agentleFS
moiz-73moiz-73/glassbox/skills/adversarial-review/SKILL.md

adversarial-review

Skill0 starsChanged 44 days agoMITView on GitHub

Run an independent, checkable review before work ships. Use when an artifact, document, skill, or change is about to be published or merged; when reviewing your own work adversarially before others see it; when findings need adjudication into fixes or written rulings; when a fix loop needs a bound and a breaker; or when running a publication checklist over text that could leak specifics. Three lenses: security and leakage, correctness and adversary simulation, method and clarity. Not for producing evidence about behavior (`evidence-protocol`) or reconnaissance.

111 other projects use a version of this.

Adopted by

Most recently taken up by

Used with

Projects with it also use rethink, root-cause-analysis and 10 more.

Other files people use alongside this one.

Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
  • fabioespindulaSkill

    rethink

    Reframe a plan, idea, or decision before the user commits to it. Use when the user asks to rethink, step back, pressure-test, gut-check, decide, right-size, simplify, think bigger about, or evaluate a concrete plan or decision in product, strategy, technical architecture, career, life, company direction, or personal priorities. This skill improves decision quality before action: clarify the real question, challenge assumptions, compare alternatives, calibrate ambition, classify the operating mode, surface risks, and recommend a next move. Do not use for broad future exploration without a concrete decision; that belongs to a think-big or research-style skill.

    Also in 3 projects that use this

  • MicrosoftSkill

    root-cause-analysis

    Use when a Regional Retail Manager asks I'm seeing unusual inventory movement at our Northwest stores. Can you help me understand what's happening?

    Also in 3 projects that use this

  • anasahmed07Skill

    better-auth-best-practices

    Skill for integrating Better Auth - the comprehensive TypeScript authentication framework.

    Also in 4 projects that use this

  • DockerSkill

    fix-issue

    Fix a single GitHub issue end-to-end: triage, research, write the fix, review, and create a PR. Use when asked to fix an issue: "fix issue 1234", "resolve #500", "create a PR for issue 200".

    Also in 3 projects that use this

  • MicrosoftSkill

    pull-request

    Write concise, focused pull request descriptions for Wassette — at most three sentences on the what and why, with an explicit note on any breaking public-API change and how users adapt. Use when opening or editing a Wassette pull request.

    Also in 3 projects that use this

  • juriwiensSkill

    codex-review

    Run code review via Codex CLI

    Also in 3 projects that use this

  • +2 more from microsoft

The file

We split this skill into the rules it gives an agent

7 rules. 0 are common, 7 are distinctive, and 0 are very specific to this project.

  • 0

    Common

    Most of the 100 edited versions we read keep these.

  • 7

    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
  1. Distinctive:

    Review the delta, not the world: After a fix, re-review the fix's diff in a fresh context.

    line 61
  2. Distinctive:

    Use a separate context for every judgment, review, and final gate.

    line 101
  3. Distinctive:

    When no external judge exists, or before one sees the work, run this drill: write a one-page threat model of your own deliverable.

    line 110
  4. Distinctive:

    Every finding ends with exactly one of two rulings:

    line 133
  5. Distinctive:

    Before anything ships: scan for secret-shaped patterns and for person-identifying or target-specific content.

    line 146
  6. Distinctive:

    Read only the description and predict everything the artifact does.

    line 157
  7. Distinctive:

    Exit gate: All three lenses have verdicts.

    line 163

“Common” is measured: kept by at least 6 in 10 of the 100 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

Adversarial Review

Review is not a reading; it is an adversarial simulation with a verdict. The author knows what they meant, so the author is the worst available judge of what the artifact actually says. This skill defines the pre-ship protocol: separate read-only judges, three defined lenses, checkable rubrics, bounded fix loops with a breaker, and adjudication that never silently discards a finding.

The principle that drives the whole protocol:

Assume a hostile reader exists. Find what they will do with this before they do it, then either remove it or write down why it stays.

Decision table

SituationActionLoad
Work is about to ship, merge, or publishRun all three lenses with independent judgesMethod 2-3
You must review your own artifactThreat-model it as a hostile reader firstMethod 4
A judge returned findingsAdjudicate each: fix, or park with a written rulingMethod 6
A fix landed in response to findingsRun a scoped re-review of the fix delta onlyMethod 5
A fix loop is not convergingApply the breaker: stop the line, adjudicateMethod 5
Preparing text for publicationRun the publication checklistMethod 7
Checking a skill or instruction setRun the lack-of-surprise testMethod 8
The artifact is still movingFreeze it, or review the plan insteadMethod 1

Golden rules

  1. The author never approves. Self-review carries author bias, which is why the most predictable defects survive it. Judges are separate contexts, and where possible read-only roles that cannot edit the artifact.
  2. Rubrics are checkable statements. "Covers the three failure classes" is a criterion; "high quality" is a mood. PASS means every criterion is met, because a blended score with no per-criterion feedback cannot be acted on.
  3. Every finding gets a ruling. Fixed, or parked with written reasoning. Silent discards are forbidden: an unrecorded decision is indistinguishable from a missed defect, and the next reviewer pays for it.
  4. Fix loops are bounded. One fix dispatch per wave, then a scoped re-review. At the cap, adjudicate. Unbounded loops converge on exhaustion, not on quality.
  5. The breaker protects the line. A load-bearing finding that later work would build on stops the advance even if the budget is exhausted; parking it silently would ship the defect with compounded interest.
  6. Review the delta, not the world. After a fix, re-review the fix's diff in a fresh context; re-reading everything reintroduces review that already passed and burns the loop's budget.
  7. Lack of surprise is a safety property. A reasonable reader of the description should be able to predict what the artifact does. Hidden side effects, boundary escapes, and unexplained destructive behavior are findings, not details.

