Ivan Charapanau

Use when the user requests integration testing, feature validation, or test plan execution

Use when the user wants to systematically fix AI code slop — duplicated logic, over-engineering, silent error swallowing, convention drift, cargo-cult patterns, and other LLM-introduced architectural decay — over a specified duration

Produce a researched long-form article from a topic prompt via an orchestrated pipeline - research agent (first-person sources, working-definition gate), narrative-architecture outline, writer/cold-reviewer loop with an explicit ACCEPT/REVISE verdict contract, then a catalog-deslop pass with a regression gate. The orchestrator dispatches subagents only; the writer never judges its own draft. Use when the user says "article factory", "write an article about X", "run the article pipeline", or asks for a researched long-form piece produced end-to-end. For essays and micro posts in the user's own voice without a research stage, use the prose skill instead.

Runs autonomous keep/discard experiments on a codebase to optimize a single metric for a fixed duration, in the style of karpathy/autoresearch. Use when the user says "autoresearch" (optionally with a focus, e.g. "autoresearch the optimizer"), asks to run experiments on a repo overnight, to hill-climb or optimize a metric autonomously, or points at a repo with a karpathy-style program.md.

Create custom modules for [Harbor Boost](https://github.com/av/harbor/tree/main/boost), an optimizing LLM proxy. Use when building Python modules that intercept/transform LLM chat completions—reasoning chains, prompt injection, structured outputs, artifacts, or custom workflows. Triggers on requests to create Boost modules, extend LLM behavior via proxy, or implement chat completion middleware.

Systematically explore and test any software project (CLI, API, Backend, Library, etc.) to find bugs, usability issues, and edge cases. Produces a structured report with full reproduction evidence (exact commands, inputs, logs, and tracebacks) for every issue.

Fully autonomous bug hunting pipeline — discover bugs in a scoped area using parallel subagents, independently triage each finding, fix confirmed issues with subagents, then audit all fixes against repo constraints and target platforms. Runs end-to-end without user interaction.

Catalog-then-fix slop removal on an existing prose draft. Parallel catalog subagents inventory four slop categories (theatrics, overwrought register, corrective antithesis, dramatic short sentences), fix subagents repair only what a catalog names, mechanical lint + frequency analysis bracket the run, and a regression gate (lint delta, word count, cold meaning check) decides whether the result ships. Use when the user says "deslop this draft", "catalog deslop", "strip the AI patterns from this file", or as the final stage of the article-factory and mosaic-writing skills. For writing or redrafting prose from scratch, use the prose skill instead - this skill only repairs an existing draft in place.

Bulletproof agent operating protocol. 15 failure-prevention rules distilled from 120+ real sessions and 10 agent definitions. Covers fabrication, constraint tracking, verification, scoping, retry discipline, and communication. Load before any task to prevent the most common agent failure modes.

Restate content as a flat list of atomic facts in the style of av/facts fact sheets. Use when the user says "just facts", "facts only", "as facts", "give me the facts", or replies with a bare "facts" / "facts?" to a previous message — then the scope is that message's content.

Timeboxed ideation on a topic using propose-and-critique subagent pairs. Use when the user wants to brainstorm, explore ideas, discover features, generate options, or think through possibilities for a specified duration. Triggers on requests like "brainstorm X for 30 minutes", "ideate on X", "spend an hour thinking about X", "what features should we build", "explore options for X".

Improve an existing agent SKILL.md through one gated cycle: select the target (default this skill), snapshot the committed baseline, diagnose against skill-design-principles plus four rubric dimensions, apply exactly one focused change, and ship only if a different agent than the author PASSes with quoted evidence. Use when the user says "improve this skill", "improve-skill", "iterate on SKILL.md", or "apply skill-design-principles to a skill". Do not use for code slop over a duration (anneal), generic timeboxed work (timeboxed-iterating), graph orchestration (workgraph), metric hill-climb on a codebase (autoresearch), or stripping slop from existing prose (catalog-deslop).

Build a HyperFrames video composition from a brief/script autonomously. Uses separated builder and critic subagents — builders never review their own work. Quality decisions are grounded in project context (identity, audience, goals) and established visual/motion design principles. Measurable criteria (text size, contrast, overflow) are verified with tooling, not visual judgment. Use when the user says "make this video", "implement this brief", "build this composition", or provides a video script/brief to implement.

Assemble a finished piece from human-written fragments via an orchestrated two-loop pipeline - an assembly loop (assembler + cold reviewer) that sequences the fragments into a draft, then a polish loop (polisher + cold reviewer), then frequency analysis and a catalog-deslop pass with a regression gate. Two modes with opposite fidelity contracts - stitch mode preserves the author's exact wording, narrative mode preserves every idea but may rewrite freely. Use when the user says "mosaic", "assemble my fragments", "stitch these notes into a post", or provides a fragments file to turn into a finished piece. For writing from a topic with research, use article-factory; for prose from scratch, use the prose skill.

Build-less ESM frontends that run directly in the browser without bundlers. Use this skill when creating static frontends, SPAs without build tools, prototypes, or when the user explicitly wants no Vite/Webpack/bundler. Covers import maps, CDN imports, cache-busting, hash routing, and performance patterns.

Use when pulling or downloading a new llamacpp model. The active ROCm image (kyuz0/amd-strix-halo-toolboxes) fails to start in the ephemeral pull container without ROCm device access. Must temporarily switch to the standard CPU image.

