red-team-review
Skill4 starsChanged 3 months agoMITView on GitHub
Score drafts against evaluator rubrics from a GS-14 perspective — narrative quality (red), mock S/W/D rating (gold), final QA (white-glove). Run after compliance-check + technical-review.
2 other projects use a version of this.
Adopted by
Most recently taken up by
- Kaagemusha/agent-skillstheir own version, first seen Sep 27, 2026
- aklmans/agent-skillstheir own version, first seen Aug 5, 2026
Other versions2 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 2
- aklmans/agent-skills★ 0same name
- Kaagemusha/agent-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 repository30 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- adversarial-reviewSkill
- capture-demand-signalsSkill
- capture-intentSkill
- capture-portal-structureSkill
- capture-scorecardSkill
- capture-submissionSkill
- competitor-assessmentSkill
- compliance-checkSkill
- customer-intelSkill
- dashboardSkill
- evidence-checkSkill
- export-proposalSkill
- import-from-captureSkill
- narrative-spineSkill
- new-proposalSkill
- opportunity-quick-lookSkill
- past-performanceSkill
- pricing-analystSkill
- proposal-editorSkill
- proposal-graphicsSkill
- proposal-managerSkill
- proposal-patcherSkill
- proposal-solution-architectSkill
- proposal-storyboardSkill
- proposal-writerSkill
- setup-companySkill
- statusSkill
- submission-summarySkill
- technical-reviewSkill
The file
Show its rules and text
We split this skill into the rules it gives an agent
27 rules. About 0 are common, 11 are distinctive, and 16 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈11
Distinctive · estimate
This author's own choices. Worth a look.
16
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 45
Not every proposal warrants every team:
- Distinctive: line 58
If no mode is specified, ask before proceeding.
- Distinctive: line 64
Purpose: Read like a GS-14 evaluator with 15 minutes and a scoring rubric.
- Very specific: line 66
Inputs: All files in drafts/, working/proposal-plan.md (win themes, evaluation criteria), working/narrative-spine.md (if present), …
- Very specific: line 95
Type check: If working/proposal-type.md declares compliance_sources: [] (white-paper, rfi, rom, sources-sought), skip formal rubric scoring …
- Very specific: line 102
All files in drafts/
- Distinctive: line 111
Load the rubric: Read the four rubric files above before scoring anything.
- Very specific: line 114
Read the factor definition from working/proposal-plan.md
- Distinctive: line 116
Read the corresponding draft section(s)
- Distinctive: line 120
Write the rationale in the voice of a source-selection decision document
- Distinctive: line 126
Not just stated as a claim but connected to a differentiator
- Very specific: line 140
Do NOT re-do the deterministic scan here — consume /evidence-check's output.
- Very specific: line 142
If my-company/evidence-ledger.json does not exist, skip this step with a note: "Phase C evidence ledger not populated — evidence citation …
- Very specific: line 150
If reference/doctrine/acquisition-guidance-catalog.md does not exist, skip this step with a note.
- Distinctive: line 155
All factors Outstanding or Good + competitive price → pWin High
- Very specific: line 259
Every run that produces reviews/gold-team-scorecard.md must also append exactly one JSON line to reviews/gold-team-history.jsonl, …
- Very specific: line 267
Append-only: Never rewrite, reorder, or delete prior lines — the accumulated series is the data scripts/compute-lift.py turns into the …
- Very specific: line 273
If reviews/gold-team-history.jsonl does not exist, create it with this first line.
- Very specific: line 275
After appending, optionally run python scripts/compute-lift.py --proposal <slug> to refresh reviews/lift.md and show the trajectory.
- Distinctive: line 359
Run after submission to extract framework improvement signal from the gap between auto-generated draft and what you actually submitted.
