agentleFS
peterbamuhigirepeterbamuhigire/proposal-skills/skills/meta/bid-red-team-dual-review/SKILL.md

bid-red-team-dual-review

Skill11 starsChanged 17 days agoMITView on GitHub

Use when a high-stakes bid, tender response, EOI, or compliance-sensitive proposal section needs adversarial two-reviewer verification with a compliance rubric before it ships; use kaizen-improvement-system for the full engine/product audit and ai-slop-audit for the routine per-section gate.

The file

We split this skill into the rules it gives an agent

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

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈3

    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
  1. Distinctive:

    No second reviewer context (a genuinely isolated second pass, human or agent) is available — see Degraded Mode.

    line 27
  2. Distinctive:

    Run the phases in order; stop and recover at any phase whose inputs or verdicts are incomplete.

    line 40
  3. Distinctive:

    If round 3 still fails, stop and escalate to the accountable reviewer/signatory named in the relevant pipeline or domain skill — do not …

    line 88
  4. Very specific:

    Every criterion has a PASS or FAIL result with a specific cited issue or evidence location.

    line 125

≈ “Common” is an estimate: at least 1 in 5 of the 372,429 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.

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

Bid Red Team Dual Review

Acknowledgement: Shared by Peter Bamuhigire, techguypeter.com, +256 784 464178. Dual-review contract adapted from the Santa Method (origin: Ronald Skelton, Founder, RapportScore.ai), read from the ECC skill engine's skills/santa-method/SKILL.md, and scoped to this engine's tender/EOI/compliance context.

<!-- dual-compat-start -->

Use When

  • A bid, tender response, EOI, or proposal section will be submitted to a buyer, evaluator, donor, or regulator and a single self-review is not enough assurance for the stakes involved (large value, regulated sector, legal/financial commitments, reputational exposure).
  • The routine ai-slop-audit per-section gate has passed but the content still carries compliance, legal, financial, or outcome-guarantee risk that a style/slop audit does not test.
  • kaizen-improvement-system step 6 ("run one time-boxed evaluator simulation or compliance red-team") calls for a red-team pass and the stakes justify two independent reviewers instead of one.
  • Content makes claims that, if wrong or unsubstantiated, would trigger a compliance knockout, a legal exposure, or a false outcome guarantee (financial services, health, public-sector, donor-funded, or regulated-industry proposals).

Do Not Use When

  • The content is an internal draft, exploratory outline, or working note not yet intended for evaluator eyes — use anti-ai-slop as the real-time guardrail instead.
  • The check is purely mechanical (page limits, price tokens in the technical folder, file naming) — use the relevant submission checklist, e.g. skills/domain-delivery/giz-eu-local-procurement-response/references/submission-pack-checklist.md, or submission-proof-and-receipt-discipline for what-was-sent evidence.
  • No second reviewer context (a genuinely isolated second pass, human or agent) is available — see Degraded Mode.

Required Inputs

ArtefactSource/providerRequired?PurposeIf absent
The content under review, in near-final formProposal draftYesFixed target for both reviewersStop; nothing to review
The evaluation rubric (base + compliance-sensitive extension, below)This skillYesObjective pass/fail criteria for both reviewersStop; do not improvise criteria mid-review
ToR/RFP mandatory requirements and evaluation criteriaBuyer documentsYes for compliance checksAnchors "approved terminology" and "jurisdiction-appropriate language" checks in the actual tender, not a past oneMark compliance checks NOT_ASSESSED, per the never-infer-evaluation-criteria rule in rules/common/core.md
Evidence register / claim sourcesProposal working papersConditionalSupports the factual-accuracy and hallucination checksFlag unverifiable claims as findings, do not silently pass them

Workflow

Run the phases in order; stop and recover at any phase whose inputs or verdicts are incomplete.

Phase 1 — Fix the target

Freeze the exact text/version under review. Santa Method is a post-generation verification layer, not a drafting technique — do not revise while reviewers are working.

Phase 2 — Independent dual review

Spawn two reviewers with no shared context:

  1. Neither reviewer sees the other's assessment, prompt, or existence.
  2. Both receive the identical rubric (below) and the identical content under review.
  3. Each reviewer returns a structured verdict per criterion: PASS or FAIL with the specific issue cited — not prose impressions.
  4. Each reviewer's brief is explicitly adversarial: "your job is to find problems, not approve." Rubber-stamping both reviewers defeats the method.

In this engine, run reviewers as separate Agent tool invocations (or, if agents are unavailable, as fully context-reset sequential passes — see Degraded Mode) so isolation is real, not simulated.

Phase 3 — Base rubric

CriterionPass conditionFailure signal
Factual accuracyEvery claim verifiable against the evidence register or the ToR/RFPInvented statistics, wrong dates, misstated buyer requirements
Hallucination-freeNo fabricated entities, quotes, credentials, or citationsAttributed quotes with no source, invented reference projects
CompletenessEvery ToR/RFP requirement and scored criterion is addressedMissing sections, skipped mandatory items, unanswered evaluation questions
Internal consistencyNo contradictions across technical, financial, and annex contentMethodology promises what the work plan or budget does not support
Compliance (base)Passes this engine's own gates already run (anti-ai-slop, ai-slop-audit)Slop-audit grade F content reaching this review unresolved

Phase 4 — Compliance-sensitive rubric extension

Apply this extension whenever the content is regulated, legal, or financial — the default assumption for tender and EOI submissions:

