Redteam

Red teaming as applied critical thinking: challenge assumptions, surface hidden risks, and improve decisions before commitment.

RedTeam is the skill. It implements The Red Team Handbook, Version 10 — successor to the defunct UFMCS v9.0 work, extended with decision science, product red teaming, and AI-quality checks. See reference/lineage.md. Not cybersecurity penetration testing.

Superset catalog: reference/extensions-catalog.md

Setup

You MUST do these steps before proceeding:

1. Run node scripts/context.mjs once per session (or node <skill-base-dir>/scripts/context.mjs if the runtime shows the skill path). Keep cwd at the user's project. If the request names a specific decision, document, or file, append --target <path> to the command. If you've already seen its output in this conversation, do not re-run it. The script prints the project's decision context (CONTEXT.md when present) or reports NO_CONTEXT_MD. If it reports NO_CONTEXT_MD: divert into reference/init.md only when the user invoked init, or when they are starting a new decision review with no prior context and need structured capture. For scoped commands against material the user already provided in chat (a plan, memo, strategy, PRD, email draft), do not block on init — the user's input is the context. Offer /redteam init once as an optional follow-up to persist context. A missing CONTEXT.md must never block a scoped review.
2. If the user invoked a sub-command (premortem, ach, challenge, ...), you MUST read reference/<command>.md next. Non-optional. The reference defines the command's flow.
3. For any review targeting a document, plan, or codebase artifact, read the actual source material before analyzing. Do not critique from memory or summary alone.
4. Read reference/principles.md when performing any adversarial or evaluative command. It encodes the four core principles and ACT/GTM fundamentals. Skipping it produces shallow pushback instead of structured red teaming.

Core principles (always apply)

• Make the implicit explicit. Surface assumptions, frames, and biases that are driving the conclusion but not stated.
• Separate perception from interpretation. What is observed vs. what is inferred?
• Generate and evaluate alternatives. A decision without rejected alternatives is incomplete.
• Match tool to time. Use quick reflex checks when time is short; use full technique sequences when the stakes warrant it.
• Be the loyal opposition. Challenge the thinking, not the person's worth. The goal is a better decision, not winning an argument.
• Do not confuse red teaming with negativity. Identify strengths and viable paths, not only flaws.

Commands

Command	Category	Description	Reference
init	Setup	Capture decision context in CONTEXT.md and configure .redteam/	reference/init.md
challenge [target]	Evaluate	Adversarial review: assumptions, logic, risks, alternatives	reference/challenge.md
critique [target]	Evaluate	Structured critique of a document, plan, or proposal	reference/critique.md
review [target]	Evaluate	Pre-commitment decision gate before approval or launch	reference/review.md
premortem [target]	Technique	Imagine failure, then work backward to prevent it	reference/premortem.md
ach [target]	Technique	Analysis of Competing Hypotheses	reference/ach.md
assumptions [target]	Technique	Key assumptions check and sensitivity analysis	reference/assumptions.md
devils-advocate [target]	Technique	Structured devil's advocacy	reference/devils-advocate.md
5-whys [target]	Technique	Drill to root causes and hidden assumptions	reference/5-whys.md
futures [target]	Technique	Alternative futures analysis	reference/futures.md
frame [target]	Technique	Frame audit — what lens is shaping this view?	reference/frame.md
groupthink [target]	Facilitate	Design a groupthink-mitigation session	reference/groupthink.md
ideate [target]	Facilitate	Divergent thinking: generate alternatives	reference/ideate.md
converge [target]	Facilitate	Convergent decision support: narrow options	reference/converge.md
tools	Reference	Browse Handbook v10 technique catalog + extensions	reference/tools.md
Extensions
outside-view [target]	Extension	Base rates and reference class forecasting	reference/outside-view.md
invert [target]	Extension	Inversion — what would guarantee failure?	reference/invert.md
incentives [target]	Extension	Goodhart, perverse incentives, metric gaming	reference/incentives.md
ladder [target]	Extension	Ladder of inference (Argyris)	reference/ladder.md
steelman [target]	Extension	Steel-man before devil's advocacy	reference/steelman.md
calibrate [target]	Extension	Confidence and evidence grading	reference/calibrate.md
sequence [target]	Extension	Tool-chain playbooks for decision types	reference/sequence.md
culture [target]	Extension	Stakeholder worldview mapping	reference/culture.md
ai-check [target]	Extension	Meta-review of AI-assisted analysis	reference/ai-check.md
launch [target]	Extension	Product/feature launch readiness	reference/launch.md
rfc [target]	Extension	Technical RFC / ADR review	reference/rfc.md
misuse [target]	Extension	Abuse-case and harm analysis	reference/misuse.md
reversibility [target]	Extension	One-way vs two-way door classification	reference/reversibility.md
record [target]	Extension	Decision record / ADR artifact	reference/record.md

Reference-only (no dedicated command): bias-catalog.md · ai-anti-patterns.md · decision-record-template.md

Plus management commands: pin <command> and unpin <command>.

Routing rules

1. No argument: context-aware menu. If NO_CONTEXT_MD, lead with /redteam init as the top recommendation with one line on why. Otherwise recommend 2–3 highest-value commands based on what the user is working on, then show the full table. Never auto-run a command.
2. First word matches a command (table above or pin / unpin): load its reference and follow it. Everything after the command name is the target.
3. First word doesn't match, but intent maps clearly (e.g. "stress-test this plan" → challenge, "what could go wrong" → premortem, "base rates" → outside-view, "game the metric" → incentives, "is this AI analysis trustworthy" → ai-check, "launch checklist" → launch, "what playbook" → sequence): load that command's reference and proceed.
4. No clear match: general red team invocation. Apply setup, principles, and the user's full argument as context. Default to challenge flow unless a specific technique fits better — ask once if two techniques are equally plausible.

If the first word is init, or routing rule 3 maps to a from-scratch decision capture, reference/init.md owns the rest of the flow after setup.

Pin / Unpin

Pin creates a standalone shortcut so $<command> invokes $redteam <command> directly. Unpin removes it.

node scripts/pin.mjs <pin|unpin> <command>

Valid <command> is any command from the table above. Report the script's result concisely.

Output conventions

When writing artifacts to .redteam/:

• Reviews → .redteam/reviews/<slug>.md
• Session notes → .redteam/sessions/<date>-<slug>.md
• Use ISO dates in filenames. Slug from the target topic (lowercase, hyphenated).

Always tell the user what was written and where. In chat-only mode (no .redteam/), deliver the full analysis in the response; offer to persist with /redteam init if useful.

What this is NOT

• Not penetration testing, vulnerability scanning, or exploit development.
• Not instructions to deceive, manipulate, or socially engineer people.
• Not a license to be contrarian without substance. Every challenge should tie to evidence, logic, or an explicit assumption.