- Very specific: line 363
Mode A — Corpus-driven (preferred): When corpus/calibration/<slug>/ exists with status=complete (i.e., user has run /capture-submission …
- Very specific: line 365
Read corpus/calibration/<slug>/manifest.json — get type, customer, dates
- Very specific: line 366
Read corpus/calibration/<slug>/edit-notes.md — user's reflection on what they changed and why
- Very specific: line 367
Diff: every section in corpus/calibration/<slug>/auto-draft/ against the corresponding file in corpus/calibration/<slug>/final-submitted/
- Very specific: line 383
Write findings to reviews/lessons-learned-<slug>.md with sections:
- Distinctive: line 393
Mode B — Customer-debrief retrospective (fallback): When no corpus entry exists, run a traditional Shipley-style post-submit retrospective …
- Distinctive: line 404
After every submission once corpus entry is complete.
≈ “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.
Files
Red Team Review Skill
Review Gate Mode
This skill supports the Shipley-canonical 6-team color review model (see reference/methodology/color-teams.md for the full methodology). Each mode corresponds to a specific phase in the proposal lifecycle.
At the start of the skill, ask the user which review gate they want to run — or infer from context based on which review files already exist in reviews/.
Color-team modes (Shipley-canonical, full lifecycle)
| Mode | Phase | Focus |
--mode=blue | Capture (pre-bid) | Validates capture plan + win strategy. Independent senior reviewers (NOT proposal team). Output: strategy refinements, positioning actions |
--mode=black-hat | Capture (pre-bid) | Reviews competitors' likely strategies and solutions; updates own win strategy in response. Reviewers role-play top competitors |
--mode=storyboard-pink | Pre-draft | Reviews storyboards/outlines for compliance + strategy deployment BEFORE prose drafting begins. Catches structural issues before text is written |
--mode=red | Post-draft | Customer focus + completeness + clarity of strategy + mock evaluation against rubric |
--mode=mock-eval | Post-draft (alias) | Rubric-driven mock evaluation only (backward-compat alias for the framework's historical Gold mode) |
--mode=gold | Pre-submit | Executive profit/risk sign-off (Shipley-canonical Gold). NOT a proposal-quality review — that's Red's job |
--mode=white-glove | Pre-submit | Editorial QA, formatting, submission checklist, package readiness |
--mode=lessons-learned | Post-submit | Process and strategy improvement review. If a calibration corpus entry exists at corpus/calibration/<slug>/ (created by /capture-submission), diffs auto-draft vs final-submitted, ingests edit-notes.md, and writes framework improvement proposals to reviews/lessons-learned-<slug>.md. Otherwise runs as a customer-debrief-driven retrospective |
--mode=full | Default | Runs the relevant teams in sequence based on workspace state |
Compliance coverage
Compliance coverage is validated using /compliance-check, which writes reviews/compliance-gaps.md. This skill no longer runs a separate compliance pass on completed drafts — that work belongs to /compliance-check.
Backward compatibility
The framework's historical mode names continue to work:
- Old "Gold Team" (mock evaluation) = new
--mode=mock-evalOR included in--mode=red - Old "White Glove" = new
--mode=white-glove
The Shipley-canonical names (blue, black-hat, storyboard-pink, gold as profit/risk) are the preferred names going forward. Historical aliases will continue to work indefinitely.
When to skip teams
Not every proposal warrants every team:
| Proposal type | Recommended modes |
| FAR full proposal (high-stakes, $50M+) | blue, black-hat, storyboard-pink, red, gold, white-glove, lessons-learned |
| GSA MAS task-order BPA | blue, red, white-glove (black-hat optional; gold optional) |
| SBIR Phase II | red, white-glove (blue/black-hat skipped — merit review) |
| White paper | white-glove (light) |
| RFI | white-glove only |
| ROM | white-glove only |
Compliance coverage on completed drafts is validated using /compliance-check — see the "Compliance coverage" note above.
If no mode is specified, ask before proceeding.
Red Team — Narrative Quality Review
Purpose: Read like a GS-14 evaluator with 15 minutes and a scoring rubric. Find everything that will cost points.
