Codex AI Skills
AI skills for OpenAI's autonomous coding agent
Codex is OpenAI's cloud-based coding agent that executes tasks in sandboxed environments. AgentPowers skills extend Codex's capabilities with specialized tools for code generation, testing, deployment automation, and domain-specific development workflows.
Top Codex Skills
Canadian Bookkeeping Skills for AI Agents: CRA Mileage Log, T2125 & GST/HST
Four Agent Skills that help your AI assistant do Canadian sole-proprietor bookkeeping using CRA's rules, with current CRA rates, each with its source and as-of date, from the Tallyfern Rates MCP. - CRA mileage log: CRA trip fields, business-use %, the 3-month logbook test, current per-km rates. - T2125 expense sorting: turns bank, Etsy or Shopify CSVs into T2125 lines, applies the 50% meals limit, flags unclear rows. - GST/HST $30,000 check: CRA's small-supplier tests, and when you must register. - Year-end package: a tidy folder for your accountant or tax software. Works with Claude, Claude Code, Codex and Cursor. Single-buyer licence, no redistribution. Not tax advice. Designed by Tallyfern in Ontario with the help of AI tools. The examples are tested by automated checks, and every CRA rate is tested against CRA's published figures before release.
local-jev
Make Laya trustworthy. Fit its confidence to your own labeled data, keep the decisions it gets right local and free on Apple Silicon, and send only the ones it does not know to hosted Jev (TypeSafe System One) or any OpenAI-compatible LLM. Includes a Jev-compatible local server (laya-mlx), per-type temperature calibration with before/after ECE and accuracy, confidence thresholds for a target accuracy, an escalation-rate and cost report, a fail-open TypeScript decide() client, and a playbook to find LLM calls that are really classifications. Use for ticket routing, yes/no gates, rubric scores, replacing prompt-and-parse LLM classification, cutting Jev or LLM spend, and private offline typed decisions.
Notice Trap Pack
Build a SaaS vendor inventory from your CSV (or your own Ramp), catch renewals, score Switch / Renegotiate / Waste with real $, and get an approval-ready vendor counter — never auto-sends. Inventory → score → approval-ready draft. Never auto-emails vendors. Includes vendor-ops-setup, build-spend-inventory, score-savings-opportunity, research-alternatives, draft-vendor-counter, google-sheets-final-check. Missing evidence stays LEAD. Start free: Notice Trap Lite — https://agentpowers.ai/skills/notice-trap-lite
BrandLoom: Project Logos & Covers
Need a logo and cover that feel like they belong to the same project? BrandLoom helps you turn your project idea into a clear, consistent visual set. It reads the project information, files, links, and brand materials you provide, then asks one simple question at a time. You choose the wording, style, fonts, characters, layout, language, and image size before anything is created. BrandLoom can make a square project logo image, a wide cover, and a GitHub social preview in Chinese, English, or both. It can reuse your existing company logo and approved brand assets, keep important text clear, save new versions without replacing old ones, and help with later edits, translations, and style changes. It is made for solo founders, indie developers, AI builders, creators, and anyone launching a GitHub project, Skill, app, or digital product. Made for Codex. Full image creation requires Python 3.12 or newer and access to Codex's built-in image tool.
What Codex skills can do for you
Cloud-Native Execution
Codex runs tasks in sandboxed cloud environments. Skills integrate seamlessly, adding specialized capabilities without compromising the secure execution model.
Autonomous Task Completion
Codex works autonomously on multi-step coding tasks. Skills give it domain-specific knowledge to handle complex workflows from planning through implementation.
OpenAI Ecosystem Integration
Skills work with OpenAI's tool-use and function-calling features. Integrate them into your existing OpenAI-powered applications and workflows.
Codex Skills
Canadian Bookkeeping Skills for AI Agents: CRA Mileage Log, T2125 & GST/HST
Four Agent Skills that help your AI assistant do Canadian sole-proprietor bookkeeping using CRA's rules, with current CRA rates, each with its source and as-of date, from the Tallyfern Rates MCP. - CRA mileage log: CRA trip fields, business-use %, the 3-month logbook test, current per-km rates. - T2125 expense sorting: turns bank, Etsy or Shopify CSVs into T2125 lines, applies the 50% meals limit, flags unclear rows. - GST/HST $30,000 check: CRA's small-supplier tests, and when you must register. - Year-end package: a tidy folder for your accountant or tax software. Works with Claude, Claude Code, Codex and Cursor. Single-buyer licence, no redistribution. Not tax advice. Designed by Tallyfern in Ontario with the help of AI tools. The examples are tested by automated checks, and every CRA rate is tested against CRA's published figures before release.
local-jev
Make Laya trustworthy. Fit its confidence to your own labeled data, keep the decisions it gets right local and free on Apple Silicon, and send only the ones it does not know to hosted Jev (TypeSafe System One) or any OpenAI-compatible LLM. Includes a Jev-compatible local server (laya-mlx), per-type temperature calibration with before/after ECE and accuracy, confidence thresholds for a target accuracy, an escalation-rate and cost report, a fail-open TypeScript decide() client, and a playbook to find LLM calls that are really classifications. Use for ticket routing, yes/no gates, rubric scores, replacing prompt-and-parse LLM classification, cutting Jev or LLM spend, and private offline typed decisions.
