Three weeks ago my Claude Max session jumped from 21% to 100% on a normal-sized prompt. Two cache bugs were inflating token consumption 10 to 20x, After that I installed Codex. Now I run both.
Here are the skills I use in Codex.
A skill is a SKILL.md file in ~/.agents/skills/, loaded automatically when the task matches.
npm i -g /codex codex 1. WarpGrep
Codex grepping a large codebase burns 75 seconds loading context the main model doesn't need. WarpGrep is a reinforcement learning trained search subagent in an isolated context window, 8 parallel tool calls per turn, up to 36 calls in under 5 seconds. Returns only the file:line-range spans needed.
Median search drops from 75s to 5s. Software Engineering Bench Pro hits 59.1% (+3.1 points), 17% fewer input tokens, 15.6% lower cost per task.
# Add to ~/.codex/config.toml [mcp_servers.morph-mcp] command = "npx" args = ["-y", "@morphllm/morphmcp"] [mcp_servers.morph-mcp.env] MORPH_API_KEY = "your-api-key" Key at morphllm.com. Install this first, it's the only one that moves benchmarks.
2. create-plan
Forces a written plan before Codex opens a file. Which files change, what approach, what edge cases, what tests pass. You approve, then it executes.
$skill-installer create-plan Wrong-direction sessions are the most expensive thing in agentic coding.
3. gh-fix-ci
Reads the failing GitHub Actions output, identifies the cause, commits the fix. Handles flaky imports, missing mocks, test ordering, lint, environment variable mismatches.
$skill-installer gh-fix-ci 4. Valyu
Model Context Protocol server connecting Codex to ArXiv, GitHub search, docs search, and major academic sources through one integration. Optimized for fresh queries and time-sensitive question answering.
# Add to ~/.codex/config.toml [mcp_servers.valyu] command = "npx" args = ["-y", "@valyu/mcp-server"] [mcp_servers.valyu.env] VALYU_API_KEY = "your-api-key" Key at platform.valyu.ai.
5. gh-address-comments
Reads every pull request review comment, groups by type, addresses each in one session. Commits changes, responds inline, reads surrounding code per comment.
$skill-installer gh-address-comments 6. Coding CLI
What broke me on plain Codex was wiring up auth, a database, and API keys for the 40th side project. Half the session gone before any product code lands.
This handles the agent a sandboxed runtime with auth, database, storage, 30+ pre-authenticated Application Programming Interfaces (no keys to manage), and one-shot deploy to a custom domain or the App Store. Codex runs inside the sandbox, so the build-and-test loop doesn't touch your machine. Works with Codex, Claude Code, Cursor, and Gemini.
# Follow setup at github.com/vibecode/vibecode-cli # Then paste the install snippet into your agent's chat 7. frontend-skill
Bans Inter, neutral grays, and default 8px border-radius. Requires a typography rationale and color palette before the first Cascading Style Sheets line.
mkdir -p ~/.agents/skills git clone https://github.com/vipulgupta2048/codex-skills.git cp -r codex-skills/frontend-design ~/.agents/skills/ 8. stop-slop
Strips em-dashes, throat-clearing openers, binary contrasts, and passive voice from READMEs, commit messages, and comments.
mkdir -p ~/.codex/skills git clone https://github.com/hardikpandya/stop-slop.git ~/.codex/skills/stop-slop 9. Superpowers
Subagent-driven development. Agents work each task, inspect their work, continue forward.
/plugins Search Superpowers, Install Plugin.
10. Codex Security
Codex Cloud feature, not a skill. Launched March 6, 2026. Maps trust boundaries, generates an editable threat model, scans for vulnerabilities in sandboxed environments. Beta scanned 1.2 million commits, found 792 critical and 10,561 high-severity issues. Pro, Enterprise, Business, and Edu plans.
How I split the two
Claude Code for large-codebase reasoning (1M context on Sonnet 4.6 and Opus 4.7 holds up, Opus 4.6 scored 78.3% on Multi-Round Coreference Resolution v2), interactive debugging, multi-file refactors. It uses ~3-4x more tokens but wins blind code-quality reviews ~67% of the time.
Codex for terminal work (GPT-5.3-Codex leads Terminal-Bench 2.0 at 77.3%, Opus 4.7 at 69.4%), background tasks via Codex Cloud, high-volume sessions, and anywhere the ten skills run automatically.
Migration
cp CLAUDE.md AGENTS.md AGENTS.md is identical to CLAUDE.md. Rebuild Model Context Protocol configs in ~/.codex/config.toml. Codex uses Tom's Obvious Minimal Language, not JavaScript Object Notation, so config.json gets ignored.
codex mcp add server-name -- npx -y u/package/name Reinstall skills in ~/.agents/skills/. For complex setups, the cc2codex tool handles the rest. Rate limits run a 5-hour and weekly window in parallel, check /status in the Command Line Interface.
[link] [comments]