The method

1. Prepare the review

State the artifact's intent in one paragraph, write the rubric as a list of checkable criteria, define what "done" means for the review, then freeze the artifact and record its hash. A review of a moving target is a conversation, not a verdict. If the artifact is too large to freeze, split it — reviewable units are a deliverable of the split, not a compromise.

2. Run the three lenses

Each lens is a separate pass with its own rubric. A judge may cover more than one lens only if each verdict comes from a genuinely separate pass.

  • Security and leakage. Does it contain secret-shaped values, personal identifiers, or target-specific operational detail? Does it instruct or enable boundary crossing? Are destructive operations behind an explicit human gate? What does it leak by structure, not just by content?
  • Correctness and adversary simulation. Simulate a hostile reader and a hostile input: empty, oversized, malformed, duplicated, out of order, concurrent. Which instructions fail on a novel input? Which claims are unsupported? What would someone do with this that the author did not intend?
  • Method and clarity. Is the method correct and complete? Does the decision table route real situations? Do the golden rules explain why? Are examples generic and reserved-range only? Does anything depend on a specific target that would stop working tomorrow?

The reference file carries the full checklists per lens.

3. Keep judges independent and read-only

Use a separate context for every judgment, review, and final gate; where the tooling allows, a role that physically cannot edit the artifact. Structured verdicts — criterion, pass or fail, evidence for the verdict — let the orchestrator route work mechanically. Judge prompts state the rubric, never a desired conclusion; a prompt that says "do not flag X" has destroyed its own review.

4. Self-review as a hostile reader

When no external judge exists, or before one sees the work, run this drill: write a one-page threat model of your own deliverable. Who is the hostile reader? What is the worst thing they could do with this text, code, or instruction set? What would they quote out of context? Which assumption of yours is visibly untested? Answer each in the artifact or in a recorded ruling. This is not a substitute for independent review, but it is where most real findings originate: the author who simulates hostility finds what politeness hides.

5. Bound the fix loop and install the breaker

Findings go to one fix dispatch per wave, not one per finding. Resume the original author first, because it has the context; escalate to a fresh implementer on a stronger tier after repeated failures. Every round ends with a scoped re-review of only the fix delta, in a fresh context. Set both a round cap and a patience counter, and stop at the first of: all criteria pass, the cap, or the breaker. The breaker fires when a load-bearing finding is neither fixed nor safely parkable — it stops the line and hands the decision to a human rather than shipping a known weak foundation. Record every termination cause.

6. Adjudicate findings in writing

Every finding ends with exactly one of two rulings:

  • Fix — the finding was addressed, and the scoped re-review confirms it.
  • Park — the finding is deliberately not addressed, with a written ruling that names the finding, the reasoning, the risk accepted, and the condition that would reopen it.

Deferred minor findings go to the final review, not fixed inline mid-loop. Adjudicating early to end a loop is the failure this step exists to prevent; the ruling must be defensible to a reader who never saw the discussion.

7. Run the publication checklist

Before anything ships: scan for secret-shaped patterns and for person-identifying or target-specific content; run an abstraction pass that replaces every specific with its class; confirm examples use reserved names and reserved test ranges only; run the privacy test — a reader who sees only the description should not be able to reconstruct operational detail; and get an independent review whose findings are fixed or parked with rulings. Clean scans do not replace review, and review does not replace scans; they catch different classes of defect.

8. Apply the lack-of-surprise test

Read only the description and predict everything the artifact does. Any discrepancy between the prediction and the artifact is a finding: hidden side effects, tools that reach outside their working area, destructive operations without a gate, or intent that is obfuscated rather than stated. The test is cheap and catches an entire class of trust failures early.

Exit gate. All three lenses have verdicts; every finding has a ruling; the publication checklist passes; and the final hash of the reviewed artifact matches the hash of the artifact that ships.

When NOT to use

  • There is no frozen artifact to judge yet — review the plan, or return when the work is frozen.
  • You need evidence about how something behaves — evidence-protocol runs the probes; this skill judges the result.
  • The task is to map or decode a system — use recon or protocol-cartography.
  • A one-line typo fix in a living document: fix it and move on. Full ceremony on trivial changes spends the review budget that risky changes need.

References

  • references/review-lenses.md — the complete checklists for all three lenses, the adjudication record template, the fix-loop and breaker rules, the publication checklist, and the hostile-reader questionnaire. Read it when running a review end to end.
  • evidence-protocol — the evidence a review inspects.
  • honest-fences — authorization and boundary doctrine.
  • self-evolving-skills — turning recurring findings into lessons.

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.