Backend

11 Codex Skills That Fix the 'Looks Fine, Then Blows Up' Problem

By 苏三说技术 · Nov 7, 2026

Read original on juejin.cn ↗ Google Translate ↗ Alt translation

The 80-95% token reduction changes what AI agents can attempt: long multi-session refactors and automated PR review become practical. Because SKILL.md is an open standard, the same skill works across Codex, Claude Code, Cursor, and Copilot, removing the vendor lock-in that has slowed enterprise adoption of AI coding agents.

Summary

Codex Skills are instruction packs built around a single SKILL.md file. At startup, Codex reads only each skill's name and description, about 2% of the context budget, and loads the full instructions only when a task matches. That progressive loading cuts token consumption by 80-95% compared to traditional prompting, and the open standard means the same SKILL.md runs in Codex, Claude Code, Cursor, GitHub Copilot, and Gemini CLI.

The 11 skills span the whole development lifecycle. create-plan forces a reviewable execution plan before any code is written; Superpowers (90k+ community stars) enforces TDD and automatic code review across 14 sub-skills; gh-fix-ci turns a 20-minute GitHub Actions debug into seconds with an 8-step automated flow; frontend-skill encodes design principles so pages don't look like default AI output; webapp-testing runs natural-language-described Playwright tests matched to the project's existing framework; mcp-builder guides four-stage MCP server construction; brooks-lint audits code against six classic software engineering books; pr-review automates PR reviews with multiple AI models; Planning-with-Files uses Markdown files as external memory to survive context clears; DocumentSkills parses PDFs, Word, Excel, and PPT; Context-Engineering monitors and compresses token usage to prevent amnesia in long agent runs.

A benchmark in the article shows processing 100,000 customer service requests: 15M tokens with traditional prompts, 12M with MCP, and 2-3M with Skills. Codex also uses BFS to search up to 6 directory levels and 2000 directories for SKILL.md files, keeping discovery efficient. The author recommends starting with create-plan and Superpowers, then adding targeted skills like gh-fix-ci and frontend-skill as needed.

Takeaways

— Codex Skills are instruction packs centered on SKILL.md files; Codex reads only names and descriptions at startup, using roughly 2% of the context budget.

— Progressive loading matches task descriptions to skills, saving 80-95% of token consumption compared to traditional prompts.

— SKILL.md is an open standard, portable across Codex CLI, Claude Code, Gemini CLI, Cursor, and GitHub Copilot without modification.

— Installation methods include git cloning to ~/.agents/skills, OpenAI's $skill-installer command, and community tools like npx @supercorks/skills-installer.

— create-plan forces a structured execution plan (saved as <NAME>_PLAN.md) before code writing; since January it is partially built into Codex CLI with Shift+Tab toggling Plan/Code modes.

— Superpowers, the highest-starred framework at 90k+, enforces TDD and automatic code review across 14 skills and 1 agent.

— gh-fix-ci automates an 8-step CI failure diagnosis flow, cutting a typical 20-minute debug to seconds.

— frontend-skill encodes design principles such as spacing, color, typography, and border radius to avoid generic AI-style interfaces.

— webapp-testing uses Playwright to execute natural-language-described tests and detects the project's existing test framework (Jest, Mocha, Playwright, Cypress).

— mcp-builder uses a four-stage workflow (plan, implement with FastMCP or MCP SDK, evaluate, optimize) to build MCP Servers.

— brooks-lint derives 6 decay-risk dimensions from six classic software engineering books, including The Mythical Man-Month and Code Complete.

— Planning-with-Files creates task_plan.md, findings.md, and progress.md as external memory, allowing state recovery after /clear.

— Context-Engineering monitors token usage, compresses non-core context, and manages priority to prevent long-session amnesia.

— A benchmark for 100,000 customer service requests: 15M tokens with prompts, 12M with MCP, 2-3M with Skills.

— Codex startup uses BFS traversal up to 6 directory levels and 2000 directories maximum to find SKILL.md files.

— Knowledge consolidation from Skills can improve output consistency by 300%.

Conclusions

The progressive loading mechanism is what makes multi-session tasks possible; without the 2% preview and on-demand loading, context windows would fill up long before a large refactor finishes.

The popularity of TDD-forcing skills like Superpowers is a direct acknowledgment that AI-generated code needs guardrails; the community is treating AI output as something to be corrected rather than trusted.

The token benchmark (15M vs 12M vs 2-3M) comes from a single test, not a formal benchmark suite, so treat the numbers as directional; the mechanism itself is sound regardless of exact figures.

The open SKILL.md standard reduces switching costs across AI coding tools, which likely speeds up skill ecosystem growth and makes the format a de facto standard.

Concepts & terms

SKILL.md

A Markdown file that defines a skill's instructions, trigger keywords, and prerequisites. It is the core unit of the Codex Skills format and is designed to be portable across multiple AI coding tools.

Progressive loading

A mechanism where the AI reads only the names and short descriptions of installed skills at startup (about 2% of context budget) and loads full instructions only when a task matches a skill's description, saving 80-95% of token consumption.

MCP (Model Context Protocol)

An open standard released in 2024 that establishes a standardized communication channel between AI models and external systems using JSON-RPC 2.0 over a client-server model, solving the M×N integration problem.

TDD (Test-Driven Development)

A development practice where tests are written before implementation. The Superpowers skill enforces this on AI-generated code, requiring test cases before functional code and code review after.

Source: juejin.cn ↗ Google Translate ↗ Backup ↗