Codex Think About
Purpose
Use this skill for peer reasoning, not code review. Claude and Codex are equal analytical peers; Claude orchestrates the debate loop and final synthesis.
Prerequisites
- A clear question or decision topic from the user.
codexCLI installed and authenticated.codex-reviewskill pack is installed (npx github:lploc94/codex_skill).
Runner
RUNNER="{{RUNNER_PATH}}"Workflow
- Ask user to choose reasoning effort level:
low,medium,high, orxhigh(default:high). Gather factual context only (no premature opinion). SetEFFORT. - Build round-1 prompt from
references/prompts.md. - Start Codex + Claude Independent Analysis (parallel): a. Start Codex thread:
node "$RUNNER" start --working-dir "$PWD" --effort "$EFFORT" --sandbox danger-full-access. b. Claude Independent Analysis (IMMEDIATELY, before polling): Analyze the question independently using own knowledge and optionally MCP tools. Follow the structured format inreferences/claude-analysis-template.md. Complete this BEFORE reading any Codex output. Seereferences/workflow.mdStep 2.5 for detailed instructions. c. INFORMATION BARRIER: Do NOT read$STATE_DIR/review.mdor interpret Codex's conclusions until Step 5. Poll activity telemetry (file reads, URLs, topics) is allowed for progress reporting. - Poll Codex with adaptive intervals (Round 1: 90s/60s/30s/15s..., Round 2+: 45s/30s/15s...). After each poll, report specific activities from poll output. See
references/workflow.mdfor parsing guide. NEVER report generic "Codex is running". - Cross-Analysis: After Codex completes, compare Claude's independent analysis with Codex output. Identify genuine agreements, genuine disagreements, and unique perspectives from each side. See
references/workflow.mdStep 4. - Resume via
--thread-idand loop until consensus, stalemate, or hard cap (5 rounds). - Present user-facing synthesis with agreements, disagreements, cited sources, and confidence.
Effort Level Guide
| Level | Depth | Best for |
|---|---|---|
low | Surface check | Quick sanity check |
medium | Standard review | Most day-to-day work |
high | Deep analysis | Important features |
xhigh | Exhaustive | Critical/security-sensitive |
Required References
- Execution loop:
references/workflow.md - Prompt templates:
references/prompts.md - Output contract:
references/output-format.md - Claude analysis format:
references/claude-analysis-template.md
Rules
- Keep roles as peers; no reviewer/implementer framing.
- Codex must NOT modify, create, or delete ANY project files.
danger-full-accesssandbox is used SOLELY for web search. Prompt contains strict guardrails. - Codex MUST cite sources (URL) for factual claims from web.
- Separate researched facts (with sources) from opinions.
- Detect stalemate when arguments repeat with no new evidence.
- End with clear recommendations, source list, and open questions.
- Information barrier: Claude MUST complete its independent analysis (Step 3b) before reading Codex output. This prevents anchoring bias.