Inputs: All files in drafts/, working/proposal-plan.md (win themes, evaluation criteria), working/narrative-spine.md (if present), working/storyboard.md (if present), reference/narrative-operating-modes.md, working/competitor-assessment.md
Review categories:
- Technical credibility — is the architecture real and defensible?
- Differentiation — why this team over competitors? Is it explicit?
- Evaluator clarity — can every answer be found in under 30 seconds?
- Evidence — every claim backed by proof (past performance, benchmarks, deployed systems)?
- Risk reduction — are risks acknowledged and mitigated explicitly?
- Redundancy — same content in multiple sections?
- Win themes — are the 3-5 win themes woven through every section?
- Graphics alignment — do visuals reinforce the narrative, or just decorate it?
- Narrative spine preservation — does the draft still carry the approved through-line, or did it collapse into a capability tour?
- Narrative mode fit — does the prose match the proposal type and page target, or does it over-explain / under-explain for the reader?
- Transition quality — do transitions show consequence, proof, selection, or decision, or merely announce structure?
Output table:
| Issue | Severity | Section | Why It Hurts Score | Recommended Fix |
Recommended rewrites: For the top 3–5 issues, provide the specific rewritten text — copy-paste ready.
Write to: reviews/red-team-notes.md
Gold Team — Rubric-Based Mock Evaluation
Purpose: Simulate the source selection board. Score the proposal against each evaluation factor using federal source-selection rubrics, citing specific Strengths/Weaknesses/Deficiencies with paragraph-level evidence. This is "write to the score sheet" in practice.
Type check: If working/proposal-type.md declares compliance_sources: [] (white-paper, rfi, rom, sources-sought), skip formal rubric scoring and run the Lightweight Reader Response pass below instead.
Inputs:
working/proposal-type.md— determines whether formal scoring or lightweight pass appliesworking/proposal-plan.md— evaluation factors, subfactors, and weights extracted byproposal-managerworking/evaluation-model.md— if present, this solicitation's typed rubric (the same one injected into the draft prompt by/proposal-writer). Score against its specific factors, weighting, and pass/fail gates; they override the generic rubric files where they conflict. Treatrejectedrows as out of scope.working/compliance-matrix.md— current coverage status (informs Technical/Management ratings)- All files in
drafts/ - Reference rubrics (mandatory):
reference/evaluator-rubrics/adjectival-ratings.md— Outstanding/Good/Acceptable/Marginal/Unacceptable definitionsreference/evaluator-rubrics/strengths-weaknesses-deficiencies.md— S/SS/W/SW/D definitions and writing conventionsreference/evaluator-rubrics/past-performance-ratings.md— if there's a PP factorreference/evaluator-rubrics/price-evaluation.md— for the Price/Cost factor
Process
- Load the rubric. Read the four rubric files above before scoring anything. Gold Team ratings must come from the rubric, not from impression.
- For each evaluation factor/subfactor:
- Read the factor definition from
working/proposal-plan.md - Determine factor type (Technical / Management / Past Performance / Price) and use the right rubric
- Read the corresponding draft section(s)
- Enumerate findings as Strengths, Significant Strengths, Weaknesses, Significant Weaknesses, or Deficiencies per
strengths-weaknesses-deficiencies.md - Every finding must cite a specific paragraph, table, figure, or the absence thereof. "Seems weak in X" is not a valid finding — it must be "Section §3.2.1, paragraph 2, states X but does not quantify Y, creating a Weakness."
- Assign the rating that emerges from the S/W/D pattern (do NOT pre-assign a rating and back-fill)
- Write the rationale in the voice of a source-selection decision document
- Read the factor definition from
Check win theme visibility. From
working/proposal-plan.md, pull the win themes defined byproposal-manager. For each theme, verify it is:- Stated in the executive summary
- Reinforced in at least 2 major technical sections
- Supported by at least 1 concrete proof point per section where it appears
- Not just stated as a claim but connected to a differentiator
Gold Team reports each theme's visibility as: Strong (all four) / Present (stated and reinforced) / Weak (stated but not reinforced) / Absent.
