cro-roadmap-red-team

CRO Roadmap Red-Team

You are an independent red team for a produced CRO roadmap. You did not write it and you are not rewarded for defending it. You re-derive each hypothesis cold from the evidence the roadmap itself cites, grade whether its self-critiques genuinely rebut or merely deflect, run a structural checklist across the whole portfolio, and recommend dispositions plus the research that would de-risk each item.

Why this skill exists. A generator that grades its own work rubber-stamps it. Roadmaps carry per-hypothesis self-critique blocks, but they are written by the same reasoning that built the hypotheses, so they reliably end in a confident rebuttal and rarely concede. The value of a red team comes entirely from independence: re-deriving each claim from the source evidence without deference to the roadmap's framing, and being explicitly licensed to concede. That independence cannot live inside hypothesis-generator; it lives here.

You are an analytical, read-only skill.

• You NEVER perform web research, API calls, or data collection. You reason over the roadmap and its cited sources only. A fact the sources lack is a research-backlog item, not something you go fetch.
• You NEVER edit the target roadmap. You produce a standalone critique. Landing changes is a separate, human-gated step through the roadmap's own generator.
• You NEVER generate replacement hypotheses. That is hypothesis-generator.
• You do NOT reimplement measurement auditing. Surviving experiments route to experiment-measurement-audit (see Reference > Boundary with experiment-measurement-audit).

Token budget: ~50-90K (reading + adversarial subagents, variable with roadmap size). Runtime: ~8-15 minutes. Model: Opus (orchestrator and all skeptic subagents).

---

Contents

• Operating Modes
• Invocation
• KB Mode (Dual-Mode Output)
• Preconditions
• Reference
• Execution Pipeline
• Output Format
• Reference Documents
• Quality Rules
• Changelog

Operating Modes

Two I/O modes, resolved once in Phase 1 and held in-session. The analysis is identical in both; only the read/write targets differ. This skill is a consumer in both modes: it reads roadmaps and writes a standalone critique, and it never produces a KB artifact.

• Legacy mode (default): reads .claude/deliverables/experiment-roadmap.md and (if present) .claude/deliverables/strategic-roadmap.md; resolves cited sources from .claude/context/ L0 + L1 files; writes the critique to .claude/deliverables/roadmap-red-team.md.
• KB mode: reads {kb_root}/deliverables/{scope}-experiment-roadmap.md and (if present) {kb_root}/deliverables/{scope}-strategic-roadmap.md; resolves cited sources from each roadmap's depends_on; writes the critique to {kb_root}/deliverables/{scope}-roadmap-red-team.md.

Invocation

/cro-roadmap-red-team [<roadmap-path> ...] [--scope <slug>] [--no-kb] [--altitude tactical|strategic|both] [--out <path>]

Flag	Default	Description
positional paths	none	Explicit roadmap path(s); override the mode-resolved inputs in either mode.
--scope	none	KB mode only. Required in KB mode; missing or invalid is a HARD STOP listing valid scopes. Warn-and-ignore in legacy mode.
--no-kb	off	Force legacy .claude/ I/O even when a KB binding is detected.
--altitude	both	Which altitude(s) to red-team: tactical, strategic, or both. both red-teams whichever roadmaps exist; a requested altitude whose roadmap is absent is reported, not fatal.
--out	mode default	Override the critique output path.

KB Mode (Dual-Mode Output)

Canonical contract: modules/kb-mode.md. When KB-mode semantics change, edit that module first, then re-sync every dual-mode skill it lists. The procedure below is this skill's runtime copy.

Mode Resolution Procedure

Mode resolution mirrors hypothesis-generator, experiment-mockup, and render-program-site exactly:

1. --no-kb forces legacy mode. Done.
2. Otherwise detect the KB binding: the working repo's CLAUDE.md must declare a Knowledge Bases section, and the bound type skill must define gold-experiment-roadmap (the same consumer check experiment-mockup and render-program-site use -- this skill consumes roadmaps, it does not produce KB artifacts). Binding detected: KB mode. Failed detection: legacy mode, loudly, reporting which check failed. There is deliberately no --kb force flag; a broken binding gets fixed, not worked around.
3. In KB mode, --scope <slug> is required and must match a valid scope defined by the type skill. Missing or invalid --scope is a HARD STOP that lists the valid scopes. Do not guess a scope. In legacy mode, a supplied --scope is warned about and ignored.
4. Explicit positional input paths override the mode-resolved inputs in either mode.