Set up and run local LLMs with Harbor. Use when the user wants to run models locally, install Harbor, start Open WebUI, llama.cpp, Ollama, vLLM, Docker Model Runner, MLX, or oMLX, pull GGUF or HuggingFace models, add SearXNG web search, Speaches TTS/STT, or Open Terminal code execution, launch Codex/Claude/Grok/OpenCode against a Harbor backend, or troubleshoot GPU, VRAM, and service startup.

Orchestrate AI coding agents (Claude Code, Codex, OpenCode) in isolated sandboxes using the @ai-hero/sandcastle SDK. Use when the user needs to (1) run agents AFK in Docker/Podman containers, (2) build multi-agent pipelines with plan-execute-review patterns, (3) run parallel agents on separate worktrees, (4) create iterative agent loops with maxIterations, (5) extract structured output from agent runs, (6) set up sandcastle in a new or existing project, or (7) write prompt files with template args and shell expressions.

Configure and operate the Claude Code harness for large codebases. Builds CLAUDE.md hierarchies, scoped test/lint commands, file exclusions, codebase maps, hooks, skills, subagent strategies, and LSP/MCP wiring. Use when setting up Claude Code for a new repo, auditing an existing configuration, onboarding a team, or scaling from single-developer to org-wide deployment. Triggers on "set up Claude Code for this repo", "optimize my Claude Code config", "audit my CLAUDE.md", "make this codebase navigable", "configure hooks/skills/plugins".

Run a task iteratively over a user-specified duration by dispatching subagents. The orchestrator stays strictly linear and time-checked, but each iteration fans out MULTIPLE subagents in parallel over independent units (and those subagents may fan out further). Shared state — a compact progress digest and a growing environment cheatsheet — lives on disk and is passed to every subagent BY REFERENCE, killing the per-subagent rediscovery tax. Every subagent first reads an initialiser preamble that points it at that shared state and makes it write findings back, so the cheatsheet populates itself; role prompts are produced ONCE on the filesystem by a single scaffold command and dispatched by path, never re-typed per dispatch. Subagents persist their results to disk and return only a tiny status, so the orchestrator's context stays small and lasts. The orchestrator classifies the goal as FINITE or OPEN-ENDED: finite lists may finish early; open-ended goals use the full duration, never idle, never manufacture busywork. Use when the user gives a task and a duration and wants it ground out iteratively by subagents over that time.

Deep learning framework development with tinygrad - a minimal tensor library with autograd, JIT compilation, and multi-device support. Use when writing neural networks, training models, implementing tensor operations, working with UOps/PatternMatcher for graph transformations, or contributing to tinygrad internals. Triggers on tinygrad imports, Tensor operations, nn modules, optimizer usage, schedule/codegen work, or device backends.

Install, configure, and work with Turso DB — an in-process SQLite-compatible relational database engine written in Rust. Use when the user needs to (1) install Turso DB, (2) create or query databases with the tursodb CLI shell, (3) use Turso from JavaScript/Node.js via @tursodatabase/database, (4) work with vector search or embeddings in Turso, (5) set up full-text search with FTS indexes, (6) configure transactions including MVCC concurrent transactions, (7) enable encryption at rest, or (8) use Change Data Capture (CDC) for audit logging.

Orchestrate work as a cyclic directed graph of subagent-executed nodes with transition criteria on edges, budgeted cycles, and per-node gates (runtime verification, metric, or artifact). Use when the user says "workgraph", "run this as a graph", "graph this work", or when a goal has parallel branches, feedback loops, or mixed gate types that a linear loop cannot express. Every node is sized to a 64k context budget, so small goals run as a two-node do/verify graph and large goals scale out without caps. For clock-driven or metric-driven single loops, timeboxed-iterating and autoresearch remain lighter.

Use when the user requests integration testing, feature validation, or test plan execution

Systematically explore and test any software project (CLI, API, Backend, Library, etc.) to find bugs, usability issues, and edge cases. Produces a structured report with full reproduction evidence (exact commands, inputs, logs, and tracebacks) for every issue.

Bulletproof agent operating protocol. 15 failure-prevention rules distilled from 120+ real sessions and 10 agent definitions. Covers fabrication, constraint tracking, verification, scoping, retry discipline, and communication. Load before any task to prevent the most common agent failure modes.

Manage .facts files — atomic, validatable truth statements about a project. Install, check, list, add, edit, remove, and lint facts via the CLI. ALWAYS read this skill when the user mentions facts in any capacity.

Scan the codebase and classify every fact by lifecycle stage — tag @draft, @spec, or @implemented based on what the code actually shows. Add missing facts, fix inaccurate ones, remove obsolete ones. Use when asked to discover facts, bootstrap or update a fact sheet, scan the codebase for truths, sync facts to match the code, or audit the fact sheet for accuracy.

Operate on @spec facts — implement them in code, then tag @implemented. Use when asked to implement facts, implement the spec, build from the fact sheet, make facts true, or work through unimplemented requirements.

Operate on @draft facts — collaboratively refine them into precise, actionable @spec facts. Resolve ambiguities, fill gaps, eliminate contradictions, and sharpen labels until every fact is ready to implement. Use when asked to refine facts, clarify the spec, review facts for quality, or "work on facts" with the user.

Use when the user specifies a task and a duration, and the work should be done iteratively by subagents over that time period