Claude Skills by korchasa

Run all framework acceptance tests. Default is sequential (single `deno task acceptance-tests`). Use `--parallel` for 5-subagent parallel mode. Triggers on "run all acceptance tests", "acceptance-tests all", "/acceptance-tests-all".

Extract Codex system prompt from its compiled binary or JS bundle into a structured template file. Produces a document mirroring the real API request with exact prompt text, ant-only variants, feature-gated sections, and minified-variable mappings. Use when extracting, documenting, or comparing Codex prompt versions.

Fully autonomous improvement loop for flowai primitives — run the flowai SWE-bench arm, root-cause failures, generalize fixes beyond code, implement via Acceptance-Test TDD; gated by a critic subagent instead of the user. Use when asked to improve primitives from benchmark results or diagnose benchmark failures.

Autonomous investigate → root-cause → fix → verify loop for this repo's own defects. Decides first whether the product or the measuring instrument is broken, fixes that layer, proves it with the test that would have caught it, and writes what it learned back into itself. Use when a scenario, check or benchmark fails and the cause is not known.

Analyzes past AI IDE sessions (Claude Code, OpenAI Codex, OpenCode) for LLM behavior errors: hallucinated diagnostics, context decay, format loss, language leakage, and excessive reading. Use when the user asks to inspect session history, find agent errors, or build a session error index.

Run all framework acceptance tests. Default is sequential (single `deno task acceptance-tests`). Use `--parallel` for 5-subagent parallel mode. Triggers on "run all acceptance tests", "acceptance-tests all", "/acceptance-tests-all".

Fully autonomous improvement loop for flowai primitives — run the flowai SWE-bench arm, root-cause failures, generalize fixes beyond code, implement via Acceptance-Test TDD; gated by a critic subagent instead of the user. Use when asked to improve primitives from benchmark results or diagnose benchmark failures.

Autonomous investigate → root-cause → fix → verify loop for this repo's own defects. Decides first whether the product or the measuring instrument is broken, fixes that layer, proves it with the test that would have caught it, and writes what it learned back into itself. Use when a scenario, check or benchmark fails and the cause is not known.

Analyzes past AI IDE sessions (Claude Code, OpenAI Codex, OpenCode) for LLM behavior errors: hallucinated diagnostics, context decay, format loss, language leakage, and excessive reading. Use when the user asks to inspect session history, find agent errors, or build a session error index.

Give developers a repeatable, evidence-backed way to pick the right LLM for a concrete task instead of guessing from memory or a single stale leaderboard. The skill ingests a free-form task description, derives which capability axes matter (intelligence, coding, agentic-coding, reasoning, tool-use, price, speed), and returns a ranked shortlist of models with per-axis rationale and source citations. Ships in the opt-in `beta` pack as a first, experimental capability not yet promoted to `core`.

Run prompts in Claude Code, OpenCode, Cursor, or Codex CLIs from the current session — pick one IDE, fan out across several, or compare models. You are a courier that relays the other runtime's stdout verbatim, do not synthesise your own answer. Use on "run in <ide>", "compare <ide> vs <ide>", "try on <model>", "which IDE handles X better", "run across models".

Delegate a task to another AI IDE's CLI (codex / claude / opencode / cursor-agent) through an isolated-context subagent. Use when the user says "delegate to <ide>", "have <ide> do <task>", "execute <task> in <ide>", "offload to <ide>" — running that CLI from a shell yourself is the wrong path, the isolated context is the point. For one-shot relay or fan-out use `ai-ide-runner`.

Recommend which LLM model to use for a task. Use when asked "which model / best LLM for X", "pick a model for this task", or for a model shortlist ranked by live leaderboard evidence (coding, reasoning, agentic, tool-use, price, speed). Live-fetches public leaderboards and ranks models with per-axis rationale and citations.

Scaffold a complete AI agent application skeleton with LLM integration, tool calling, observability, cost tracking, session management, and content fetching using @korchasa/ai-skel-ts. Use when the user asks to create an AI agent, add LLM integration, scaffold an AI application, or mentions "ai-skel", "agent skeleton", or "AI scaffold". Works with any programming language.

