implement

name	implement
model	claude-sonnet-5
effort	low
argument-hint	[--model <gpt-model>] [--effort <level>] <task to implement>
description	Delegate a coding task to a Codex subagent (GPT 5.6 Sol by default) that writes files directly in the working directory. Use when the user wants Codex — or "GPT-5.5" — to build, implement, write, add, refactor, fix, or otherwise change code, whether they say "have Codex do X", "get Codex to…", "hand this to Codex", or type `/codex:implement`. Do NOT use for read-only code review where Codex inspects a diff, branch, PR, or commit without editing (use codex:review); for driving desktop or browser UI (use codex:computer); when the user wants Claude to make the change itself rather than delegate; or for a trivial edit Claude can just do in one step.

Codex Implement

Delegate an implementation task to a Codex subagent running GPT 5.6 Sol (or another GPT model the user names). The subagent will edit files in the current working directory.

Choosing the model

The -m flag selects the model. Default to gpt-5.6-sol, but honor any specific model the user asks for:

• Use gpt-5.6-sol unless the user names a different model.
• If the user specifies a model — e.g. "run gpt-5.6-terra", "use gpt-5.5-codex", "with the <name> model" — pass that exact string to -m instead. Don't validate or second-guess the name; Codex will error if it's unknown.
• If they typed /codex:implement --model <name> <task> (or -m <name>), strip that flag from the prompt and use <name> as the model.

Choosing the effort

Reasoning effort is set with -c model_reasoning_effort="<level>". Default to high, but honor any level the user asks for:

• Use high unless the user names a different level.
• If the user asks in prose — e.g. "low effort", "medium effort" — substitute that level.
• If they typed /codex:implement --effort <level> <task>, strip that flag from the prompt and use <level> as the effort.

Everywhere below shows -m gpt-5.6-sol and high effort; substitute the chosen model and effort.

How to invoke

1. Take the user's prompt (either the args passed to this skill, or the surrounding request if they said "have Codex do X").

2. Run the Bash tool with:

   codex exec \
     -m gpt-5.6-sol \
     -c model_reasoning_effort="high" \
     -s workspace-write \
     --skip-git-repo-check \
     -C "$PWD" \
     "<PROMPT>"
   

   • -m gpt-5.6-sol — the model; default gpt-5.6-sol, or the model the user named (see Choosing the model).
   • -c model_reasoning_effort="high" — reasoning effort; default high, or the level the user named (see Choosing the effort).
   • -s workspace-write — Codex may write files under the workdir + $TMPDIR; this is the safe default. codex exec runs non-interactively with approval mode never (there is no -a/--ask-for-approval flag), so Codex cannot prompt to escalate — it simply stays within the sandbox. Only switch to --dangerously-bypass-approvals-and-sandbox if the user explicitly asks for autonomous, unsandboxed execution.
   • -C "$PWD" — pin the workspace to Claude's current directory.
   • --skip-git-repo-check — allow running outside a git repo. Drop this if the user's task is git-related.
   • Always set an explicit Bash timeout. Implementation runs are slow and the Bash default (120000 ms / 2 min) will kill Codex mid-run. Pass timeout: 600000 (10 min — the maximum the Bash tool allows) on every codex exec call.
   • If the task is large enough that even 10 minutes may not be enough, run the Bash call with run_in_background: true instead and poll its output rather than blocking — a foreground call cannot exceed the 600000 ms cap.

3. Quote the prompt safely. For multi-line prompts, pipe via stdin instead:

   codex exec -m gpt-5.6-sol -c model_reasoning_effort="high" -s workspace-write --skip-git-repo-check -C "$PWD" - <<'EOF'
   <PROMPT>
   EOF
   

4. After Codex finishes, briefly summarize what it changed (from its stdout tail and, if useful, git status / git diff). Do not paste Codex's full transcript back to the user.

Passing prompts

• If the user typed /codex:implement <text>, use <text> verbatim as the prompt.
• If the user said "have Codex implement X", assemble a clear self-contained prompt describing the task, relevant files, and constraints — Codex has no memory of this Claude conversation.
• Include specific file paths and function names when known. Codex will explore, but a targeted prompt is faster and cheaper.

What Codex sees

Codex starts fresh in $PWD with no context from this conversation. Anything Claude has already figured out (file paths, decisions, constraints) must be in the prompt or Codex will re-discover it.

When NOT to use this skill

• The user wants a code review with no changes → use codex:review.
• The user wants Codex to click through UI or drive a browser → use codex:computer.
• The task is trivial and Claude can do it directly in one edit → just do it; don't spawn a subagent.