agentleFS
rajasdevelrajasdevel/outcomebound/skills/review-council/SKILL.md

review-council

Skill0 starsChanged 15 days agoApache-2.0View on GitHub

Use when a second opinion from other models or harnesses is worthwhile (a design, release candidate or change whose miss would reach users). Plan first; run only after the person's yes.

7 other projects use a version of this.

Adopted by

Most recently taken up by

The file

We split this skill into the rules it gives an agent

3 rules. 0 are common, 1 is distinctive, and 2 are very specific to this project.

  • 0

    Common

    Most of the 7 edited versions we read keep these.

  • 1

    Distinctive

    This author's own choices. Worth a look.

  • 2

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Every reviewer spends the person's credits, so ask whether a second reader would change a decision.

    line 14
  2. Very specific:

    Run outcomebound review council --plan with what to review (--range A..B, --base REF, or --file PATH for a design or a spec), a lens (what …

    line 23
  3. Very specific:

    Run the same command with --accept in place of --plan and with --plan-id set to the id the plan printed, in the background, and wait once …

    line 43

“Common” is measured: kept by at least 6 in 10 of the 7 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.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Review council

You are done when the person saw the plan and said yes before any reviewer ran, each reviewer that answered has a review file, each reviewer that did not is named with its reason, and every finding you kept has one disposition with the evidence for it.

When it is worth asking

Every reviewer spends the person's credits, so ask whether a second reader would change a decision. It would for a design that others will build on, for a release candidate, and for a change that would hurt users if wrong and that no check you can run would catch. It would not for a change a test or the project's gate already covers, a fix to text for people, or a doubt you can settle by reading the code. A reader from another model family adds more than another run of your own, so name reviewers that are not your model.

Plan, and ask

Run outcomebound review council --plan with what to review (--range A..B, --base REF, or --file PATH for a design or a spec), a lens (what the reviewers look at, one lens for a run so that the answers compare), and a --reviewer harness:model[:effort] for each reviewer. With none named, the reviewers are those of the project's or the person's council file, else the engine's built-in default of one reviewer, and the plan says which. Those names are examples, so check them against what the person's logins offer. Name more reviewers when the question needs them. The plan runs nothing.

Show the person the plan as printed, so that their yes is to the argv, the timeout and the cap and not to your summary of them. Then say in plain words who reviews, what they read, that it spends their credits, any program the plan says is not found, which login each Claude reviewer bills, any reviewer the plan marks as read-only not confirmed (one is run only if the person names it), and which safeguards no run has shown. Then ask for a plain yes. Only that yes lets you run --accept. A yes to an earlier plan, or to reviews in general, is not a yes to this plan: a changed reviewer, range, lens or cap is a new plan. A failed reviewer you want to run again is a new plan too. If you get no yes, you stop at the plan. You never start codex, claude -p or another reviewer by hand, because that skips the safeguards the plan lists.

Run

Run the same command with --accept in place of --plan and with --plan-id set to the id the plan printed, in the background, and wait once for it to end. It runs the reviewers in parallel, read-only, each on a snapshot of the reviewed revision. A run that the plan id refuses differs from the plan the person approved, so plan again and ask again. A reviewer that fails or times out is reported and the rest go on; do not read a missing file as a clean review. Report to the person which reviewers answered, which did not and why, and the spend the run printed, which is an estimate.

Merge and dispose

Open each review file: the run prints each reviewer's verdict and finding count, and a verdict of pass with a body that says the reviewer could not read the code is a failed review. Take the findings with the review-findings skill. Where reviewers raise one cause, keep it once and note who raised it; agreement is no proof and one reviewer's finding stands or falls on its evidence. A Citation check: line under a finding is the engine's: a place the finding names is not in the reviewed revision, is ambiguous, or the finding names none. It shows that the finding needs your reading, and does not show it wrong. Write the dispositions into the review files and run outcomebound review check on each. A reviewer reading again after your fix is a new plan.

Download now (ZIP)

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
What did you use it for? e.g. Reviewing pull requests in a Rails app

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.