claude-agent-flow

Use when updating documentation, README sections, CHANGELOG entries, docstrings, or CLAUDE.md conventions after a feature is complete.

Use when the task involves server-side API routes, middleware, business logic, authentication, database queries, or infrastructure configuration.

Task and document management with Backlog.md. Use this skill whenever working with tasks, tracking work, creating subtasks, searching or viewing existing tasks, or any time work should be preserved across sessions or handed off. Triggers include: "create a task", "track this", "add to backlog", "what tasks do I have", "mark as done", "continue on task", "break this into tasks", or any request where the outcome has a concrete deliverable worth tracking. Always use this skill when a project includes a Backlog.md setup and work needs to be logged. Prefer CLI commands unless MCP tools are confirmed working in the current environment.

Use when adversarial code review is needed — actively constructing failure scenarios, edge cases, and security violations to break the implementation.

Run code review via an external LLM API using external-review.sh. Outputs structured JSON with verdict, summary, and per-file concerns.

Use when the task involves UI components, HTML, CSS, styling, React, Vue, Svelte, client-side state, accessibility, or browser APIs.

Use when reviewing code for correctness, security, performance, or style — including /review tasks, PR reviews, and post-build quality passes.

Use when the task involves database schema changes, SQL migrations, RLS policies, cloud storage buckets, file system persistence, or data access control.

Sync installed Claude Code plugin skills into the repo so they work in web environments. Run this from a local CLI session where plugins are installed.

Use when writing, reviewing, or validating tests — including unit tests, integration tests, BATS tests, and visual verification tasks.

Analyse Claude Code session token costs inline

Agent team best practices — philosophy, routing, execution model, adversarial loop, cost management, failure modes, and brief writing.

Optional web search skill using a LiteLLM-compatible grounded-search endpoint (e.g. Gemini via LiteLLM). When AGENT_FLOW_WEB_SEARCH_ENABLED=true, agents can shell out to this skill for real-time search results as an alternative to the native WebSearch tool. Outputs synthesised answer with citations (answer mode), raw citation list (search mode), or full API JSON (raw mode).

Guard skill that halts command execution if the agent-flow system is not initialized. Checks for .claude-agent-flow/sync-state.json (installed repos) or .claude-agent-flow/publish-plugin-manifest.yml (master source repo) and stops with a clear message to run /install if neither is present.

Create visually appealing tables and diagrams for display. Use when the user asks to create tables, diagrams, status displays, dashboards, or structured text layouts that should render nicely in monospace fonts. Triggers include requests for tables, terminal-style output, box diagrams, or recreating visual layouts in text form.

Technical project management for Backlog.md workstreams using coordinated sub-agents. Use only when the user explicitly asks Codex to take over one or more existing Backlog.md tasks (for example: "act as TPM", "coordinate these tasks", "delegate to sub-agents"). Orchestrate planning, implementation, and finalization; manage dependencies and overlapping code areas; review and approve sub-agent plans before implementation.

You MUST use this before any creative work - creating features, building components, adding functionality, or modifying behavior. Explores user intent, requirements and design before implementation.

Convert a Claude Code session JSONL log into an enriched visualiser JSON file for docs/public/logs/. Runs log_converter.py for the base structure, then enriches with correct agent models, brainstorming Q&A (question events), adversarial loop transitions, parallel timing, and a story-focused title/subtitle.

Default verification tool for local HTML files. Use visual-check.sh as the FIRST choice whenever you need to verify, test, or check any local HTML page — whether confirming elements render correctly, checking layout, detecting overflow, or asserting uniform widths. Prefer this over raw playwright-cli commands. Also the authoritative reference for playwright-cli setup and gotchas in this project — always load this skill before using playwright-cli directly.