codex-skills

Audit and improve a public GitHub repository's discoverability and any linked canonical site without confusing repository presentation with website SEO. Use for repository SEO, GitHub discoverability, topics/homepage/README optimization, GitHub Pages SEO, AEO, or agent-friendly public documentation.

Create, edit, analyze, or format spreadsheets (`.xlsx`, `.csv`, `.tsv`) with `openpyxl` and `pandas`, especially when formulas, references, and formatting must survive the change. Use when spreadsheet integrity matters more than ad hoc CSV hacking.

Analyze and reshape spreadsheet, CSV, or TSV data with `pandas` and stable output files. Use when the job is aggregation, filtering, joins, metrics, or tabular transformations rather than workbook-style editing.

Create or edit `.xlsx` workbooks while preserving formulas, references, formatting, and layout. Use when workbook integrity matters more than raw data analysis.

Produce a fast plain-text transcript with the bundled CLI using the light transcription model. Use when speaker diarization is not needed.

Produce a speaker-aware transcript with diarization and optional known-speaker hints using the bundled CLI. Use when the user explicitly needs speaker labels.

Transcribe audio files to text with optional diarization and known-speaker hints using the bundled CLI. Use when the user wants text from audio or video, especially when speaker labeling or repeatable output matters.

Create short narrated videos from text scripts, simple timelines, or still images using Python and ffmpeg, with optional edge-tts voiceover. Use when the user asks for an MP4 render, narration generation, or a simple explainer video build.

Render and validate an MP4 from an existing `video-builder` spec JSON using the bundled Python pipeline. Use when a spec already exists and the job is actual rendering, not story or spec design.

Build a minimal JSON render spec for the local video pipeline from script lines, still images, or narration assets. Use when the spec artifact must exist before any MP4 render can run.

Build a Codex execution pack from a raw task or PRD: `AGENTS.md` plus `docs/PLAN.md`, `docs/STATUS.md`, `docs/TEST_PLAN.md`, and `docs/BACKLOG.md`, with optional prompt surfaces and a spec companion when the work is large enough to justify them.

Validate a raw product idea before PRD or implementation: buyer, pain, workaround, promised outcome, smallest wedge, distribution path, proof sought, kill criteria, and the next honest lane. Use when an idea still needs market or wedge truth.

Shape product work with the minimum useful framework after an idea clears basic validation. Use for assumptions, experiments, opportunity-solution trees, strategy, value proposition, or pre-mortems when the next product decision is still fuzzy.

Create a self-contained HTML explainer for systems, workflows, diagrams, timelines, comparisons, or data-heavy summaries when plain terminal text would be painful to read. Use when the output naturally wants a visual artifact instead of chat prose.

Keep substantial work coherent across long sessions or compaction with one short factual `CONTINUITY.md` covering sources of truth, goal, constraints, decisions, evidence, and done/now/next state. Use when session drift would otherwise become a real risk.

Turn a GitHub Issue into a clean human-agent control loop with canonical issue state, deterministic PRD intake, explicit machine markers, and clear handoff points. Use when GitHub should be the durable control surface instead of chat alone.

Turn a completed coding session, git diff, notes, or transcript into a development diary draft or end-of-session writeup without rewriting the story by hand. Use near the end of meaningful work when the user wants a durable wrap-up.

Build an implementation-ready spec bundle when a plain PRD is no longer enough: minimum useful contracts, schema, test plan, blueprint, gates, ADRs, boundaries, and GitHub-ready task breakdown. Use when ambiguity or architecture risk would otherwise slow execution.

Decide how much process a task deserves before coding: tiny vs substantial, chat vs GitHub, checklist vs execution brief, architecture lane vs full bundle, operator review vs council, and private diary vs public draft. Use when the shape of the work is the blocker.

Record significant architecture or execution decisions so their reasoning survives the session. Trigger on requests like record this decision, write an ADR, capture the trade-off, or supersede the old decision.

Remove obvious AI-writing tics from drafts so the text sounds credible, calm, and human. Trigger on requests like make this sound less AI, detox this draft, or clean up public-facing copy before publishing.