Adapt project-local flowai primitives (skills, agents, AGENTS.md artifacts, hooks) to project specifics. Standalone adaptation — independent of update.

Initialize project with AGENTS.md and rules, handling both Greenfield (new) and Brownfield (existing) projects.

Reconcile project-owned AGENTS.md, CLAUDE.md, and scaffolded artifacts with the current flowai framework templates.

Configure and maintain Deno development commands (check, test, dev, prod). Use when the user wants to set up or update the standard command interface in deno.json and scripts/ directory.

Use when the user asks to plan a LARGE feature that spans multiple sessions or phases — produces an epic under the `tasks` role from AGENTS.md with dependency-ordered phases, atomic tasks, and per-phase status tracking. Use plan for single-session tasks. Do NOT trigger on generic planning, roadmap, or brainstorming requests.

Use when the user asks to diagnose a bug, find the root cause of a failing behavior, or run a controlled hypothesis-by-hypothesis investigation with experiments and evidence. Produces a diagnostic report with a recommended fix (but does not apply the fix). Do NOT trigger on "fix this bug" when the cause is obvious, or on simple error-message lookups.

Use when the user asks for a project health audit, maintenance sweep, or multi-category lead-engineer scan followed by interactive issue-by-issue resolution with user approval. Do NOT trigger on routine lint/test runs, single-file cleanups, or standard "check project" requests.

Manage Django database migrations (makemigrations, migrate, squash).

Use when the user asks to review PAST sessions (not the current one), find recurring patterns across multiple sessions, or improve project primitives (rules, skills, hooks, docs) based on historical IDE transcripts. Do NOT trigger on current-session self-review — use reflect instead.

Use when the user asks to reflect on the CURRENT session, analyze the agent's own process, audit context usage in this conversation, or extract lessons from how the current task was executed. Do NOT trigger for historical multi-session analysis — use reflect-by-history for that. Do NOT trigger on plain "what did you do?" style recaps.

Generate a .devcontainer (devcontainer.json plus optional Dockerfile) for AI-IDE development with AI CLI integration, skill mounting, and security hardening. Use when the user wants to set up or containerize a local dev environment for Claude Code, OpenCode, or flowai. Not for cloud deployment (e.g. Deno Deploy).

Supplies the accepted task format — GODS, meaning Goal, Overview, Definition of Done, Solution — and its file template. Use when the user asks what that format is, or to write up an already-decided piece of work in it. NOT for planning work before coding — weighing variants and critiquing a plan belong to the planning skill, which writes the file itself.

Run Deno from the command line with the right flags — permissions for run/test, `deno task` over raw commands, unstable-feature opt-ins, and dependencies via `deno add`/`outdated`/`remove`. Use when wiring or running any Deno command — these flags are not guessable. Not for Deno Deploy (use deploy) or npm/package.json projects.

Manage Deno Deploy cloud services using both `deno deploy` and `deployctl`. Use this skill for deploying projects, managing cloud environments (Build/Dev/Prod contexts), monitoring logs, and troubleshooting deployment issues (like private npm dependencies).

Guide for creating effective flowai commands. This skill should be used when users want to create a new command (or update an existing command) that extends flowai's capabilities with specialized knowledge, workflows, or tool integrations. Works across IDEs (Cursor, Claude Code, OpenCode).

Creation and configuration of event hooks/plugins to manage agent behavior, command filtering, auditing, and automation. Works across IDEs (Cursor, Claude Code, OpenCode). Use when you need to: (1) Create a new hook (e.g., for formatting or security checks), (2) Configure hooks/plugins, (3) Implement logic for blocking or modifying agent actions via scripts.

Guide for creating persistent AI rules (coding standards, project conventions, file-specific patterns). Use when users want to create a rule, add coding standards, set up project conventions, configure file-specific patterns, or ask about rules placement. Works across IDEs (Cursor, Claude Code, OpenCode).

