JoaquinCampo

Use when building or improving HTML/CSS artifacts, React pages, dashboards, landing pages, explainers, demos, SVG diagrams, UI prototypes, or visual outputs where design quality, subject-specific treatment, responsive behavior, and render verification matter.

General Python development skill for designing, implementing, refactoring, debugging, testing, validating, and reviewing Python code. Use for Python 3.11+ projects, especially modern pyproject-based repos using uv, Ruff, pytest, mypy, Pydantic v2, pydantic-settings, Typer, and Loguru. Applies the house Python stack uniformly while discovering local contracts.

Use when the user asks to use a kanban process, track development work on a GitHub board, work from the agent board, proactively track tasks and deferred findings, runs `/kanban-process`, or when the project is already configured to use this workflow.

Use at the start of every Fable session. Fable is the brain: decompose, dispatch, judge, and talk to the user. Codex CLI does the work: exploration, bulk reads, tests, builds, implementation, reviews, and other token-heavy units.

Use when spawning Codex subagents or Codex threads and choosing among GPT 5.6 Luna, Terra, and Sol with the appropriate effort. Use this with orchestrating-subagents whenever delegation involves model overrides, reasoning effort, Explore agents, checker/fixer loops, or fan-out work.

Use when an AI agent is itself performing a pull request review (not coaching a human, not fetching the PR). Covers agent-specific tactics: context budget, tool escalation triggers, Explore subagent dispatch (single sidecar and parallel fan-out), when to emit "I cannot review this responsibly", auto-memory integration, and agent failure modes (confabulation, tool-call theater, delegated certainty laundering, evidence laundering, patch sycophancy, premature fan-out). Complements reviewing-pull-requests (universal doctrine) and pr-review (operational PR fetch + per-layer comparison).

Use when the user asks to consult the advisor, wants transcript-aware strategic guidance, wants a second opinion on ongoing agent work, provides a transcript for review, is stuck, is choosing between candidates, or wants to know whether completed work is ready. Always run the advisor as a separate GPT 5.6 Sol xhigh agent with the exact recovered advisor system prompt.

Use preemptively at the start of any session whose work decomposes into delegable units and whose main-thread context is worth protecting. A report, a large feature, a migration are examples, not the full set; the trigger is the shape (many token-heavy steps, compact results), not the task type. Also use whenever the user asks to delegate work to subagents. Invoke before context fills, not after.

GSM8K evaluation protocol: answer extraction (####, \boxed, CoT), accuracy scoring, prompt formatting, few-shot exemplars, dataset loading, pitfalls. Use when: GSM8K, grade school math, openai/gsm8k, #### delimiter, parse_gsm8k_answer, detect_answer_failure, load_gsm8k, format_chat, math benchmark scoring, gsm8k few-shot, chain-of-thought eval.

Use when implementing labeling.py, features.py, train.py, or code involving hazard/survival modeling, person-period data expansion, horizon labels, catastrophe prediction, XGBoost survival (survival:cox, survival:aft, binary:logistic), discrete-time survival, censoring, competing risks, C-index, Brier score, scale_pos_weight, or GroupKFold for sequences.

HF Transformers generate() internals — scores vs logits, LogitsProcessors, KV cache, StoppingCriteria, chat templates. Use when code calls model.generate(), output_scores, output_logits, return_dict_in_generate, GenerateDecoderOnlyOutput, LogitsProcessor, StoppingCriteria, past_key_values, DynamicCache, apply_chat_template, do_sample, or num_beams.

kvpress (NVIDIA) KV-cache compression for HuggingFace LLMs. Use when: kvpress imports, compression_ratio, press(model) context managers, StreamingLLMPress, SnapKVPress, ExpectedAttentionPress, TOVAPress, KnormPress, KV-cache eviction, token pruning during generation, or attention sink methods.

Use when writing, reviewing, or planning time series forecasting code. Triggers on: ARIMA, ETS, Theta, SARIMA, statsforecast, mlforecast, neuralforecast, XGBoost/LightGBM/CatBoost for time series, PatchTST, N-BEATS, TFT, Chronos, TimesFM, Moirai, MASE, MAPE, CRPS, temporal CV, walk-forward validation, prediction intervals, conformal prediction, data leakage in time series, demand forecasting, hierarchical forecasting, lag features, rolling features.

Use when: worktrunk, `wt` commands, `.config/wt.toml`, git worktrees for parallel agents, worktree hooks, LLM commit messages, agent handoffs, `hash_port`/`sanitize` filters, "run agents in parallel", "set up worktrees", managing multiple Claude Code sessions.

Browser automation CLI for AI agents. Use when the user needs to interact with websites, including navigating pages, filling forms, clicking buttons, taking screenshots, extracting data, testing web apps, or automating any browser task. Triggers include requests to "open a website", "fill out a form", "click a button", "take a screenshot", "scrape data from a page", "test this web app", "login to a site", "automate browser actions", or any task requiring programmatic web interaction.

