am-will/codex-skills
18 skills20.9k installs16.9k starsGitHub
Install
npx skills add https://github.com/am-will/codex-skillsSkills in this repo
1Frontend Designfrontend-design is an agent skill that forces a creative direction before any UI code lands, helping solo builders escape interchangeable AI-generated interfaces. When someone asks for a component, page, or application, the skill walks through purpose, audience, technical constraints, and a deliberately extreme tone—then implements production-grade markup and styles that execute that vision with precision. It is aimed at founders and indie developers who need memorable marketing sites, dashboards, or tools that feel authored rather than assembled from the same purple-gradient playbook. Use it during Validate when a prototype must communicate brand, and again in Build when you are implementing the real frontend. The emphasis is on differentiation: one memorable detail, coherent typography and spacing, and aesthetics that match the stated tone. It does not replace accessibility audits or backend architecture, but it elevates the visual and experiential layer your users judge in the first five seconds.1.3kinstalls2Frontend Responsive Design StandardsFrontend Responsive Design Standards is an agent skill package that encodes how solo builders should implement adaptive web layouts: start mobile-first, layer breakpoints with media queries, use flexible containers and relative units, and keep touch targets large enough for phones and tablets. It is aimed at indie developers shipping SaaS dashboards, marketing sites, or browser extensions where one codebase must read well from narrow viewports to wide desktops. Invoke it when you are creating or refactoring page structure, responsive grids, navigation that collapses on small screens, or any CSS that mixes fixed pixels with breakpoint overrides. The skill does not replace a design system or component library; it standardizes the procedural choices—what to optimize first, which units to prefer, and how to test mentally across device classes—so agent-generated UI work stays consistent and shippable without a separate design review every time.1.2kinstalls3Context7Context7 is a Codex-oriented agent skill that wraps the Context7 CLI to search libraries and retrieve fresh documentation snippets before your agent writes integration code. Solo builders should treat it as mandatory hygiene whenever they add or upgrade dependencies, debug vendor-specific errors, or implement features against moving SaaS APIs. The workflow is deliberately small: search by name, capture the library id, then request contextual doc passages for a focused question. Because stale training data is a common source of silent bugs, the skill explicitly tells agents to prefer Context7 over memory when versions matter. Setup requires placing CONTEXT7_API_KEY in the skill’s local .env, which agents must locate in hidden skill directories. It is journey-wide procedural knowledge—usable during Build, Ship debugging, and Operate incident fixes—any time external packages are involved.1.2kinstalls4Plannerplanner is a Codex-oriented agent skill that runs a disciplined planning ritual before implementation. When you say you want a plan, want steps outlined, or invoke /planner, the agent first researches your repository—patterns, neighbors, frameworks, and related modules—then pauses to ask clarifying questions (on the order of five to ten) about goals, scope, users, environments, integrations, auth, and validation. That ordering reduces wrong assumptions that often waste solo-builder time. The output is a phased implementation plan broken into sprints and atomic tasks, with explicit attention to edge cases and non-functional requirements. It fits indie workflows where one person both specs and ships: you get a shareable backlog the same agent (or you) can execute next. It is not a substitute for product discovery or automated code review; it front-loads structure so Build and Ship phases proceed with fewer reversals.1.2kinstalls5Read GithubRead GitHub is a Codex-oriented integration skill that lets solo builders and agent workflows fetch repository documentation through GitMCP instead of scraping GitHub by hand. A Python CLI normalizes full GitHub URLs or shorthand owner/repo strings into gitmcp.io targets, trimming branch and blob paths that would break MCP lookups. That makes it practical when you are wiring an agent to learn a library, compare fork docs, or answer questions against the latest README while you integrate a dependency. It spans build-time integration work and earlier discovery when you are still evaluating repos. Treat it as a network-backed doc reader companion to MCP hosts, not a substitute for cloning, issuing tokens, or running local code search.1.2kinstalls6Parallel TaskParallel Task is a Codex-oriented orchestration skill for solo builders who already have a markdown implementation plan and want execution split across parallel subagents instead of a single long agent session. It activates only on explicit /parallel-task commands, reads the plan file, extracts task IDs, names, depends_on edges, descriptions, locations, acceptance criteria, and validation steps, then schedules work in waves: every task whose dependencies are satisfied launches concurrently with a standardized subagent prompt. You can run the full plan or a subset of task IDs (automatically expanding to prerequisite tasks). The orchestrator’s job is ordering, delegation, and verification—not replacing specialists—while keeping plan documentation current with logs after each task finishes. It pairs naturally with plan-first workflows in the am-will codex-skills stack when specs are broken into checklisted tasks with clear dependency chains.1.2kinstalls7Plan HarderPlan Harder is a Codex-oriented planner agent skill for solo and indie builders who already have