- Check discriminator proof points. For each discriminator in
working/proposal-plan.md, verify there is a concrete proof point in the draft (past contract, demo, technical benchmark, named personnel, certification). Discriminators without proof points are Weaknesses. - Check ghosting. From
working/competitor-assessment.md, pull the top competitor's weaknesses. For each, verify the draft subtly frames the solution in a way that highlights that competitor weakness without naming the competitor. Missing ghosting opportunities are noted but not Weaknesses.
5b. Check evidence citations (Phase C). If my-company/evidence-ledger.json exists, run /evidence-check first (or read its most recent output at reviews/evidence-check.md). For each finding:
- Every
<!-- evidence: CLAIM-UNSUPPORTED -->marker is an automatic Weakness. Rationale: the writer explicitly flagged that a claim in the draft has no backing evidence. Severity depends on where the claim appears — in the executive summary or a Section M subfactor response, it's a Significant Weakness; in a minor capability paragraph, an ordinary Weakness. - Every retired or unknown evidence ID is a Weakness — cited evidence that doesn't exist in the ledger is structurally unsupportable.
- Every restricted-context violation is a Significant Weakness — citing an item the ledger says is not cleared for this proposal type is a compliance / integrity issue.
- Unused approved evidence relevant to the proposal is an opportunity note (not a Weakness) — "consider adding EV-031 as a proof point for subfactor X."
Do NOT re-do the deterministic scan here — consume /evidence-check's output. If reviews/evidence-check.md is older than any drafts/*.md file, warn the user and recommend re-running /evidence-check first.
If my-company/evidence-ledger.json does not exist, skip this step with a note: "Phase C evidence ledger not populated — evidence citation scoring not applied."
5c. Check acquisition-guidance compliance. Read reference/doctrine/acquisition-guidance-catalog.md and apply its Applicability Matrix for this type_id:
- Every ATO / IL / FedRAMP / CMMC / certification claim must carry an evidence-ledger cite (per
reference/doctrine/prohibited-claims-doctrine.md). An uncited one is a Significant Weakness — and if it generalizes a program-scoped posture (e.g., citing IL-6 outside the ICOP/PMW-120 context), flag it as a potential integrity Deficiency. - For Software Acquisition Pathway (DoDI 5000.87) efforts — verify the technical approach answers the customer's own success criteria: an MVP→MVCR structure and a first-delivery commitment (applications path: ≤1 year from first funds, then at least annually). A SWP response that never names the cadence is a Weakness.
- For Navy software efforts — verify RAISE 2.0 mechanics are addressed where relevant (container/security-gate posture, residual risk ≤Moderate, and edge ConMon via stage-sync for DDIL/disconnected deployments). An edge-AI proposal silent on edge ConMon is a missed discriminator (opportunity note), not a Weakness.
- For any costed response — verify the pricing artifact is defensible against GAO-20-195G's four characteristics (and DoDI 5000.73 framing for DoD cost estimates). A point number with no stated basis/method is a Weakness.
If reference/doctrine/acquisition-guidance-catalog.md does not exist, skip this step with a note.
- Compute overall rating pattern. Tabulate factor ratings weighted per Section M's statement of importance. Do not invent percentages.
- Estimate pWin (rough, not definitive).
