Red-Team Technical Plan

Challenge a proposal constructively before resources or operational trust are committed. Produce a stronger plan, not criticism for its own sake.

Establish the decision

Identify:

• objective, success criteria, and non-goals
• active repository and system boundaries
• known constraints and instruction authorities
• proposed architecture, sequence, and irreversible choices
• users, operators, data, systems, and environments exposed
• evidence already available and claims still assumed

Separate observations, verified facts, inferences, preferences, and speculation. Do not treat popularity, credentials, a previous model's answer, or the existence of an implementation as proof.

Choose review depth

Scale the review to consequence:

• light: local, reversible, low-cost change
• standard: persistent project change or meaningful maintenance burden
• deep: security boundary, production data, broad distribution, expensive migration, irreversible state, or multi-system automation

Do not manufacture blockers for trivial work. State the selected depth and why.

Attack the proposal

Test applicable dimensions:

1. Problem fit — Does it solve the actual objective, or a convenient proxy?
2. Scope — What hidden systems, repositories, users, and lifecycle work are implied?
3. Assumptions — Which prerequisites are unverified, unstable, or mutually inconsistent?
4. Correctness — What inputs, edge cases, concurrency, partial failures, or stale state break it?
5. Security and privacy — What trust boundaries, permissions, secrets, supply-chain inputs, personal data, or unsafe defaults are introduced?
6. Compatibility — Which versions, platforms, clients, formats, or APIs are untested?
7. Operations — How are installation, observability, recovery, ownership, upgrades, and support handled?
8. Maintenance — What duplicates an existing authority or creates a second system that will drift?
9. Economics — What token, compute, storage, network, licensing, and human review costs recur?
10. Exit — Can the change be disabled, downgraded, migrated, or removed without losing unrelated state?

For high-impact automation, also examine idempotency, transaction boundaries, retries, rate limits, duplicate processes, configuration corruption, and operator interruption.

Compare alternatives

Include the strongest practical alternatives, including:

• do nothing yet
• narrow pilot
• strengthen an existing capability
• simpler manual or deterministic mechanism
• staged migration
• proposed approach with revised safeguards

Compare them against the same criteria. Do not use a straw alternative to make the preferred proposal look better.

Verify material claims

Inspect current project evidence and use authoritative, current sources for version-sensitive claims. Verify critical findings against source or runtime behavior when possible. Label unverified risks and avoid converting a possible failure into a claimed fact.

Do not mutate the project unless the user separately authorizes implementation.

Produce the decision review

Return:

• verdict: proceed, proceed with changes, pilot first, defer, or reject
• strongest aspects worth preserving
• blocking flaws and why they block
• material non-blocking risks
• assumptions requiring evidence
• alternative comparison
• revised design and acceptance criteria
• validation, monitoring, rollback, and stop conditions
• unresolved questions that materially change the decision

Rank findings by consequence and likelihood. End with the smallest defensible next step.