CriterionPass conditionFailure signal
No outcome guarantees or unsubstantiated claimsResults are stated as targets, ranges, or evidenced track record — never as guaranteed outcomes"We will guarantee a 30% cost reduction" with no basis
Required disclaimers presentEvery place the ToR/RFP or sector norm requires a disclaimer, qualification, or assumption statement has oneA financial projection with no assumptions stated; a health/safety claim with no scope limit
Approved terminology onlyLanguage matches the buyer's own terms (from the ToR/RFP, evaluation grid, or sector skill) rather than the proposer's marketing vocabularyRenaming a buyer's defined deliverable, inventing a certification name
Jurisdiction-appropriate languageLegal, tax, licensing, and regulatory statements match the jurisdiction named in the ToR/RFP, sourced per operational-readiness-and-localisation and Digital Research, not assumed from another marketA claim correct in one country asserted without qualification in another

Phase 5 — Verdict gate

Both reviewers must return PASS on every criterion for the content to ship. One reviewer catching an issue the other misses means the issue is real — that asymmetry is exactly what independent review exists to surface, not a tie-break to resolve by majority.

Phase 6 — Fix and reconverge

  1. Merge and deduplicate both reviewers' findings.
  2. Fix only the flagged issues — no unrequested rewriting.
  3. Re-run both reviewers as fresh instances with no memory of the previous round, on the same rubric, against the fixed content.
  4. Repeat for a maximum of 3 iterations total.
  5. If round 3 still fails, stop and escalate to the accountable reviewer/signatory named in the relevant pipeline or domain skill — do not ship on a fourth attempt or on partial credit.

Decision Rules

ConditionActionFailure or risk avoided
Both reviewers PASS all criteriaShip — record the result as this skill's evidence artefactAn undocumented release decision
Either reviewer FAILs any criterionFix and reconverge (Phase 6); never ship on one PASS and one FAILMajority voting hiding a material finding
3 iterations exhausted without full PASSStop, escalate to a human signatory, do not submitA fourth attempt turning an unresolved blocker into false assurance
No genuinely isolated second reviewer availableDowngrade to Degraded Mode and label the result accordingly — never present a single-reviewer pass as a dual-review PASSContext bleed being reported as independent assurance
Compliance-sensitive extension criterion failsTreat as a knockout-equivalent finding — same severity as a kaizen-improvement-system compliance blockerA legal, financial, or compliance defect being treated as style feedback

Outputs

ArtefactConsumerAcceptance condition
Dual-review verdict record (both reviewers' structured checks, iteration count, final verdict)Proposal lead, reviewer, release ownerShows both reviewers' criteria-by-criteria results and the round they converged on, or the escalation
Findings register (deduplicated, with fix applied per item)Bid leadEvery FAIL traces to a specific fix or an explicit accepted-risk note from the accountable owner

Evidence Produced

EvidenceConsumerAcceptance condition
Reviewer A and Reviewer B structured verdictsReviewer and release ownerIndependently reproducible — a third party could re-run the same rubric against the same content and reach a comparable verdict
Iteration logRelease ownerStates round count, what changed each round, and whether convergence or escalation occurred

Capability Contract

Read and search are required. This is a read-only review by default: it may inspect the frozen draft, rubric, ToR/RFP, and evidence register, but may not alter source files, submit a bid, contact a buyer, or certify compliance. Running two independent reviewer agents requires the Agent tool or an equivalent isolated second pass; fixing flagged content requires the same explicit drafting authority as the section being reviewed. Final submission remains gated by kaizen-improvement-system and this engine's own release rules — this skill produces one input to that gate, not a standalone ship decision.

Degraded Mode

If the ToR/RFP, evidence register, frozen draft, or isolated reviewer is unavailable, mark the affected criterion NOT_ASSESSED and report the missing input before any release decision. Without a genuinely isolated second reviewer (no Agent tool, no second human), simulate isolation with an explicit context reset between passes: record Reviewer A's findings verbatim, clear context completely, then run Reviewer B fresh against the same rubric and content with no visibility into Reviewer A's output. Label the result "single-session simulated dual review" rather than a true dual review, qualify the limitation to the release owner, and do not represent it as a dual-review PASS.

Quality Standards

  • Both reviewers use the same frozen version, rubric, mandatory requirements, and evidence boundary; the verdict record names each input version.
  • Every criterion has a PASS or FAIL result with a specific cited issue or evidence location; unsupported claims remain findings.
  • The final record states the iteration count, convergence or escalation, unresolved risks, and whether any criterion is NOT_ASSESSED.
  • The compliance extension is applied whenever the proposal is regulated, legal, financial, donor-funded, or otherwise compliance-sensitive; no outcome guarantee is accepted without a stated basis.
  • A reviewer result is reproducible from the retained draft, rubric, and evidence register, while submission authority remains outside this skill.

References

  • Proposal engine core rules
  • Kaizen improvement system
  • Skill authoring standard
  • Submission proof and receipt discipline

Domain Anti-Patterns

  • Letting one reviewer see the other's findings before completing their own pass (anchoring bias). Fix: freeze each brief and retain the independent verdict before reconvergence.
  • Treating a compliance-sensitive rubric failure as a style note rather than a knockout-equivalent finding. Fix: carry the failed criterion into the findings register and release decision.
  • Reusing the same reviewer instance across fix rounds instead of spawning fresh reviewers each round. Fix: start a fresh isolated pass for every iteration.
  • Inferring "approved terminology" or "jurisdiction-appropriate language" from a similar past tender instead of the current ToR/RFP (see rules/common/core.md). Fix: mark the check NOT_ASSESSED until the current buyer document is available.
  • Shipping on a fourth iteration instead of escalating after round 3. Fix: stop at the third failed round and escalate to the accountable signatory.
  • Presenting a single-reviewer or self-review pass as a dual-review PASS. Fix: label simulated isolation and retain its limitation explicitly.
<!-- dual-compat-end -->

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.