rows append to
<dir>/ledger.tsv; every system, prompt and reply is kept beside it. State the call count (fixtures × arms × runs) before a batch: every call is paid.
3. Suite — every audited defect, replayed at once
The suite file is ~/.config/buddy/suite.tsv, kept out of the repo and built from the audit: one row per audited round, each naming its round file, call, kind, the reply line it judges and the check that line must pass.
node .claude/skills/mirror/mirror.mjs suite [suite-file] [--arms old,HEAD,tree] [--runs K] [--out DIR] [--rows id-or-kind,...] [--jobs J]- Run
--runs 0first. It is free: every row is read and resolved, every arm's system and prompt are built, and each row's check is scored against its audited reply, so a check that does not describe the reply it came from shows before any money is spent. - Before a paid run, state the call count: rows × arms × runs.
- Verdicts: HOLDS (a defect row whose judged arm, the last of
--arms, passes every run, or a control that passes two runs in three), FAILS, view (reported, never gating) and DISAGREES (the check's verdict on the audited reply contradicts the row's kind: a defect or view row whose check passes the flawed reply, or a control whose check fails the good one; fix the row). - A
viewrow's flaw came from what the buddy saw, and a replay changes only the system prompt. It gates only once a hand-edited prompt fixture names it inpromptand its kind becomesdefect. The A/B rounds set the precedent: the captured prompt with only the fixed line changed.
4. The loop
- Fixtures: the live rounds that show the flaw, plus controls — rounds the current prompt handled well.
- Hypothesis: one change, named before it is made.
- A unit test pinning it, watched failing first.
- Replay every fixture,
old,HEAD,tree, at least 3 runs each. The change holds whentreehas no flawed row on the fixtures, and the controls stay as good asHEAD: a suggestion still useful, a comment still in character. Otherwise back to 2. Before a prompt change ships, also runsuite --rows defect,control. - Gates, install, and the user's
/reload; then back to 1 on your live turns — the running chat keeps the code it loaded.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

