Claude codex-cli skill: 10 ways to bridge Claude Code and OpenAI Codex CLI

The prompt

Write me a shell wrapper scripts/codex-task.sh that takes a prompt file and runs `codex exec` with --json, --output-last-message, --sandbox workspace-write. Save the assistant's last message to .codex-out/<timestamp>.md. The wrapper should also log the JSONL events to .codex-out/<timestamp>.jsonl so I can grep tool calls later. Pin gpt-5.3-codex via -m. Skip git-repo-check off (we want it on). Bonus: a Makefile target `make codex` that reads .codex/prompt.md.

What slides.md looks like

#!/usr/bin/env bash
# scripts/codex-task.sh
set -euo pipefail
PROMPT_FILE="${1:-.codex/prompt.md}"
TS="$(date -u +%Y%m%dT%H%M%SZ)"
mkdir -p .codex-out
codex exec \
  -m gpt-5.3-codex \
  --sandbox workspace-write \
  --json \
  --output-last-message ".codex-out/${TS}.md" \
  "$(cat "$PROMPT_FILE")" \
  | tee ".codex-out/${TS}.jsonl"

Pairs with

The prompt

I'm in Claude Code. Add a slash command /second-opinion that: (1) runs `git diff origin/main...HEAD > /tmp/diff.patch`, (2) shells out to `codex exec --sandbox read-only -m gpt-5.3-codex 'Review this diff for SQL injection, missing auth checks, and race conditions: \\n\\n' \"$(cat /tmp/diff.patch)\" -o /tmp/codex.md`, (3) reads /tmp/codex.md back into context, (4) prints the union of Claude's review and Codex's review with disagreements flagged.

What slides.md looks like

# .claude/commands/second-opinion.md
---
description: Adversarial review by OpenAI Codex CLI on the current branch diff
allowed-tools: Bash(git diff:*), Bash(codex exec:*), Read
---

Run `git diff origin/main...HEAD > /tmp/diff.patch` then:

!`codex exec --sandbox read-only -m gpt-5.3-codex \
  "Review this diff for SQL injection, missing auth, race conditions:" \
  "$(cat /tmp/diff.patch)" \
  -o /tmp/codex-review.md`

Read /tmp/codex-review.md and reconcile with your own review. Flag every disagreement.

Pairs with

The prompt

I have ~/.claude/skills/<slug>/SKILL.md files that I want Codex to honor too. Generate scripts/sync-skills-to-codex.sh that (1) finds every SKILL.md under ~/.claude/skills, (2) writes a single ~/.codex/AGENTS.md whose body lists each skill's name + description + trigger phrases, plus a one-line `When the user asks for X, follow ~/.claude/skills/<slug>/SKILL.md`. Keep it idempotent (overwrite, don't append). Print a count.

What slides.md looks like

#!/usr/bin/env bash
# scripts/sync-skills-to-codex.sh
set -euo pipefail
SRC="${HOME}/.claude/skills"
DST="${HOME}/.codex/AGENTS.md"
mkdir -p "$(dirname "$DST")"
{
  echo "# Skills (auto-synced from ~/.claude/skills)"
  echo
  for f in "$SRC"/*/SKILL.md; do
    slug="$(basename "$(dirname "$f")")"
    desc="$(awk '/^description:/{$1=""; print substr($0,2); exit}' "$f")"
    echo "## ${slug}"
    echo "- description: ${desc}"
    echo "- on match, follow: ${f}"
    echo
  done
} > "$DST"
echo "Wrote ${DST} with $(grep -c '^## ' "$DST") skills."

Pairs with

The prompt

Generate scripts/bench-cli.sh that takes a prompt file and runs it under both Claude (`claude -p` headless mode with --output-format=json) and Codex (`codex exec --json -o`). Save both last-messages to bench/<timestamp>/{claude.md,codex.md} and write a bench/<timestamp>/diff.md that runs `diff -u` on the two files. Print the wall-clock for each run.

What slides.md looks like

#!/usr/bin/env bash
# scripts/bench-cli.sh
set -euo pipefail
PROMPT="${1:?prompt file required}"
RUN="bench/$(date -u +%Y%m%dT%H%M%SZ)"
mkdir -p "$RUN"
time claude -p --output-format=json "$(cat "$PROMPT")" \
  | jq -r '.result' > "$RUN/claude.md"
time codex exec --json -o "$RUN/codex.md" \
  --sandbox read-only "$(cat "$PROMPT")"
diff -u "$RUN/claude.md" "$RUN/codex.md" > "$RUN/diff.md" || true
echo "$(wc -l < "$RUN/diff.md") differing lines — see $RUN/"

Pairs with

The prompt

In Claude Code: produce a SPEC.md for the change I'm asking for, plus a TASKS.md with 6-10 atomic tasks (each one file, ~50 LOC). Then write a Makefile target `make codex-implement` that loops over TASKS.md and runs `codex exec --full-auto --sandbox workspace-write -o .codex-out/task-N.md` once per task, passing the SPEC.md + the single task body as the prompt. After the loop, run `pytest` and feed any failures back to Claude.

