evidence-bundle
Skill1 starsChanged 33 days agoMITView on GitHub
What an agent proves before the owner sees its PR or outward draft, and who decides next. Tests on the pushed head, before/after renders, every claim checked against its source, a fresh Sonnet refuter, cross-review (on the sandbox, from a cloud thread through Remote Control), gathered in one PR comment with a trust level named for who acts: Do (the agent merges), Pick (the owner picks on one card, work goes on with the default) or Yours (the owner, or the agent right after his go). Use when finishing any change or draft that reports to the owner, in a Claude project thread or any session, and before merging without the owner.
1 other project uses a version of this.
Adopted by
Most recently taken up by
- manhds82/claude-code-anyllmtheir own version, first seen Oct 13, 2026
Other versions1 project uses a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 1
- manhds82/claude-code-anyllm★ 16same 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 repository18 other files it gives its agents.
- AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- blast-radiusSkill
- claude-remote-control-serverSkill
- correctSkill
- councilSkill
- cross-reviewSkill
- docs-reviewSkill
- herdrSkill
- intent-cardSkill
- one-termSkill
- owner-charterSkill
- quotaSkill
- starbridgeSkill
- test-auditSkill
- unnarrateSkill
- unslopSkill
- verifySkill
The file
We split this skill into the rules it gives an agent
2 rules. About 0 are common, 1 is distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈1
Distinctive · estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 14
Push first: Every proof runs on the pushed head SHA.
- Very specific: line 16
Author's proof: Write the PR's Blast radius section (blast-radius).
≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Evidence bundle
The owner reads the bundle, not the diff. It runs blast-radius (author),
verify (second agent) and cross-review (other vendor), and adds one
summary, a trust level, and how a cloud thread gets each piece.
Happy path
- Push first. Every proof runs on the pushed head SHA. A proof on local work or an older head counts for nothing.
- Author's proof. Write the PR's Blast radius section (
blast-radius). Take each claim in the PR body to "ran it" where it can run: a bug fix's test fails on the base and passes on the head; anything that renders gets before and after at the same viewport, light and dark; motion gets a recording. Open every image and say what it shows. - Claims check. For prose, posts and outward items, list every number,
name, date, link and quote with the source you took it from. A Haiku
subagent gets only that list and opens each source; it returns, per claim,
matches or what the source says instead. Name each source by repo, ref and
path: one it can't open comes back "no source", which reads like a gap.
Anything public also names no private repo of the owner's: no link, no
repo #12reference, branch, file path or artifact built from one. - Fresh refuter. A Sonnet subagent (never Fable) with the
verifyskill gets the repo, PR number, head SHA and a checkout at that SHA, and the instruction to find what's wrong. Nothing else: no transcript, no summary of why the change is right. In the cloud, tell it to read the PR with the GitHub MCP tools orgh api, sincegh prfails there. Recheck each finding yourself before acting on it. Fix what holds and push; a second refuter round isn't needed. A finding you can't fix goes to the owner as Pick. - Cross-review. Skip it only where
cross-review§1 allows, and say so. On the sandbox, run it yourself. From a project thread, start one Remote Control session per batch (start_remote_control, not one per PR; it answers a message of the owner's, so it can't start unprompted) in the one folder the owner approved for all sandbox sessions,/home/dev/work/T0mSIlver, andcdfrom there; never ask for another folder. Tell it, per PR: check out the head SHA, write a 5–15 line brief to a file (task, constraints, claim, how it was verified), runcross-review --brief FILE, and post the findings as one PR comment; for a merge-by-default PR, alsocross-review --proof <n>. Not run when it was due: a "Not verified" line, not a reason to ask. - Post the bundle as one PR comment, edited on every push, then act on its trust level. The thread's reply to the owner is the comment's first four lines and a link.
<!-- evidence-bundle -->
## Evidence at <short head SHA>
**Claim:** <what is now true, one sentence>
**Trust:** Do, merging | Pick, your call (default: …) | Yours, yours to do. <why this level>
**Confidence:** high | medium | low, resting on <the one fact>
**Not verified:** <each gap and what would close it>
| Claim | Evidence (run link, file:line, image) | Got to |
|---|---|---|
**Refuter** (Sonnet, fresh, at <sha>): VERIFIED | NOT VERIFIED | INCONCLUSIVE. <findings: fixed, or rejected because>
**Cross-review:** <vendor>, N findings, fixed X, rejected Y because … | skipped: <rule> | not run: <why>
**Claims checked:** N of N against their sources <link to the table>
**You decide:** <one question, with the default> (Pick and Yours only)Trust levels
owner-charter "Who acts" sets the level: Do (the agent merges, then
one line), Pick (one card with a default, work goes on meanwhile) or
Yours (the owner, or the agent right after his own yes naming the item, in any words).
The repo's AGENTS.md and CONTRIBUTING.md win (the fast-lane repos,
vidtheque and working-set, push to main and are reviewed nightly). Do
merges once every check is green at the head SHA and the bundle is posted;
edit the bundle to "Do, merged" after.
Work the owner didn't ask for goes up as a draft PR labelled proposal, with
its bundle and a 30-second demo. It never merges without him. Close it
unanswered after 7 days.
Gotchas
- A refuter who read your reasoning agrees with it. Pass the PR and SHA, never the summary. Same for the claims checker: the source, not your notes.
- "Not verified: none" from a cloud thread is almost always false. No Mac, no device, no sandbox: name each one the change needed.
- Green as a draft is not green. Some lanes skip drafts (localvoxtral
mac-lanes); count each check's state on the head SHA. - A push makes the bundle stale. Merge only when its SHA is the head's.
- Green alone isn't green with main. Two PRs green on their own heads have broken main together. Before merging, check what landed on main after the PR's CI run; if it touches the same code, merge main in and push. A rerun reuses the old merge ref, so push instead of rerunning.
- Remote Control turns off the thread's folder access and shows the owner an approval prompt. Ask once a night, never on Sunday, and keep working; no answer leaves cross-review as a "Not verified" line, never a reason to wait.
- Haiku misreads long sources. Give it the page section, not the site; a mismatch it reports gets rechecked by you before you change the text.
Not possible
- Cross-review inside a cloud container: opencode and Codex live on the sandbox.
- Proving a visual change without a render.
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.

