skills

Preferred first-step workflow for creating a UI from whole cloth. Use when asked to create, design, build, prototype, mock up, or redesign a website, web app, dashboard, landing page, product screen, internal tool, or application UI without an approved visual direction. Generates exactly three image concepts, stops for selection or iteration, then converts the approved direction into a reusable HTML/CSS style guide and representative demo. Use before frontend implementation unless the user supplies an existing design, requests a narrow edit, or explicitly skips visual exploration.

Use when building Django performance regression coverage with pytest-benchmark for ORM, views, schema, transactions, or backend behavior. Match the project's production database engine and topology; never prescribe one. Excludes one-off profiling, browser rendering, APM, and non-Django services. Return deterministic scenarios, correctness and query/operation guards, calibrated budgets, reports, and baseline maintenance.

Use when creating, updating, validating, or reviewing a site-root `llms.txt` file as a concise map to canonical, LLM-friendly content. Trigger only when `llms.txt` is explicitly in scope, not for robots.txt, sitemaps, SEO metadata, READMEs, OpenAPI documentation, or AGENTS.md. Return a proposal-aligned file or evidence-backed review with useful links and validation notes.

Use when building, wiring, or debugging asynchronous OpenAI Responses API workflows in Django: background requests, signed webhooks, response correlation, persistence, and HTMX status polling. Do not use for generic OpenAI calls, non-Django consumers, unrelated webhook providers, or frontend-only polling. Return scoped Django code or a plan covering security, lifecycle, failures, and validation.

Use when designing or implementing HTML-first web interfaces with native browser features, progressive enhancement, server-rendered updates, and minimal JavaScript. Covers forms, dialogs, popovers, disclosure, URL state, and replacing SPA-style interactions. Excludes visual direction, framework architecture, games, and browser debugging. Produces repo-specific code or a plan with accessibility, fallbacks, and validation.

Use when planning, auditing, or executing a Django 5.x to Django 6 upgrade. Covers runtime and dependency compatibility, deprecations, settings, databases, migrations, tests, deployment, and feature adoption. Excludes generic Django work, earlier upgrades, frontend styling, and unrelated maintenance. Produces a repo-grounded checklist, risk report, or patch with verified requirements, staged validation, and rollback guidance.

Use when integrating or refactoring Tailwind CSS and DaisyUI in a Django application. Covers the asset pipeline, template and component boundaries, theme tokens, optional theme switching, migration, and verification. Excludes isolated template edits, non-Django work, visual direction, and backend-only tasks. Produces repo-specific code or a plan that preserves existing conventions and identifies ownership and checks.

Audit or refactor a repository so coding agents can discover its purpose, architecture, constraints, workflows, validation, and sources of truth. Use when essential knowledge is hidden, stale, or unenforced. Do not use for general code review, marketing-only README work, product docs, or narrow implementation. Produce a prioritized audit or proportionate repo-local maps, docs, links, and checks.

Use when configuring or troubleshooting PyPI publishing for PDM packages with GitHub Actions and Trusted Publishing, including release events, OIDC permissions, builds, and repository/PyPI settings. Do not use for applications, non-Python projects, generic CI, containers, or token uploads unless requested. Return scoped workflow changes, required settings, validation results, and external steps.

Use when reviewing an OpenAI Agents SDK design, prototype, eval harness, or deployment plan for production readiness: architecture fit, tool safety, approvals, failures, evals, observability, rollout, cost, privacy, and human oversight. Do not use as the source for current SDK syntax or for non-agent systems. Return evidence-backed, severity-ordered blockers, gaps, and hardening actions.

Use when creating, updating, validating, or wiring Codex skill icons, including the small SVG, 100x100 PNG, and `agents/openai.yaml` fields. Covers generation, conversion, repair, metadata wiring, and visual checks. Do not use for app icons, favicons, badges, unrelated artwork, or general skill authoring without an icon task. Return correctly placed assets, skill-relative paths, deterministic checks, and rendered verification.

Design, implement, or review typed LLM output with constrained generation, validation, bounded retries, and failure handling. Use when extraction, classification, or generated objects feed code. Do not use for ordinary prompts, unstructured chat, non-LLM validation, or API-key setup. Produce a schema and pipeline, tests, and production risks.

Use when the user asks for a broad review of code elegance, beauty, developer joy, cognitive clarity, idiomatic language fit, naming, locality, or contemplative craft in one project. Evaluate code as a humane maintenance experience rather than a defect inventory. Do not use for ordinary bug, security, small-diff, or frontend visual reviews, or for deep-codebase-review unless the user explicitly combines them.

Use when reviewing an entire codebase, architecture, technical debt, structural quality, cleanup opportunities, or a substantial PR or branch. Produce evidence-backed findings about boundaries, lifecycle and concurrency, duplication, tests, plans, and maintainability. Do not use for small diffs, narrow bug fixes, frontend visual QA, repo-onboarding documentation, or OpenAI Agents SDK production-readiness review.