What slides.md looks like

# Makefile
codex-implement:
	@mkdir -p .codex-out
	@n=0; while IFS= read -r task; do \
	  n=$$((n+1)); \
	  echo "→ task $$n: $$task"; \
	  codex exec --full-auto --sandbox workspace-write \
	    -m gpt-5.3-codex \
	    -o .codex-out/task-$$n.md \
	    "Spec: $$(cat SPEC.md)\n\nTask: $$task" ; \
	done < TASKS.md
	@pytest -x || (echo "Failures — feed .codex-out/*.md back to Claude" && exit 1)

Pairs with

The prompt

Write scripts/route-by-size.sh that: (1) computes the patch size from `git diff --stat`, (2) if files-touched <= 2 AND lines-changed < 80, run `codex exec --full-auto`, (3) otherwise hand off to `claude -p`. Print which CLI ran and why. Read the prompt from .ai/prompt.md so I can iterate without retyping.

What slides.md looks like

#!/usr/bin/env bash
# scripts/route-by-size.sh
set -euo pipefail
PROMPT="$(cat .ai/prompt.md)"
files="$(git diff --stat | tail -n1 | awk '{print $1}')"
lines="$(git diff --shortstat | awk '{print $4+$6}')"
if [[ "${files:-0}" -le 2 && "${lines:-0}" -lt 80 ]]; then
  echo "→ codex (small: ${files}f / ${lines}l)"
  codex exec --full-auto --sandbox workspace-write \
    -m gpt-5.3-codex -o .codex-out/last.md "$PROMPT"
else
  echo "→ claude (large: ${files}f / ${lines}l)"
  claude -p --output-format=json "$PROMPT" \
    | jq -r '.result' > .claude-out/last.md
fi

Pairs with

The prompt

I want to run Claude on feature/auth-refactor and Codex on feature/migrations from the same repo, in two terminals, no merge conflicts. Generate scripts/dual-worktree.sh that creates ../wt-claude (Claude branch) and ../wt-codex (Codex branch), prints the cd commands, and writes a tiny .codex/config.toml in the Codex worktree pinning gpt-5.3-codex with sandbox=workspace-write.

What slides.md looks like

#!/usr/bin/env bash
# scripts/dual-worktree.sh
set -euo pipefail
git worktree add ../wt-claude -b feature/auth-refactor
git worktree add ../wt-codex  -b feature/migrations
mkdir -p ../wt-codex/.codex
cat > ../wt-codex/.codex/config.toml <<'TOML'
model = "gpt-5.3-codex"
sandbox_mode = "workspace-write"
approval_policy = "on-request"
TOML
cat <<EOF
Open two terminals:
  cd ../wt-claude && claude
  cd ../wt-codex  && codex
Merge later with: git switch main && git merge feature/auth-refactor feature/migrations
EOF

Pairs with

The prompt

I want to rename `User.email_address` to `User.primary_email` across 40+ files. Sequence: (1) ask Claude to write a pytest that captures the externally visible behaviour I care about (one test per public method that returns email); commit. (2) Ask Codex CLI: `codex exec --full-auto --sandbox workspace-write 'Rename User.email_address to User.primary_email everywhere it's referenced. Update SQL columns, fixtures, factories. Don't touch tests.'`. (3) Run pytest. If green, commit. If red, hand the failure back to Claude.

What slides.md looks like

# scripts/refactor-with-tests.sh
set -euo pipefail
echo "→ Claude: write characterization tests"
claude -p "Write pytest tests/test_user_email_behaviour.py that locks in current externally-visible behaviour of User.email_address. One test per public getter."
git add tests/ && git commit -m "test: characterize User.email behaviour"

echo "→ Codex: do the rename"
codex exec --full-auto --sandbox workspace-write \
  -m gpt-5.3-codex \
  "Rename User.email_address to User.primary_email everywhere it's referenced. Update SQL columns, fixtures, factories. Don't touch tests."

echo "→ Pytest gate"
pytest tests/test_user_email_behaviour.py

Pairs with

The prompt

Build scripts/codex-then-claude.sh that: (1) runs `codex exec --json --output-last-message /tmp/codex.md '<task>'`, (2) pipes the output into `claude -p 'Critique this draft. Find logic errors, untested edge cases, and style violations. Rewrite if you can do better.'`, (3) saves Claude's reply to /tmp/final.md and prints a one-line stat (lines changed by the critique pass).

What slides.md looks like

#!/usr/bin/env bash
# scripts/codex-then-claude.sh
set -euo pipefail
TASK="${1:?task required}"
codex exec --json -o /tmp/codex.md \
  --sandbox read-only -m gpt-5.3-codex "$TASK"

claude -p --output-format=json \
  "Critique this draft. Find logic errors, untested edge cases, style. Rewrite if better:\n\n$(cat /tmp/codex.md)" \
  | jq -r '.result' > /tmp/final.md

delta="$(diff /tmp/codex.md /tmp/final.md | grep -c '^[<>]' || true)"
echo "critique pass changed ~${delta} lines"

Pairs with

The prompt

Add a slash command /codex-continue that runs `codex resume --last "$ARGUMENTS"` — i.e. lets me append a follow-up to the most recent Codex session without leaving Claude Code. Save the new last-message to .codex-out/resumed.md. Print the session id Codex resumed.

What slides.md looks like

# .claude/commands/codex-continue.md
---
description: Resume the most recent Codex CLI session and append $ARGUMENTS as the next turn
allowed-tools: Bash(codex resume:*), Read
---

Run:

!`codex resume --last "$ARGUMENTS" --json -o .codex-out/resumed.md 2>&1 | tee .codex-out/resumed.jsonl`

Then read .codex-out/resumed.md and summarize what Codex did. Print the resumed SESSION_ID from the JSONL header.

Pairs with