Notice Trap Pack
Build a SaaS vendor inventory from your CSV (or your own Ramp), catch renewals, score Switch / Renegotiate / Waste with real $, and get an approval-ready vendor counter — never auto-sends. Inventory → score → approval-ready draft. Never auto-emails vendors. Includes vendor-ops-setup, build-spend-inventory, score-savings-opportunity, research-alternatives, draft-vendor-counter, google-sheets-final-check. Missing evidence stays LEAD. Start free: Notice Trap Lite — https://agentpowers.ai/skills/notice-trap-lite
BrandLoom: Project Logos & Covers
Need a logo and cover that feel like they belong to the same project? BrandLoom helps you turn your project idea into a clear, consistent visual set. It reads the project information, files, links, and brand materials you provide, then asks one simple question at a time. You choose the wording, style, fonts, characters, layout, language, and image size before anything is created. BrandLoom can make a square project logo image, a wide cover, and a GitHub social preview in Chinese, English, or both. It can reuse your existing company logo and approved brand assets, keep important text clear, save new versions without replacing old ones, and help with later edits, translations, and style changes. It is made for solo founders, indie developers, AI builders, creators, and anyone launching a GitHub project, Skill, app, or digital product. Made for Codex. Full image creation requires Python 3.12 or newer and access to Codex's built-in image tool.
Cognitive Anchor Sketcher for Chinese Articles
Turn Chinese articles into hand-drawn visuals that explain the idea instead of simply decorating the page. Cognitive Anchor Sketcher reads the article, finds the strongest judgments, turning points, structures, states, and metaphors, then builds a focused shot list before image generation. Each illustration communicates one idea in a clean 16:9 composition with a white background, black hand-drawn linework, generous whitespace, and short Chinese annotations. Use it for newsletters, blogs, Notion pages, product thinking, workflow documents, and methodology content. Choose between minimal-line and emotion-doodle styles, use the included Xiaohei, Tuotuo, and Xingbi characters, or work with an authorized custom IP. A guided QA flow confirms the direction, style, IP, shot list, and output settings. It can also provide cognitive-anchor analysis and illustration planning without generating images. Image generation requires a ChatGPT-authenticated Codex CLI session.
Popular Platform Data Collection Skill
Popular Platform Data Collection Skill is an agent workflow for authorized content research, content archiving, and first-party account backups. It guides users to manually export public, owned, or clearly authorized videos, cover images, page copy, tags, engagement data, and comments through DATATOOL. It validates the input files, uses local OpenAI Whisper to transcribe the complete video audio track, anonymizes comment data when requested, and produces a verifiable Chinese ZIP package containing Excel files, text, subtitles, segmented JSON, and a SHA-256 manifest. Users handle DATATOOL installation, login, verification, authorization, and content export in a normal browser. The Skill does not control DATATOOL, store cookies or passwords, bypass captchas, paywalls, private-content restrictions, or anti-bot controls. Videos are transcribed locally by default and are not uploaded to third-party transcription services.
Project Sentinel - Multi-Project Development Progress Tracker
ENHE Project Sentinel is a local-first, read-only Skill for tracking multiple development projects. It reads authorized Git repositories, project documents, reports, quality signals, and available Codex usage data to generate bilingual project portfolio reports. Reports include progress, elapsed development time, ETA ranges, risks, quality signals, Codex context, token usage, next actions, and agent-ready prompts. It is designed for indie developers, AI product builders, product owners, and small R&D teams. Supports Windows 10/11. By default, it does not modify project code or upload source code, and it clearly marks insufficient evidence instead of inventing precise ETAs.
Verify Paid Work Before Starting
Paid-looking work can fail before implementation: the payer may be unverified, the issue may already be saturated, the rules may prohibit the submission, the requested result may be impossible to validate, or payment may depend on an undefined review. This gate instructs the agent to prove the commercial path before it commits time, credentials, reputation, or money. It checks buyer authority, amount and payment rail, assistance and eligibility rules, full deliverability, repository safety, competition, acceptance evidence, and time-to-cash. Four competitive levers produce a 0-100 score, but hard safety and payability gates always win. The result is one clear PROCEED or BLOCK decision with mandatory scope, exact blockers, a proof plan, the next bounded action, a stop condition, and the official evidence that would justify reconsideration.
Common questions about Codex skills
How do skills work with Codex?
Skills integrate via the MCP protocol with Codex's tool system. They extend Codex's capabilities with specialized, verified functionality for specific development tasks.
Do I need an OpenAI subscription to use these skills?
You need an active OpenAI subscription that includes Codex access. The AgentPowers skill license is purchased separately from your OpenAI subscription.
Are skills tested with Codex's sandbox environment?
Yes. Skills are tested to work within Codex's sandboxed execution model, ensuring compatibility with its security constraints and cloud-native architecture.
Connect AgentPowers to Codex
Ready to supercharge your
Codex workflow?
Browse verified skills and agents compatible with Codex. One-command install, instant results.