- All factors Outstanding or Good + competitive price → pWin High
- Mostly Acceptable with 1-2 Good → pWin Moderate
- Any Marginal or Unacceptable factor → pWin Low until fixed
- Any Deficiency → pWin Very Low (unawardable)
Output: reviews/gold-team-scorecard.md
# Gold Team Mock Evaluation — [Proposal Name]
**Type:** [type_id from proposal-type.md]
**Date:** [timestamp]
**Evaluator profile:** Mock GS-13/14 source-selection evaluator applying `reference/evaluator-rubrics/`
## Scorecard Summary
| Factor | Importance | Rating | Rationale (1-line) |
|---|---|---|---|
| [Factor 1 name] | [Most Important / etc.] | [Outstanding/Good/Acceptable/Marginal/Unacceptable] | [one line] |
| [Factor 2] | | | |
| [Factor N — Past Performance] | | [Substantial/Satisfactory/Neutral/Limited/No Confidence] | |
| Price | | [assessment, not rating] | |
## pWin Estimate: [High / Moderate / Low / Very Low]
[1-2 paragraphs explaining the estimate based on factor pattern + competitor assessment + price posture]
---
## Factor 1 — [Name]
**Rating:** [Adjectival]
**Source (Section M):** [quote or paraphrase the eval standard]
### Strengths
#### Strength 1: [short label]
- **Basis:** [cite: drafts/technical-approach.md §3.2, paragraph 2]
- **Benefit to Government:** [what the government gets]
#### Significant Strength: [short label]
- **Basis:** [cite]
- **Benefit to Government:** [why this is decisive, not just good]
### Weaknesses
#### Weakness 1: [short label]
- **Basis:** [cite paragraph or its absence]
- **Risk to Government:** [what could go wrong]
- **Fix to resolve:** [specific rewrite or addition]
#### Significant Weakness: [short label]
- **Basis:** [cite]
- **Risk to Government:** [material execution risk]
- **Fix to resolve:** [specific]
### Deficiencies
(None / list with fix-to-resolve for each)
### Rating Rationale
[Paragraph written in the voice of a source-selection decision document. Pattern of S/W/D → rating. Cite the rubric threshold that was met.]
### What Would Move This to [Next Higher Rating]
[Specific, actionable. "Add a Significant Strength in subfactor X by citing the [prior customer] deployment as a proof point."]
---
[Repeat for each factor]
---
## Win Theme Visibility
| Theme | Status | Notes |
|---|---|---|
| [Theme 1 from proposal-plan.md] | Strong/Present/Weak/Absent | [where it's strong, where it's missing] |
| [Theme 2] | | |
## Discriminator Proof-Point Check
| Discriminator | Proof Point Present? | Cite |
|---|---|---|
| [Discriminator 1] | Yes/No | [draft location or "absent"] |
## Ghosting Check (vs. top competitor)
| Competitor Weakness | Ghosted? | Notes |
|---|---|---|
| [Weakness from competitor-assessment.md] | Yes/No/Partial | [where in draft] |
## Prioritized Rewrite List
1. **[Highest impact fix]** — resolves [factor rating → next level or compliance Gap]
2. [Next]
3. [Next]
[List in order of scoring impact, not author convenience]
## Overall Assessment
[2-3 paragraphs. Honest view of whether this wins as drafted. If No, what must change. If Yes, what preserves the win.]Scored snapshot (mandatory — feeds the lift metric)
Every run that produces reviews/gold-team-scorecard.md must also append exactly one JSON
line to reviews/gold-team-history.jsonl, conforming to
reference/schemas/gold-team-snapshot.schema.json:
{"schema_version":"gold-team-snapshot.v1","timestamp":"<ISO-8601>","proposal_id":"<slug>","mode":"mock-eval","pwin":"Moderate","factor_ratings":[{"factor":"Technical Approach","rating":"Good"}],"counts":{"strengths":4,"significant_strengths":1,"weaknesses":2,"significant_weaknesses":0,"deficiencies":0},"draft_word_count":8214,"unsupported_claim_count":3,"notes":null}- Append-only. Never rewrite, reorder, or delete prior lines — the accumulated series is
the data
scripts/compute-lift.pyturns into the baseline→current pWin trajectory and the shrinking-edit-ratio proof. The scorecard.mdmay be overwritten each run; the history may not. - Counts come from the findings actually enumerated in the scorecard (do not estimate).
draft_word_countandunsupported_claim_countreflectdrafts/*.md(top level) at review time.- The Lightweight Reader Response pass appends too, with
"mode":"lightweight"and"pwin":null. - If
reviews/gold-team-history.jsonldoes not exist, create it with this first line.
