amplify

Create a plan for a coding task. Interviews to extract intent, thinks deeply before planning, writes a plan file. Use when asked to "create a plan", "make a plan for", "plan out", "help me plan", or when user needs structured approach to a task.

Execute PLAN.md tasks one at a time with human oversight. Triggers on "execute", "execute plan", "execute the plan", "continue", "continue the plan", "continue to plan", "next task", "work on the plan", "pick up where I left off", "resume", or any request to make progress on PLAN.md.

Deep analysis of PLAN.md before execution. Validates against codebase reality, checks task decomposition, identifies risks and gaps. Use after create-plan, before execute-plan.

Pauses to ask focused clarifying questions when a request lacks enough information to execute correctly. Use when the task is vague, has multiple plausible interpretations, is missing acceptance criteria, or when wrong assumptions would produce significant wasted work. Don't use when the request is already well-specified, when requirements were clarified earlier in the same conversation, when the task is trivially unambiguous (e.g., fix a specific typo), or when general clarification techniques are being discussed rather than applied to a specific task.

Navigates and interacts with websites using the agent-browser CLI, including clicking elements, filling forms, taking screenshots, extracting page content, and automating multi-step browser workflows. Use when the user asks to open a URL, click something on a page, fill out a web form, take a page screenshot, scrape or extract data from a website, or automate a browser task. Don't use for Playwright, Puppeteer, or Selenium scripts; don't use for browser extension development; don't use for questions about browser APIs, JavaScript, or frontend code that don't require live browser interaction.

Refines existing code for clarity, readability, and maintainability without changing behavior, interfaces, or outputs. Use when asked to "simplify", "clean up", "refactor for readability", "reduce complexity", or polish code without adding features. Don't use when the request involves changing behavior, adding functionality, migrating to a new pattern or library, renaming public APIs, or any structural change that alters interfaces or outputs.

Creates a structured PLAN.md file for a software engineering task by interviewing the user to extract intent, thinking through the approach, and writing an actionable plan with scope, success criteria, and a sequenced checklist. Use when the user asks to "create a plan", "make a plan for [engineering task]", "plan out [feature or change]", "help me plan [implementation]", "write a plan", or when they need a written plan document before starting implementation work. Also use when the user asks to advance to the next phase of an existing plan. Don't use for general brainstorming sessions, architecture diagrams, meeting or project management planning, non-software scheduling, or when the user wants to execute a task immediately without a planning document. Don't use for "make a plan for lunch", team retrospectives, or sprint planning.

Creates or updates a SPEC.md specification document by structuring requirements, notes, or interview output into a consistent format with sections for goals, design, edge cases, security, testing, and success criteria. Use when the user asks to write a spec, create a specification, turn notes or requirements into a spec, document a feature design, or structure technical requirements. Don't use for requirements gathering or discovery interviews (use the interview skill instead), for Architecture Decision Records or RFCs, for general documentation or README files, or for implementing code based on a spec.

Creates an Ampi-ready SQLite vault from a folder of documents (md, markdown, txt, docx, doc) or an existing SQLite source table. Builds keyword search (FTS5), optional sparse semantic search, a search_schema contract, and an amplify_search_manifest so Ampi tools (search_vault_keyword, search_vault_semantic, search_vault_deep, lookup_vault_records) work immediately. Use when the user wants to build a new Ampi vault, ingest documents into a searchable SQLite database, check whether an existing vault passes the Ampi search contract, or re-build a vault from updated documents. Don't use for general SQLite database creation, non-Ampi document stores, CSV imports, querying an existing vault, or adding documents to an already-built vault without re-ingesting.

