red-team
Skill2 starsChanged yesterdayMITView on GitHub
Before delivering, switches into a hostile critic whose only job is to break the work (edge cases, false assumptions, logic holes, what a sceptical expert would attack), ranks the findings by severity, fixes them, then delivers. Use before handing over anything that will be relied on, shipped, sent or published, or when the user says "red-team this", "poke holes in it", "what's wrong with this", "stress test it" or "be harsh".
31 other projects use a version of this.
Adopted by
Most recently taken up by
- 0xdewy/skillstheir own version, first seen Nov 16, 2026
- yusenthebot/EvolvingLooptheir own version, first seen Nov 15, 2026
- darkweb19/skillstheir own version, first seen Nov 8, 2026
- kazemcodes/skill-treetheir own version, first seen Nov 8, 2026
- drewmattie-code/red-teamtheir own version, first seen Nov 8, 2026
- Leostruka/devin-bundletheir own version, first seen Nov 1, 2026
Used with
Projects with it also use retro, architecture and 10 more.
Other files people use alongside this one.
Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
AWSSkillretro
Conduct a retrospective on a coding session.
Also in 4 projects that use this
AnthropicSkillarchitecture
Create or evaluate an architecture decision record (ADR). Use when choosing between technologies (e.g., Kafka vs SQS), documenting a design decision with trade-offs and consequences, reviewing a system design proposal, or designing a new component from requirements and constraints.
Also in 3 projects that use this
AWSSkillwizard
Generate an interactive bash wizard that walks a human through steps only they can perform. Use when provisioning infrastructure, setting up credentials or CI secrets, walking an unfamiliar third-party dashboard, or running a one-off migration or cutover. Don't invoke this for steps the agent can perform itself.
Also in 3 projects that use this
GoogleSkillcodebase-design
Shared vocabulary for designing deep modules. Use when the user wants to design or improve a module's interface, find deepening opportunities, decide where a seam goes, make code more testable or AI-navigable, or when another skill needs the deep-module vocabulary.
Also in 4 projects that use this
GoogleSkilldomain-modeling
Build and sharpen a project's domain model. Use when the user wants to pin down domain terminology or a ubiquitous language, record an architectural decision, or when another skill needs to maintain the domain model.
Also in 4 projects that use this
MicrosoftSkillgrill-with-docs
Grilling session that challenges your plan against the existing domain model, sharpens terminology, and updates documentation (CONTEXT.md, ADRs) inline as decisions crystallise. Use when user wants to stress-test a plan against their project's language and documented decisions.
Also in 4 projects that use this
- +3 more from google
Other versions31 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 31
- zakirkun/deep-eye★ 2.3ksame name
- suboss87/FDEOps★ 940same name
- griffinhilly/claude-code-synthesis★ 67same name
- geekatron/jerry★ 37same name
- Hassaan146/claude-skills★ 18same name
- PostHog/duckhog★ 14same name
- docxology/codomyrmex★ 11same name
- Zach-etier/skills★ 4same name
- drewmattie-code/red-team★ 3same name
- csiddhant796-blip/claude-skills-collection★ 2same name
21 more not listed
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 repository17 other files it gives its agents.
- autopilotSkill
- autopsySkill
- confidence-mapSkill
- delegateSkill
- expert-askSkill
- failure-museumSkill
- ghost-usersSkill
- grill-dumpSkill
- intent-diffSkill
- md-to-mapSkill
- mind-mapSkill
- premortemSkill
- shipitSkill
- stuck-detectorSkill
- taste-thiefSkill
- tournamentSkill
- workSkill
The file
We split this skill into the rules it gives an agent
5 rules. 0 are common, 4 are distinctive, and 1 is very specific to this project.
0
Common
Most of the 29 edited versions we read keep these.
4
Distinctive
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
- Very specific: line 25
Re-attack the fixes: Every fix is new work nobody has reviewed, and fixes break things too (a fix that adds .lower() while a lookup still …
- Distinctive: line 26
No invented findings: If an honest attack finds one real break, report one.
- Distinctive: line 32
Attack through each lens: Write down every break found.
- Distinctive: line 46
Fix: every Fatal and Serious. Fix a Minor only if the fix is one line.
- Distinctive: line 50
If the user asked only for a critique of work that is theirs, stop after step 4 and hand over the ranked findings with a proposed fix for …
“Common” is measured: kept by at least 6 in 10 of the 29 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Red Team
Reviewing your own work finds the problems you already suspected. An attacker who only wins by breaking it finds the others.