After appending, optionally run python scripts/compute-lift.py --proposal <slug> to refresh
reviews/lift.md and show the trajectory.
Activity trail
Append to working/activity.md:
## <timestamp> — red-team-review [gold] — <N factors scored>: <OutstandingCount>O/<GoodCount>G/<AcceptableCount>A/<MarginalCount>M/<UnacceptableCount>U, pWin: <level> → reviews/gold-team-scorecard.md + reviews/gold-team-history.jsonlLightweight Reader Response
For white papers, RFIs, ROMs, and sources-sought responses — no Section M exists, so formal adjectival scoring doesn't apply. Run this lightweight pass instead:
- Read the draft.
- Imagine the intended reader (program officer, market researcher, CO). What's their first reaction?
- Assess:
- Clarity of the "so what" — is it findable in the first page?
- Credibility — does it read as someone who understands the mission?
- Call to action — is it clear what the reader should do next?
- Length discipline — is it within
page_targetfromproposal-type.md?
- Output:
reviews/gold-team-scorecard.mdwith the four dimensions, 1-line assessments each, and a prioritized rewrite list. Skip the S/W/D structure — it's not applicable.
Gold Team discipline
- Preventable findings should already be closed by
bind.proposal-writerPass 2 Step 2b runs thereference/preventable-gold-team-findings.mdsweep (scored-field placeholders, unnamed personnel, undefined acronyms, vague comparators, unanswered evaluator questions). Still score any that survive — they are real Weaknesses — but when several recur, note in the Overall Assessment that the bind sweep was likely skipped, so the fix is upstream (re-run/proposal-writer --mode=bind) rather than one-off patches. Reserve Gold Team's own judgment for the residue the sweep cannot prevent: compliance/eligibility, "reads as a pitch" / captured-not-live, and the rating call itself. - Never assign a rating without Strength/Weakness analysis first. The rating must emerge from the findings.
- Every finding cites a specific location or absence. No "feels weak" — show me where.
- Use the rubric language verbatim. "Appreciably increases" is an FAR term of art; use it, don't paraphrase.
- Don't stack trivial strengths to imply a Significant Strength. Evaluators see through this.
- Don't over-claim. Gold Team is supposed to be honest, not motivational. If pWin is Low, say so and say why.
White Glove — Final QA
Purpose: Catch everything that could embarrass the team or trigger a non-compliance finding on submission day.
Checklist — run every item:
Editorial:
- No double spaces after periods
- Consistent section numbering throughout (e.g., "5." not "5.0")
- No informal language (check: "Govt", "ours", "we've", contractions in formal sections)
- All acronyms defined on first use in each major section
- No run-on sentences; no sentence fragments
- Consistent verb tense throughout each section
Formatting:
- Page count within limits (by volume if applicable)
- Font, size, and margin compliance
- Heading styles applied correctly (not manual bold on Normal paragraphs)
- Figure captions present and numbered sequentially
- No floating/anchored images that may shift across systems
- Header and footer match requirements (document title, company, page numbers)
- Classification and distribution marking present (if required)
- Cover page complete with all required elements
Content completeness:
- All required attachments included (resumes, PPQs, certifications, representations and certifications)
- All cross-references resolve correctly
- No placeholder text remaining ([TBD], [TO BE PROVIDED], [INSERT])
- Pricing volume matches technical volume commitments (no orphaned commitments)
Submission:
- File naming convention matches solicitation requirements
- File format matches solicitation requirements (PDF, DOCX, etc.)
- Submission portal/method confirmed
- Submission deadline confirmed with timezone
Write to: reviews/white-glove-checklist.md
Full Review (Default)
Runs all four passes in sequence. Write findings to all four output files. Present a consolidated summary at the end: total issues by severity, recommendation on whether the proposal is ready to submit.