Writes production-quality @moduledoc, @doc, and @typedoc annotations for Elixir modules. Scales doc shape to module complexity: one-line docs for simple wrappers, multi-section guides with examples for core concepts. Applies ExDoc conventions including cross-references, admonitions, doctests, tables, options lists, and return contracts. Use when asked to "document modules", "add moduledocs", "write elixir docs", "doc standard", "add @doc", "improve documentation", or when Elixir modules have thin, missing, or inconsistent documentation. Don't use for README files, user-facing guides, ExDoc configuration, hex.pm publishing, mix docs generation, non-Elixir projects, or LiveBook notebooks.

Executes tasks from a PLAN.md file one at a time with human oversight, handling task splitting, clarifying questions, tests, and learning persistence. Use when the user says "execute the plan", "work on the plan", "next task", "pick up where I left off", "continue the plan", "resume the plan", or asks to make progress on a specific PLAN.md. Don't use for general task execution without a PLAN.md, creating a new plan (use create-plan instead), answering questions about what is in a plan without executing it, one-off commands or scripts with no plan file, or continuing an unrelated conversation.

Fetches and displays source code from GitHub file and directory URLs using the gh CLI. Use when the user shares a github.com/owner/repo/blob/... or github.com/owner/repo/tree/... URL, a raw.githubusercontent.com URL, a GitHub Enterprise URL (e.g. github.mycompany.com/...), or explicitly asks to read, fetch, or show the contents of a file or directory on GitHub. Supports file contents, single-line anchors (#L10), line ranges (#L10-L25), directory listings, and shallow repository clones. Don't use for GitHub pull request URLs (/pull/), issue URLs (/issues/), commit URLs (/commit/), release URLs (/releases/), GitHub Actions workflow URLs (/actions/), Gist URLs (gist.github.com), or GitHub wiki URLs (/wiki/). Don't use when the user only wants to discuss, mention, or link to a GitHub project without requesting its file contents.

Designs and builds distinctive, production-grade frontend interfaces with high aesthetic intentionality. Use when the user asks to create, design, or build a web UI from scratch, substantially redesign an existing one for visual improvement, or apply a specific design direction — including landing pages, dashboards, React/Vue/Svelte components, HTML/CSS layouts, single-page apps, or interactive artifacts. Also use when the user explicitly asks to make a UI visually striking, beautify it, give it a mood or theme, or produce a polished visual result. Don't use for: adding UI functionality without changing appearance, bug fixes to existing CSS or JS, accessibility audits, frontend performance profiling, framework migrations, writing tests for UI components, backend work that incidentally touches a frontend file, or purely structural refactors that have no visual outcome.

Creates zero-dependency, animation-rich HTML slide presentations as single self-contained files. Use when the user wants to build a presentation from scratch, convert a .ppt or .pptx file to a web-based deck, or enhance an existing HTML presentation. Don't use for Reveal.js, Marp, Slidev, Google Slides, Keynote, or PowerPoint authoring. Don't use for non-presentation HTML pages, landing pages, or dashboards.

Generates high-end, production-quality React and Next.js UI components and full pages using Tailwind CSS, Framer Motion, and premium design principles. Use when the user wants to build, design, or refactor frontend UI where visual quality matters: hero sections, SaaS dashboards, pricing pages, bento grids, navigation, forms, cards, landing pages, or interactive components with animations. Don't use for backend logic, API routes, database schemas, server-side data processing, non-React frameworks (Vue, Svelte, Angular, Astro), plain HTML/CSS without a component framework, reviewing or auditing existing UI without producing new code, accessibility-only audits, or performance profiling unrelated to UI rendering.

Coaches the user through completing a task themselves instead of doing it for them. The core signal is that the user wants to remain the actor and build understanding, not receive a finished result. Use when the user explicitly signals they want to be in the driver's seat: phrases like "guide me", "teach me", "help me learn", "I want to understand how to do this", "don't do it for me", or "let me figure it out". Don't use when the user says "show me how" or "walk me through" without pairing it with a desire to do the work themselves (these usually mean "just show me"). Don't use for requests that want a task completed directly, a bug fixed, a document produced, a one-shot explanation without back-and-forth, or general Q&A. Don't use when "help me understand X" means "explain X" without requesting interactive back-and-forth. Don't use when "help me work through" refers to debugging or problem-solving assistance, not skill-building.

