Skip to content

Latest commit

 

History

History

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
 
 
 
 
 
 
 
 

README.md

Codex Adversarial Plan Review Skill

adversarial-plan-review-codex red-teams a plan before execution. It is a companion skill for deep-planning-codex, and it can also be used directly when a plan needs hostile review for stale assumptions, missing validation, rollback gaps, sequencing errors, stakeholder risks, or unsafe execution.

What This Skill Does

  • Reads the plan, criteria, evidence catalog, assumptions, dead ends, probes, and verification plan.
  • Selects software, business, or mixed review lenses.
  • Checks evidence, order of operations, validation relevance, rollback, permissions, and stakeholder/data risks.
  • Classifies findings as BLOCKING, IMPORTANT, or NOTE.
  • Returns PASS, FAIL, or PARTIAL.

Source Provenance

This folder is a Codex-native derivative of:

  • Source gist: https://gist.github.com/dachent/cdc05151d047708c290bd4da0aaeed96
  • Source file: deep_planning.txt
  • Source revision: 6ea4c02e5aa60c9991e1e4d1c50089c01cd6ec83
  • Source phases adapted: Phase 3 grilling, Phase 6 UltraPlan review, and Phase 7 execution-plan validation
  • Source type: user-authored Claude-native deep planning prompt template
  • Port depth: new Codex-native companion skill

Codex Conversion

The source gist relies on Claude-native grilling and planning commands to challenge assumptions. The Codex version makes that challenge explicit as a reusable review skill with severity labels, pass/fail criteria, and write scope.

The conversion is Codex-native because it works without Claude slash commands, does not assume subagents are available, and can review code plans, no-Git work, business artifacts, and mixed workflows using the same artifact contract.

Files

Validation Evidence

Validation before publication covered metadata checks, static forbidden-token scans, and live subagent forward testing on a deliberately flawed plan. The live run returned FAIL with blocking findings before execution.