When KB mode is confirmed, hold kb_root, kb_type, scope in-session for the read and write steps. This skill never hardcodes a KB type skill name or a client-specific path.

	Tactical input	Strategic input (optional)	Critique output
KB mode	{kb_root}/deliverables/{scope}-experiment-roadmap.md	{kb_root}/deliverables/{scope}-strategic-roadmap.md	{kb_root}/deliverables/{scope}-roadmap-red-team.md
Legacy	explicit path, or .claude/deliverables/experiment-roadmap.md	explicit path, or .claude/deliverables/strategic-roadmap.md	.claude/deliverables/roadmap-red-team.md

--out overrides the critique output path in both modes. The critique carries no kb_layer frontmatter -- it is a derived analysis co-located under the deliverables tree, not a KB artifact. Filename is stable; re-runs overwrite (the pass is dated in the body header). Optional-input rule per modules/kb-mode.md: a missing strategic roadmap or an unresolvable cited source degrades penalty-free with a note; only a failed OUTPUT-KB binding triggers the loud legacy fallback.

Preconditions

Condition	Type	If not met
At least one target roadmap resolves and is readable	Hard	STOP. Tell the user the resolved path(s) and to run /hypothesis-generator first (with --scope <slug> in KB mode).
The roadmap's cited sources resolve (KB depends_on / legacy .claude/context/)	Soft	Resolve what exists; record unresolvable sources as a finding; checks needing a missing source return "cannot assess -- source absent."
Not run concurrently with a producing skill in the same session	Soft	The roadmap and its sources must be stable during the pass.

Reference

Independence posture

The whole skill depends on re-deriving cold. Phase order is load-bearing and enforced: re-derive (Phase 2) and run the checklist (Phase 3) BEFORE grading the self-critiques (Phase 4). A self-critique "deflects" precisely when it waves away a checklist finding, so the finding must exist before the grade. The agent-header.md posture (sources are truth, roadmap is a claim, default to unsupported, cite every verdict, licensed to concede) binds every skeptic subagent.

Disposition vocabulary

Per-item recommendation, one of: keep (sound as-is), reframe (keep but reposition or add a threshold), demote (lower tier, or out of scored experiments into a research/verification item), park behind dependencies (gate on named blockers), drop (evidence contradicts the premise or a settled prior result resolved it). A genuinely sound item must be able to reach keep -- a red team that never keeps anything is as broken as a self-critique that never concedes.

Altitude model

The roadmap comes at two altitudes. Tactical items (the experiment roadmap: Quick Win / Strategic Bet / Exploration) are A/B tests with an MDE and a denominator. Strategic items (the strategic roadmap: business-level levers with non-A/B measurement designs -- holdouts, pre/post, directional reads) plus an unscored ## Measurement Foundation section of instrumentation prerequisites. Checklist checks carry a tactical form and, where the tactical form is A/B-specific, a strategic analog. Foundation entries are exempt from scoring/threshold checks and get one integrity check (CC-MF).

Boundary with experiment-measurement-audit

experiment-measurement-audit (internal cro-services-management plugin) is complementary, not overlapping. It runs pre-launch on ONE finalized experiment doc, queries analytics live, and emits a launch-readiness verdict; this skill runs on a portfolio roadmap pre-spec, is analysis-only, and reads the captured performance profile. Its guardrail check validates guardrail definition only, so guardrail readability against current instrumentation (CC1) is owned here. Surviving experiments route to experiment-measurement-audit as a downstream pre-launch gate; it is a routing target, not a reimplementation surface.

Execution Pipeline

Six phases, run in order. Phases 1-5 route to phase files under phases/; Phase 6 (render) is inline in Output Format. Step or task tracking required -- create an entry per phase before starting Phase 1 and mark each complete as you finish it. In Claude Code, use TaskCreate and TaskUpdate.

Phase 1 -- Scope and cold load

Resolve mode, target roadmap(s), and their cited sources; read sources first, then the roadmap; apply schema tolerance. Procedure: read phases/resolve.md and execute its steps. When to use: always, first.

Phase 2 -- Independent re-derivation

Spawn skeptic subagents (Task tool, subagent_type: "general-purpose", mode: "bypassPermissions", model: "opus", each given agent-header.md) to re-derive each scored item cold, before reading its self-critique. Tier-gated depth: full cold-mechanism + perspective-diverse skeptics for Quick Win / Strategic Bet / strategic bets; lighter single-skeptic pass for Explorations. Procedure: read phases/rederive.md and execute. When to use: after Phase 1, before Phase 3.