Conducts exhaustive requirements-gathering interviews for software features or systems. Reads existing context, then asks structured numbered questions covering purpose, technical design, UI/UX, edge cases, security, and rollout. Use when the user says "interview me", "deep dive on requirements", "spec this out", "fill out the spec", "help me think through this feature", or when a feature needs thorough requirements elicited before a spec or plan can be written. Don't use for quick clarifying questions during an in-progress task (use ask-questions-if-underspecified instead), for writing or updating a SPEC.md document (use create-spec instead), for generating a task plan (use create-plan instead), or for general architectural discussions that don't need structured requirements output.

Builds and debugs Phoenix LiveView forms: changeset-backed validation, per-field error feedback with used_input?, form recovery on reconnect, nested forms with inputs_for and sort/drop params, file uploads with allow_upload, component forms with phx-target, and HTTP bridging via phx-trigger-action. Use when the user is working with LiveView forms, phx-change, phx-submit, to_form, changeset validation, error display, form errors not showing, dynamic add/remove fields, multi-step forms, file uploads, or form reconnect recovery. Don't use for non-LiveView forms (plain HTML, React, dead views), general Ecto changeset questions unrelated to a form, or LiveView topics that don't involve form input handling (streams, navigation, presence, JS hooks).

Implements optimistic UI patterns in Elixir Phoenix LiveView: instant client feedback before server confirmation, loading states, rollback/revert on failure, race condition guards, stream insert/delete animations, undo windows, and accessibility for async interactions. Use when the user needs LiveView interactions to feel immediate, wants to add loading indicators or spinners, needs to handle double-submit or double-click, wants to animate stream inserts or deletes, needs undo/rollback for destructive actions, is dealing with stale data or flicker under latency, or needs aria-live announcements for optimistic state changes. Don't use for general LiveView form validation or changeset errors (use liveview-forms instead), non-LiveView frontend frameworks (React, Vue, Svelte), or server-side performance optimization unrelated to perceived UI responsiveness.

Builds new MCP (Model Context Protocol) servers that expose external API tools to LLMs. Use when the user wants to create, scaffold, or implement an MCP server to integrate a third-party service or API, in Python (FastMCP) or Node/TypeScript (MCP TypeScript SDK). Also use when asked to add tools to an existing MCP server, write or run evaluations for an MCP server, or design the tool interface for a new MCP integration. Don't use for: connecting to or configuring an already-running MCP server, debugging MCP client connections, explaining the MCP protocol in general, or configuring Claude Desktop MCP settings.

Reads, writes, and applies durable project-specific memory stored in MEMORY.md at the repository root. Use when the user says "remember this", "save this to memory", "add to memory", "don't do this again", "project-memory", or "scratchpad", or when starting a session in a project that has a MEMORY.md file, or after correcting a recurring mistake. Don't use for one-off notes, temporary task context, saving to a different file, storing credentials, updating documentation files, or general note-taking unrelated to a specific code project.

Creates or updates a ROADMAP.md file for a software project by analyzing the codebase and defining vision, milestones, and a timeline. Use when the user asks to create, generate, write, update, or revise a project roadmap, or wants to communicate project direction and priorities to contributors or stakeholders. Don't use for updating a single section of an existing roadmap in isolation without touching the rest, creating sprint plans or Jira boards, writing a product requirements document, or generating architecture or design docs.

Low-level scaffolding toolkit for agent skills — init_skill.py creates a new skill directory from template, package_skill.py validates and packages a skill folder into a distributable .skill file, and quick_validate.py checks frontmatter. Use when the user asks to run, fix, or understand these scripts directly, or when skill-forge delegates to these scripts during scaffolding or packaging. Don't use for end-to-end skill authoring, writing SKILL.md content, or skill design guidance — use skill-forge for that. Don't use for README writing, AGENTS.md configuration, or general agent documentation.