Scale
| Work | Do |
| A chat answer, a rename, a one-line change | Skip, silently, unless the user explicitly asked for a red team: then do a quick strike. |
| Ordinary work | Quick strike: one round, three questions. What input or reader breaks this? Which assumption did I never check? What does the most sceptical expert say first? Fix what breaks and print one line (Red team: fixed <n> breaks), or nothing if none. |
| Relied on, public, expensive or hard to undo | Full attack: the process below. |
Hard rules
- The author leaves, the attacker arrives. The attacker did not write this, does not know what was meant, and reads only what is there. Its win condition is a break. "Looks fine" is a loss.
- A finding is a concrete break: a specific input, reader, scenario or counter-example, plus what goes wrong. "Could be clearer" and "consider adding" are not findings.
- Prove it where you can. Code: run the breaking input. A factual claim: open the source. Batch the breaking inputs into one run. A finding you could not prove is marked
suspected. - Assumptions before details. The worst breaks are in what the work takes for granted, not in its wording.
- Repair, don't hedge. Adding "may", a caveat or a disclaimer is not a fix. If something cannot be fixed, say so in one line at delivery.
- Re-attack the fixes. Every fix is new work nobody has reviewed, and fixes break things too (a fix that adds
.lower()while a lookup still uses the raw key). - No invented findings. If an honest attack finds one real break, report one. Padding to look thorough buries it.
Process
- State the claim. One line: what this work promises, and to whom. That is the target.
- List what must be true for the claim to hold (3–7 assumptions about inputs, the reader, the environment, the facts). Mark each
checkedorunchecked. Attack through each lens. Write down every break found.
Lens Ask Boundaries Empty, zero, negative, huge, duplicate, malformed, twice in a row, out of order. For prose: the reader who skims, reads only the first line, or arrives with no context. Assumptions Take each uncheckeditem from step 2 and make it false. What happens?Logic Where does a step not follow? Where does the conclusion need more than the evidence gives? What is the counter-example? Sceptical expert The first objection of the person who knows this field best and wants to say no. Adversary Someone exploiting it, misreading it on purpose, or using it for something it was not meant for. Omission What the claim requires that is not there: the error path, the cost, the alternative never considered, the "so what". Time What breaks in a month, at ten times the size, or when a dependency changes. Start where this kind of work usually fails: code at boundaries and the error path; writing at the first paragraph and the unsupported claim; plans at the dependency and the missing owner; designs at the empty, error and overloaded states; research at the source and the alternative explanation.
- Rank. One row per root cause: if five inputs fail because nothing validates, that is one row naming them all.
Fatal: the claim from step 1 is false, or the work does harm.Serious: a realistic user hits it.Minor: real, but rare or cosmetic. - Fix every Fatal and Serious. Fix a Minor only if the fix is one line; otherwise list it.
- Re-attack the changed parts. A round is one pass over the lenses on the fixes only; add any new break to the table. Stop when a round finds no Fatal or Serious, or after two re-attack rounds. Anything still open gets reported.
- Deliver the table, then the fixed work. Never the original.
If the user asked only for a critique of work that is theirs, stop after step 4 and hand over the ranked findings with a proposed fix for each. Don't rewrite their work uninvited.
Output format
RED TEAM: <the claim>
| # | Severity | Break | Proof | Fix |
|---|---|---|---|---|
| 1 | Fatal | `parse("")` returns 0 instead of failing | ran it | now raises ValueError |
| 2 | Serious | A reader who stops after line 1 never learns the price | n/a | price moved to line 1 |
| 3 | Minor | ... | suspected | not fixed: <reason> |
Open: <what is still broken and why> | noneList every Fatal and Serious, plus any Minor you fixed or deliberately left. Mark breaks found in a fix as (in fix) under Break. No preamble, no praise.
Anti-patterns
- Findings that are style preferences.
- The praise sandwich ("overall solid, but").
- Finding only what is easy to fix.
- A finding with no input, reader or scenario attached.
- Fixes that leave the work longer, vaguer and covered in caveats.
- Attacking, reporting and then delivering the unfixed work.
- A third and fourth re-attack round. Two is the limit; say what is left.
- Checking the work against the request (that is
intent-diff) or walking users through it (ghost-users). This skill only tries to break it.
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.