a direction but need engineering-grade execution detail. It starts with structured codebase research (architecture, similar patterns, dependencies), then uses targeted user input—up to ten questions—to nail scope, constraints, and success criteria before writing the plan. The deliverable is a phased document with an overview, ordered sprints that compound into working software, and atomic tasks with paths, tests, and parallelization hints. It sits between lightweight chat plans and jumping into implementation, making it especially useful when complexity, security, performance, or UX edge cases matter. Invoke it when you want sprint boundaries your agent can commit against, not a single blob of “implement feature X.”1.2kinstalls8Agent BrowserAgent Browser is a fast Rust-based headless browser automation CLI with a Node.js fallback, packaged so AI coding agents can drive the web through structured commands instead of brittle ad-hoc scripts. Solo builders shipping SaaS, internal tools, or agent workflows install it when they need repeatable navigation, element interaction, text extraction, and screenshots from terminal or agent sessions. Typical flows mirror the quick start: open a URL, snapshot the page, click and fill by element references, read text, capture a screenshot, then close the session. For sites that require Google, Discord, or other OAuth flows, the skill documents connecting to a real Chrome profile via remote debugging and a user data directory—critical when headless anonymous sessions cannot authenticate. Metadata notes node and npm as required binaries, aligning with the recommended npm global install and optional source build. Use it during integration work, agent prototyping, form automation, and lightweight UI verification; pair with dedicated test frameworks when you need full CI matrices rather than agent-led smoke checks.1.2kinstalls9Llm CouncilLLM Council is an agent skill for builders who want higher-quality plans without blindly trusting a single model reply. It describes a small architecture: an Orchestrator loads a task_spec JSON, spawns multiple planner CLIs in parallel via AgentRunner, validates each council_plan against a schema, anonymizes provider-identifying details, randomizes plan labels for the judge, and merges judge_output into one final_plan with warnings and metadata. Failure handling is explicit—timeouts, invalid JSON extraction, refusals, and judge errors each have retry or fallback paths so a run can complete with fewer valid plans rather than failing silently. Solo developers can use it before large refactors, security-sensitive designs, or ambiguous product decisions where a second structured pass reduces risk. It spans Validate when you are choosing scope, Build when drafting agent workflows, and Ship when you want a review gate; it is not a hosted multi-model API product itself but a pattern you wire to local planner and judge binaries. Expect intermediate complexity around JSON schemas, background shells, and operational discipline. Pair it with skills that produce the initial task_spec or implem1.2kinstalls10Markdown Urlmarkdown-url is a lightweight agent skill plus Node helper that reroutes website visits through markdown.new, a Markdown rendering proxy, so your agent reads structured text instead of wrestling with ad-heavy HTML. Whenever you would open documentation, a blog post, a changelog, or a GitHub issue, the skill requires rewriting the link: normalize to an absolute https URL, prefix with https://markdown.new/, fetch that page, and use the Markdown body in your answer or notes. Solo builders benefit across the journey because research, validation reading, implementation, and launch copy all depend on ingesting third-party pages quickly. The included script accepts a URL or host/path from argv and prints the transformed link for shell pipelines. It is not a full browser automation framework; it is a disciplined fetch pattern that pairs with network-capable agents. Use it to reduce token waste and copy-paste friction when primary sources are web pages rather than raw repo files.1.2kinstalls11Gemini Computer UseGemini Computer Use is an integration skill for solo builders and agent authors who want Gemini’s preview computer-use model to drive a real Chromium session via Playwright. The workflow is explicit: the model emits function_call actions, your client executes them in the browser, then returns function_response payloads with the latest screenshot and URL until the task completes. Supported primitives span navigation, search, clicking, hovering, typing, scrolling, and drag-and-drop—enough to exercise SaaS dashboards, onboarding flows, or staging environments during build. Configuration covers API keys and optional browser channels or executables so you can match local Chrome, Edge, or Brave instead of bundled Chromium. Because confirmations may be required for sensitive operations, the skill fits human-in-the-loop agent setups rather than unattended production bots without review.1.2kinstalls12Openai Docs SkillOpenAI Docs Skill packages a shell-first workflow for pulling authoritative OpenAI API documentation into your agent session without tab-hopping. A bash entry script exposes init, tools, search, list, fetch, endpoints, and openapi commands, building JSON-RPC payloads with jq and posting them to the configurable MCP_URL—defaulting to the hosted OpenAI developers MCP endpoint. Responses are parsed from event-stream style output so Codex or Claude Code can ground answers in current reference pages, endpoint lists, and generated client snippets. Solo builders use it when wiring chat completions, assistants, realtime, or other OpenAI surfaces and need the exact parameter names, limits, and examples from the source site. It also helps at Ship when you verify breaking changes before a release and at Launch when you document integration