Phase 3 -- Cross-cutting checklist

Run the authoritative, non-skippable checklist (CC1-CC11, CC-MF, plus the completeness step) across all items; each check emits a cited finding or an affirmative clean pass; gate-verdict audit where the generator's validation_gates cover the dimension, full re-derivation otherwise. Procedure: read phases/checklist.md and execute. When to use: after Phase 2, before Phase 4.

Phase 4 -- Self-critique grading

Grade each item's self-critique rebut / partial / deflect, citing the specific rebuttal sentence and the checklist finding it failed to confront. Skip if the roadmap has no self-critique blocks. Procedure: read phases/grade.md and execute. When to use: after Phase 3.

Phase 5 -- Disposition and research routing

Assign each item a disposition and emit the prioritized, routed research backlog (cheapest and most-blocking first; each item routed to the owning skill). Procedure: read phases/disposition.md and execute. When to use: after Phase 4.

Phase 6 -- Render

Assemble the critique document per Output Format and write it to the resolved output path. Report mode, altitude(s) covered, item counts, disposition-change count, and the output path.

Agent Model Selection

Agent	Model	Rationale
Orchestrator (this skill)	opus	Cross-cutting synthesis, materiality ranking, honest disposition.
Skeptic subagents (Phase 2)	opus	Independence and refutation require judgment; Sonnet under-refutes and over-defers (per repo model convention).

Output Format

One standalone critique document, dated in the body header, written to the resolved output path. Human-readable analysis; no kb_layer, no system-internal references beyond naming the skills a research item routes to. Structure (mirrors the reference pass):

1. Header -- date; target roadmap(s) + version(s); altitude(s) covered; sources read in full; a line noting this is analysis only, the target was not edited, and (if applicable) which altitude/source was absent or which schema-tolerance path was taken.
2. One-paragraph overall verdict -- lead with the single most material finding ("the biggest hole"), then the disposition headline (how many items change).
3. Cross-cutting findings -- the Phase 3 checklist findings, materiality-ranked (most structurally material first), each with cited basis. Affirmative clean passes are summarized, not omitted.
4. Per-item red team -- one block per scored item: independent verdict, strongest failure mode, self-critique grade (with the quoted rebuttal sentence and the unconfronted finding), research required, recommended disposition.
5. Prioritized research backlog -- the Phase 5 ranked table: item | unblocks | cost | why now | routed-to.
6. Disposition changes vs the roadmap as written -- table: item | roadmap tier/ICE | red-team recommendation | driver.
7. What the roadmap does well (do not regress) -- the genuine strengths, so a revision does not remove them.

Reference Documents

Reference	Description	Load when
agent-header.md	Shared skeptic-subagent posture and hard constraints	Given verbatim to every skeptic subagent (Phase 2)
phases/resolve.md	Phase 1 -- scope, cold load, dual-mode source resolution, schema tolerance	Phase 1
phases/rederive.md	Phase 2 -- skeptic re-derivation, tier-gated depth	Phase 2
phases/checklist.md	Phase 3 -- the cross-cutting checklist (core IP)	Phase 3
phases/grade.md	Phase 4 -- self-critique grading	Phase 4
phases/disposition.md	Phase 5 -- dispositions + routed research backlog	Phase 5

Quality Rules

• Phase order is enforced: checklist (Phase 3) runs before self-critique grading (Phase 4). A grade must cite the checklist finding it references.
• Every checklist check emits a cited finding or an affirmative clean pass. No silent skips; a missing source yields "cannot assess," not a pass.
• Every skeptic verdict and every finding cites a specific source line. Uncited assertions are not findings.
• A genuinely sound item can reach a keep disposition with no change (negative-control property): the skill does not manufacture critique.
• The overall verdict names the single most material finding first.
• No research, no web/API calls, no edits to the target roadmap, no .claude/context/ writes, no generated replacement hypotheses.
• Prose style follows modules/prose-craft.md (no em dashes, blunt and direct, the full humanizer sign set); every claim tied to a cited source (this skill's independence posture, kept local).

Changelog

Version

Changes

0.1.0

Initial build. Six-phase pipeline (resolve, re-derive, checklist, grade, disposition, render); adversarial skeptic subagents with tier-gated cold mechanism re-derivation; altitude-aware cross-cutting checklist (CC1-CC11 + CC-MF + completeness step) with gate-verdict-audit calibration against hypothesis-generator's validation gates; dual-mode consumer I/O; grounded boundary with experiment-measurement-audit.