Guide for creating agent-invocable Agent Skills (SKILL.md packages auto-discovered by description). Use when users want to create or author a skill (SKILL.md), or ask about skill structure or format. Not for user-invoked slash commands (a separate command-authoring concern). Works across IDEs.

Guide for creating effective Subagents (custom agent definitions). Use when users want to create a new subagent, set up task-specific agents, configure code reviewers, debuggers, or domain-specific assistants. Works across IDEs (Cursor, OpenCode).

Build and run benchmarks that measure an AI AGENT's behaviour — scenarios, checklists, judged runs. Use when authoring or evaluating agent benchmark scenarios. Not for performance benchmarks of code (deno bench, criterion) or ordinary test suites — those measure a program, not an agent.

Estimate total token usage and cost across conversation history, system prompts, and active rules. Use when the user asks how much context or how many tokens are in use, to audit context size, or to estimate session cost.

Automates browser interactions for web testing, form filling, screenshots, and data extraction. Use when the user needs to navigate websites, interact with web pages, fill forms, take screenshots, test web applications, or extract information from web pages.

Multi-directional deep research with sub-agents, scoring, source verification, and cited synthesis. Use for comprehensive research, deep analysis, or evidence-backed answers. Produces executive summary + markdown report. Supports search tools, playwright-cli, and MCP.

Use when a flowai benchmark fails and you need the cause from run artifacts before editing. Reads judge-evidence.md, the sandbox SKILL.md, and scenario mod.ts; classifies the failure against a known taxonomy; produces an evidence-grounded report (no fixes). Do NOT trigger for passing benchmarks or generic skill iteration.

How to conduct a Q&A session with the user

How to conduct a Q&A session with the user. Canonical reference for FR-UNIVERSAL.QA-FORMAT.

Split independent work items and run them in parallel through subagents. Use when a request names several areas that do not share files.

Split independent work items and run them in parallel through subagents. Use when a request names several areas that do not share files.

Draw and edit Mermaid diagrams in Markdown. Use when the user wants to visualize processes, flows, sequences, or asks for diagrams. Load it before writing any Mermaid block, including when you already know the syntax — knowing how is the usual reason it gets skipped.

Design and build complete AI-IDE plugins that bundle multiple components (skills, MCP tools, hooks) with packaging, manifests, marketplaces, and per-IDE validation. Use when the user wants to create, design, or package a multi-component AI-IDE plugin. Not for authoring a single standalone skill (use the skill-authoring skill instead).

Build AI-IDE plugin hook elements - shared policies plus Claude Code/Codex/OpenCode/Cursor adapters for lifecycle events, blocking, audit, and context injection. Use when adding or configuring plugin hooks.

Design AI-IDE plugin marketplaces - requirements, host-specific payloads, local dogfood, release sync, and validation. Use when the user wants to create or publish a plugin marketplace.

Build AI-IDE plugin MCP elements - stdio JSON-RPC servers, tool schemas, host wiring for Claude Code/Codex, event-name mapping, and validation. Use when building or adding an MCP server to a plugin. Not for plugin manifests or packaging when the MCP server already exists (that is a separate packaging concern).

Write or fix a prompt for a fast/cheap model (Gemini Flash, GPT-4o Mini, Haiku) so its output is stable and parseable. Use for any request to draft, tune or debug such a prompt — producing it IS the work, even when you could write one unaided. Not for reasoning models (use engineer-prompts-for-reasoning), model choice, or prompting a person.

Write, structure or fix a prompt for a reasoning/smart model (Gemini Pro, GPT-4o, Claude Sonnet) so it holds context across a multi-step task. Use for any request to draft, structure, tune or debug such a prompt — producing it IS the work, even when you could write one unaided. Not for fast models (use engineer-prompts-for-instant).

Creates interactive HTML teaching materials with clickable state diagrams. Use when the user asks to produce an explorable tutorial artifact, not to answer a question.