behavior for users. The skill is an integration helper, not a replacement for reading rate limits, data policies, or your own key management. Intermediate complexity reflects curl, jq, and MCP protocol familiarity. Keep network access scoped to the official docs endpoint you trust.1.2kinstalls13Swarm PlannerSwarm Planner is an agent skill for solo and indie builders who want implementation plans that read like a parallel work queue, not a single-threaded todo list. It follows a structured process: investigate the repo architecture and patterns, pull fresh library documentation through Context7 before locking tasks, ask scope and priority questions when ambiguity remains, and decompose work into atomic units each tagged with explicit upstream dependencies. That dependency graph is what lets multiple agents—or you plus several agent sessions—execute concurrently without merge chaos. A dedicated review subagent checks the plan for gaps before you treat it as final. Invoke it only when you explicitly want this planning ritual; it is not a passive hook. Best paired with a clear feature or refactor goal and a codebase the agent can explore. The output is a swarm-ready plan you can fan out to implementers rather than one monolithic agent thread.1.2kinstalls14Vercel React Best PracticesVercel React Best Practices packages Vercel Engineering’s performance doctrine for React and Next.js into an agent-readable rulebook. Solo builders shipping dashboards, marketing sites, or full-stack SaaS on Next use it when Lighthouse is fine but real users still wait on sequential fetches or bloated client bundles. The document ranks guidance from critical waterfall and bundle issues through incremental advanced patterns, with concrete before-and-after snippets so an agent can refactor rather than debate theory. It is optimized for automated maintenance and greenfield generation, not casual skimming. Primary placement is Build frontend, but the same rules matter during Ship performance passes before launch and when Operate incidents trace back to slow RSC or API chains.1.2kinstalls15Role CreatorRole-creator is a reference skill for defining Codex agent roles in ~/.codex/config.toml and per-role config files, grounded in codex-rs sources and the published config schema. Solo builders running multi-agent Codex workflows use it to spin up specialized roles—read-only researchers versus implementers—without stuffing unsupported keys under [agents.<role>]. The skill enforces a minimum policy profile: model, model_reasoning_effort, and developer_instructions on every role file, with sensible defaults when the user does not specify. It explains runtime merge behavior at spawn time, including that approval_policy becomes never and collab depth limits still apply, which prevents surprise permission escalation. You reach for it while configuring agent tooling in build, but the same roles support validate research, ship reviews, and operate investigations once defined. Optional sandbox, web_search, and MCP blocks stay out unless the user explicitly asks, keeping configs auditable.1.2kinstalls16Parallel Task Sparkparallel-task-spark is an orchestration skill for Codex-style agents that turns a written implementation plan into coordinated parallel work. After you trigger /parallel-task-spark, the orchestrator reads your markdown plan, extracts task IDs, dependency lists, acceptance criteria, and validation steps, then launches Sparky subagents only for tasks whose depends_on set is satisfied—running an entire wave in parallel before the next. Solo builders use it when a single agent thread would serialize hours of independent file changes across backend, frontend, and docs. It complements skills that produce the plan file but does not replace human review of diffs or security checks. Treat it as build-and-ship execution infrastructure: ideal once tasks are enumerated, risky if the plan is vague or dependencies are wrong because waves will still fire in graph order.1.1kinstalls17Super Swarm SparkSuper-swarm-spark is an explicit-command-only orchestrator for Codex-style agents. It reads a markdown plan, extracts task IDs, descriptions, locations, acceptance criteria, and validation steps, then schedules work on a rolling pool of up to fifteen Sparky subagents so new tasks start whenever a slot opens. The workflow deliberately de-prioritizes dependency graphs in favor of throughput, while compensating with rich per-task context packs and checks on each returned result. After parallel execution finishes, the orchestrator performs final integration fixes, adds or adjusts tests, runs the suite, and fixes failures. Solo builders use it when a large written plan would take days of serial agent sessions and they want continuous parallel execution with guardrails against inconsistent file naming across agents.1.1kinstalls18Tdd Test Writertdd-test-writer is a Codex-oriented agent skill that formalizes the RED step of test-driven development for solo builders who want agents to stop guessing implementation details. You invoke it when someone asks for test-first work, RED/GREEN/REFACTOR cycles, or bugfixes that need a regression spec before code changes. The workflow starts by turning the request into acceptance criteria covering happy path, edges, and negative cases, then authoring tests that fail because production behavior is absent or wrong—not because the tests are broken. It prefers focused, deterministic test commands over full-suite churn and documents ambiguity instead of silently narrowing scope. During Ship it anchors your quality gate; during Build it pairs with feature work so implementation agents have objective pass criteria. After handoff, a separate implementation pass is expected to make tests green without weakening assertions. Intermediate complexity: you need a real test runner and discipline about role separation.857installs