Create, update, or validate agent skills — SKILL.md authoring, frontmatter writing, bundled resource planning, and three-phase discoverability validation. Use when the user says "create a skill", "make a new skill", "write a skill for", "forge a skill", "update this SKILL.md", or asks for help with skill frontmatter, skill structure, or skill validation probes. Don't use for README writing, AGENTS.md configuration, general agent documentation, or anything that isn't a SKILL.md file or its bundled resources.

Analyzes all uncommitted changes (staged and unstaged), groups them into atomic commits by logical purpose, and generates conventional commit messages with user-facing bodies. Use when the user asks to "smart commit", "group my commits", "split into atomic commits", "organize my changes into commits", or when there are many uncommitted changes across multiple files that should be separated into well-scoped commits. Don't use for a single, simple commit message request, for push/pull operations, for rebasing or squashing existing commits, or when the user just says "commit this" without asking for grouping or splitting.

Transforms a technical spec file (SPEC.md, PRD, RFC, ADR, or similar structured design document) into a narrative prose retelling that preserves every technical detail while making it engaging and story-driven. Writes output to STORY.md alongside the source file. Use when the user asks to "tell the story of this spec", "make this spec fun to read", "narrative version", "spec story", "story version of the spec", "make this spec readable", or wants to experience a spec as a flowing read rather than a structured document. Don't use for summarizing a spec into bullet points or a shorter overview, creating a new spec from scratch, converting specs to slide decks or presentations, explaining arbitrary code files that are not design documents, or when the user wants a standard technical summary rather than a full narrative retelling.

Detects and removes AI writing patterns from prose to make text sound human. Use when the user asks to clean up, edit, review, or rewrite text that sounds like AI, contains AI tells, or reads as generic filler. Use when drafting new prose with explicit instructions to avoid AI patterns or sounding robotic. Don't use for grammar fixes, spell-checking, general tone rewrites, style changes unrelated to AI patterns, code review, SEO optimization, or translation. Don't use if the user wants the writing to sound more formal or professional without mentioning AI patterns specifically.

Analyze a codebase or module namespace for structural simplification opportunities. Finds dead code, shallow abstractions, misplaced concerns, unnecessary indirection, and complected state. Produces a classified report of findings with recommended actions. Use when asked to "simplify this system", "reduce complexity", "clean up architecture", "find dead code", "flatten this", "what can I delete", or when a namespace feels too large. Don't use for single-function refactors, formatting, or performance optimization. Don't use for planning or execution — hand off to create-plan and execute-plan.

Writes and rewrites technical documentation for software projects: README files, API references, architecture docs, user guides, and inline doc comments. Use when the user wants to create, improve, or restructure documentation for a codebase, library, CLI tool, or API. Also use when the user wants to audit existing docs for accuracy, clarity, or completeness. Don't use for commit messages, PR descriptions, changelogs, blog posts, marketing copy, or ad-hoc code comments that aren't part of a documentation system.

Upgrades Elixir Mix/Hex dependencies safely, covering patch/minor bumps, major-version migrations with changelog research, and post-update verification. Use when the user asks to update deps, bump packages, upgrade a specific Hex package, or run mix hex.outdated. Don't use for: adding a new dependency, removing a dependency, resolving a dependency conflict or lockfile error without upgrading, or updating npm/pip/cargo/other non-Elixir package managers.

Guides coverage-driven test writing for existing codebases: discovers untested user-facing behavior via coverage reports, writes one meaningful test per iteration, and marks low-value code with ignore annotations. Use when asked to "write tests", "add test coverage", "increase coverage", "test this feature", or when a coverage report shows gaps in a working codebase. Don't use for setting up a test framework from scratch, writing e2e or integration test suites, explaining testing theory, or reviewing existing tests without writing new ones.