Turn a vague image request into a generator-ready illustration prompt for Midjourney, DALL-E, Flux, Stable Diffusion, or similar tools. Use when the user needs visual direction, prompt structure, or reference-aware image prompting rather than image generation code.

Run a multi-lens council when a product decision is fuzzy, risky, or high-impact and one answer is not enough. Use for wedge selection, pricing, launch readiness, roadmap trade-offs, go/no-go calls, or other decisions where blind spots matter more than speed.

Turn real work into public-ready artifacts such as build diaries, release notes, launch posts, case studies, or demo scripts. Use when the goal is a public or shareable artifact grounded in completed work.

Create or rewrite `README.md` for a repo. Use when the user wants a clearer landing page, setup guide, open-source intro, project overview, or README refresh rooted in the actual repo rather than generic boilerplate.

Find current external tools, services, libraries, MCP servers, APIs, or models that could solve a concrete task faster than building from scratch. Use when the user needs ecosystem research before spending build time or money.

Triage an external finding such as a post, article, repo, screenshot, video, note, or link and decide whether it matters now. Use when the user wants a verdict, not just a summary.

Use when implementing, reviewing, or planning code where Codex should avoid overengineering: challenge whether the feature needs to exist, prefer deletion, standard library, native platform APIs, existing project dependencies, one-line solutions, decision receipts, evidence-backed GitHub prior-art checks, and adoption tradeoffs before writing minimal new code. Trigger for requests like "keep it simple", "no overengineering", "do the smallest thing", "is there already a library/repo for this", "Ponytail-style", or when building parsers, widgets, integrations, CLIs, state machines, or reusable utilities that agents often overbuild. Also use when activating, auditing, benchmarking, packaging, or publishing this anti-overengineering workflow.

Use when recording a browser walkthrough, app tutorial, screencast, PDF scroll, highlighted clicks, localized demo video, or polished browser automation MP4 with Playwright and ffmpeg.

Use when Codex needs to produce or supervise a complete short-form explainer, avatar, tutorial, or YouTube-ready video package from source material, including scripted narration, optional TTS/avatar generation, code-based motion graphics, FFmpeg assembly, QA frames, timing checks, and a verified final MP4 deliverable.

Run a second-pass review on meaningful changes with a deliberately skeptical lens: bugs, regressions, security, prompt or tool abuse, data loss, and missing tests. Use after non-trivial diffs or risky agent output.

Explain a phone-first GitHub workflow using Issues as canonical task state, Projects as the operator view, GitHub Mobile as the control surface, and chat as the fast signal path. Use when the user asks how GitHub should work from a phone or lightweight operator surface.

Default execution-planning skill for almost any non-trivial repo task. Turn a raw task, feature request, PRD, or project brief into durable plan, status, and test-plan files, then hand back a concise ready-to-execute proposal.

Delegate a large read-only cross-document or cross-file analysis to the verified long-context worker while the ordinary Codex thread stays in control. Use for big audits, due diligence, large literature synthesis, policy mapping, or other tasks that genuinely need much larger context.

Stress-test OpenClaw architecture or production-config changes with an isolated council pass, contradiction check, and independent Codex review. Use when autonomy, reliability, fallbacks, or deployment risk make a one-pass answer too weak.

Split substantial work into isolated git worktrees so coding, tests, docs, review, or experiments can proceed in separate lanes without contaminating the main checkout.

Drive a real browser from the terminal with the bundled Playwright CLI wrapper. Use when the task requires navigation, form filling, snapshots, screenshots, data extraction, or UI-flow debugging without turning the job into a Playwright test suite.

Capture desktop or system screenshots with the bundled OS helpers when the user explicitly asks for a screenshot or when tool-specific capture cannot reach the target. Supports full screen, app/window, active window, and region capture.

Create or update a Codex skill with a narrow, harness-optimized contract: functional metadata, one short purpose paragraph, at most three execution steps, hard acceptance criteria, and an explicit output format. Use when the user wants a new skill, a skill rewrite, or more reliable skill behavior.

Install Codex skills into `$CODEX_HOME/skills` from the curated OpenAI list or from an explicit GitHub repo/path. Use when the user wants to list installable skills, install one or more curated skills, or install a skill from another repo.