codex-first

name

codex-first

description

Opt-in session mode that makes Codex CLI / gpt-5.6-sol the default implementer while Claude specs, reviews, and verifies. Use only when the user explicitly opts in: invokes codex-first, asks to work codex-first, says Codex should do the implementation or coding work by default, or asks to delegate the hands-on work of a session or task to Codex. Do not trigger on ordinary implementation requests that never mention Codex-led work; without this opt-in Claude does the work directly unless another explicit delegation request or workflow applies.

Codex First

An opt-in routing posture: once the user invokes it, Codex is the default executor for hands-on work and Claude spends its tokens on specs, judgement, and verification. It stays in effect for the rest of the session unless the user says otherwise; "do this one yourself" overrides it for a single task.

Rationale: Claude tokens are metered and expensive; Codex is flat-rate and fast at writing code. Claude wins at judgement, design, spec-writing, review, and orchestration. So Codex types, Claude thinks and verifies.

Route

Delegate to Codex (the default for hands-on work):

• Implementation from a frozen spec; refactors; mechanical migrations
• Bug fixes with a known repro; test writing; coverage fills
• CI fixes, dependency bumps, scripts and tooling
• Bulk exploration or reading where raw throughput beats careful reasoning

Keep in Claude:

• Design, API design, architecture, naming, UX judgement
• Tasks where writing the spec IS the work (ambiguity means design)
• Tiny edits (roughly under 20 lines, one obvious change) — delegation overhead loses
• Anything needing session tools: MCP servers, browser sessions, secrets
• Destructive or irreversible ops, releases, pushes, GitHub mutations
• Review and verification of Codex output — never delegated, never skipped

Heuristics:

• Mixed task: Claude designs first, freezes the spec, then delegates the build-out.
• If the prompt reads as a work order, delegate; if writing it forces decisions, it is design — Claude keeps it.
• After two failed delegation rounds on the same task, take over and do it directly.

Mechanics

This skill decides whether and what to delegate; the matching codex-* skill decides how. Route through:

• codex-implementation — bounded code changes (worktree isolation, prompt contract, iteration, delivery, cleanup)
• codex-analysis — read-only bulk reading, extraction, and investigation
• codex-review — independent review of Claude-authored work
• codex-computer-use — GUI/runtime observation and verification

House invocation conventions hold in this mode: sandboxed codex exec (-s read-only / -s workspace-write), isolated worktrees for non-trivial edits, prompts via temp file, reports via -o. Do not use --dangerously-bypass-approvals-and-sandbox or equivalents.

Prompt Contract

Codex starts with zero session context. Every prompt carries: goal, exact repo and paths, constraints and non-goals, proof expected (the exact test command), and output shape. Spec quality decides success.

Verify (Claude, Always)

• Read the full diff (git status, git diff) and judge it like a contributor PR.
• Run focused tests yourself or demand proof output; Codex claims are advisory.
• Iterate with codex exec resume per codex-implementation; after two failed rounds, take over.
• The review gate still applies: Claude reviews Codex-authored diffs itself — do not send them back to codex-review, since same-model review is weak independence. Add another reviewer only when the user requests one or the selected workflow explicitly requires one.

Economics

The win is moving generation and exploration tokens to Codex while Claude spends only on spec plus diff review. Do not ping-pong trivia through delegation, and do not re-read what Codex already summarized.