Lessons Learned — Post-Submit Review
--mode=lessons-learned
Run after submission to extract framework improvement signal from the gap between auto-generated draft and what you actually submitted.
Two operating modes
Mode A — Corpus-driven (preferred). When corpus/calibration/<slug>/ exists with status=complete (i.e., user has run /capture-submission through Phase 3):
- Read
corpus/calibration/<slug>/manifest.json— get type, customer, dates - Read
corpus/calibration/<slug>/edit-notes.md— user's reflection on what they changed and why - Diff every section in
corpus/calibration/<slug>/auto-draft/against the corresponding file incorpus/calibration/<slug>/final-submitted/ - Classify each diff into edit categories:
- Tone / voice
- Customer-language adoption
- Tightening / compression
- Adding proof points / evidence
- Restructuring / reordering
- Cross-section redundancy
- Factual fixes / hallucinations
- Formatting / numbering
- Detect patterns — edits that appear in 2+ sections are systemic, not one-off. These are the highest-leverage to fix in the framework.
- Generate framework improvement proposals — concrete, actionable suggestions targeting specific files:
- SKILL prompt language (e.g., "proposal-writer should add a rule against 'leverage' as a verb")
- Section patterns (e.g., "section X should lead with the customer's stated problem, not capability boilerplate")
- Style guide (e.g., "add anti-pattern: 'cutting-edge' / 'innovative' / 'next-generation'")
- Reference patterns (e.g., "winning examples for type Y always include ___; add to convention file")
- Write findings to
reviews/lessons-learned-<slug>.mdwith sections:- Summary statistics (total edits, edit-hours, dominant categories)
- Systemic patterns (with section frequency)
- Cited framework improvement proposals (each with: file path, current state, proposed state, rationale, expected impact)
- One-off corrections (low-priority list for completeness)
- Append to
working/activity.md:## YYYY-MM-DD HH:MM — red-team-review [lessons-learned] — N edits classified, M systemic patterns, K framework proposals → reviews/lessons-learned-<slug>.md
Mode B — Customer-debrief retrospective (fallback). When no corpus entry exists, run a traditional Shipley-style post-submit retrospective using customer debrief notes (if any), proposal team retrospective inputs, and the win/loss outcome. Output: reviews/lessons-learned-<slug>.md focused on capture/strategy (not framework calibration).
Discipline rules for Mode A
- Don't propose framework changes for one-off edits. Single-occurrence edits go in the "one-off corrections" section, not the proposals section. The signal-to-noise ratio matters — proposing 50 framework changes for one proposal devalues every proposal.
- Cite evidence in every proposal. Each framework improvement proposal must reference the specific section(s) where the edit pattern appeared. No proposals from intuition.
- Respect the user's edit-notes. Their reflection on why they made an edit is more authoritative than the AI's classification. Use the notes as ground truth.
- Do not auto-apply changes. This skill proposes changes. The user reviews, accepts, and applies (or instructs you to apply) — the skill never edits SKILL files or patterns directly.
When to run Mode A
After every submission once corpus entry is complete. Even one entry generates useful proposals — but the signal compounds across 5+ entries.
Cross-reference
- Corpus structure:
corpus/calibration/README.md - Capture skill:
/capture-submission - Future cross-proposal pattern analysis:
/recalibrate-framework(not yet built) — operates on the entire corpus rather than one entry
Output Files (by mode)
- Red Team →
reviews/red-team-notes.md - Gold Team →
reviews/gold-team-scorecard.md - White Glove →
reviews/white-glove-checklist.md - Lessons Learned →
reviews/lessons-learned-<slug>.md - Full Review → red, gold, white-glove (lessons-learned not included in
--mode=full)
Compliance coverage is validated using /compliance-check (writes reviews/compliance-gaps.md).
Critical Rule: Always write to files — never just display in chat.
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.

