codex-bridgelisted

Lets Claude call the Codex CLI for a single ad-hoc task — no protocol, no rounds, no synthesis. Use this when you want a one-shot second opinion or a quick external draft without the full MAD scaffolding. Triggers: - "have Codex <verb> <thing>" (e.g. "have Codex draft X", "have Codex review file Y") - "ask Codex to <verb>" - "use Codex for <task>" - "run Codex on <file>" - Explicit "codex-bridge" mention. For structured Claude + Codex collaboration (drafts → cross-review → revisions → merge), use the mad-build skill instead. For an adversarial audit of a research document, use the mad-research skill instead.
tjhavranek/mad-research · ★ 3 · Data & Documents · score 76

Install: claude install-skill tjhavranek/mad-research

# codex-bridge skill A minimal foundation: just the verified `codex exec` invocation pattern plus a doctor for pre-flight checks. No multi-round protocol, no synthesis, no role assignment. ## When this skill is the right tool - "Have Codex draft a Python function for X." — one-shot draft. - "Have Codex review the file I just wrote." — one-shot review. - "Ask Codex to extract the tables from this PDF and write them as markdown." — one-shot transformation. - "Run Codex on this dataset and tell me what stands out." ## When this skill is NOT the right tool - You want Claude AND Codex to both draft something and then critique each other → use **`mad-build`**. - You want a structured adversarial audit of a paper or grant → use **`mad-research`**. ## When Claude should proactively propose codex-bridge The triggers above cover user-invoked use ("have Codex do X"). There is one additional case: **sequential routing**. When Claude is working alone on a user task and hits a concrete wall that codex-bridge's harness may clear, Claude should name codex-bridge as a user-controlled option — not struggle silently, partially complete, or apologize without action. Use this only for observable limits, not for "this feels hard": - Missing browser / GUI automation. - Sandbox or network operation denied by Claude Code's harness. - Package / tool installation failed for a needed dependency. - Long-running shell or background workflow that exceeds the harness. - Repeated authenticat