Use when a program crashes, a test fails, or code produces wrong results and reading the source isn't enough to see why. Lets you pause execution at any line and inspect the actual runtime state, variable values, types, call stacks, to find what went wrong. Can attach to running servers by PID — no restart or code changes needed.

Use when spawning Codex subagents or Codex threads and choosing among GPT 5.6 Luna, Terra, and Sol with the appropriate effort. Use this with orchestrating-subagents whenever delegation involves model overrides, reasoning effort, Explore agents, checker/fixer loops, or fan-out work.

Use when deploying a FastAPI or Python ASGI/WSGI app to Vercel, choosing uv vs requirements.txt, pinning Python version, or when a Vercel URL returns "Redirecting..." / 302 to SSO. Covers the auto-detect Python runtime and serverless limits.

Context Engineering for AI Agents — 81 MCP tools, 10 read modes, 95+ shell patterns, tree-sitter AST for 26 languages. Compresses LLM context by up to 99%. Use when reading files, running shell commands, searching code, or exploring directories. Auto-installs if not present.

Use preemptively at the start of any session whose work decomposes into delegable units and whose main-thread context is worth protecting. A report, a large feature, a migration are examples, not the full set; the trigger is the shape (many token-heavy steps, compact results), not the task type. Also use whenever the user asks to delegate work to subagents. Invoke before context fills, not after.

Run a deep, didactic, citation-backed PR review and emit a single self-contained HTML report. Use when the user invokes `/review-html <PR>` or asks for an HTML review of a GitHub PR.

Use when defining the CONTENT of a slide deck or presentation for handoff to a separate rendering step (e.g. Claude Design) — turning meeting notes, a topic, or raw material into a per-slide content spec with diagram specs. Triggers include "build a deck/slides", "prep a talk", "spec the slides", "didactic technical presentation", or planning what each slide says before any visuals exist.

Default Python development and review standard for this repo. Use whenever writing, changing, refactoring, debugging, testing, or reviewing Python code; when asked for modern Python, best practices, clean code, or Zen of Python guidance; and as the final self-review rubric before declaring Python work complete. Targets Python 3.12.

General Python development skill for designing, implementing, refactoring, debugging, testing, validating, and reviewing Python code. Use for Python 3.11+ projects, especially modern pyproject-based repos using uv, Ruff, pytest, mypy or pyright, Pydantic v2, Typer, or Loguru. Adapt to the target repository before applying defaults.

Code-quality rubric for this repo (miniCOIL v2): what 'good' Python looks like here, focused on the judgment-level and subtle/version-aware issues that the linter cannot catch. Use whenever reviewing, refactoring, or cleaning up Python; when asked to follow 'best practices', 'modern Python', the 'Zen of Python', or to make code less ugly; and as the standard that the code-quality review workflow loads. Targets Python 3.12.

GSM8K evaluation protocol: answer extraction (####, \boxed, CoT), accuracy scoring, prompt formatting, few-shot exemplars, dataset loading, pitfalls. Use when: GSM8K, grade school math, openai/gsm8k, #### delimiter, parse_gsm8k_answer, detect_answer_failure, load_gsm8k, format_chat, math benchmark scoring, gsm8k few-shot, chain-of-thought eval.

Use preemptively at the start of any session whose work decomposes into delegable units and whose main-thread context is worth protecting. A report, a large feature, a migration are examples, not the full set; the trigger is the shape (many token-heavy steps, compact results), not the task type. Also use whenever the user asks to delegate work to subagents. Invoke before context fills, not after.

Pydantic v2 best practices and correct syntax. Use when code imports pydantic, uses BaseModel, Field, field_validator, model_validator, ConfigDict, TypeAdapter. Triggers on: Pydantic, data validation, model serialization, API schemas. Corrects common v1 patterns LLMs generate.

Search past decisions, failures, and experiment logs for relevant context before starting a task. Use this before any significant implementation or experiment to avoid repeating mistakes.

Analyze sweep results after completion. Computes accuracy, CFR, signal statistics, and generates comparison tables. Use after a sweep completes to understand the data.

Compare results across models (Qwen vs Llama) at matching compression configurations. Generates side-by-side tables and identifies cross-model patterns.

After completing a significant task or experiment, extract lessons learned and update the project knowledge base. Captures what worked, what failed, and what to remember for next time.

Prepare and run a KV-cache compression sweep. Loads sweep configuration, validates prerequisites, and provides the exact commands needed. Use before starting any GPU experiment.

Generate and validate video_script.jsonl from a code repository (GitHub URL or local path). Use this when running phase 1 of the repo-to-video pipeline.

Generate and validate video_script.jsonl from a specific paper URL or pasted paper content. Use this when running phase 1 of the paper-to-video pipeline.

Reference for manimlib animation primitives, transform families, and timing patterns used with self.play().

ML-Ralph autonomous agent. Start ML projects, create PRDs through conversation, run autonomous experiments. Triggers: ralph, ml project, kaggle, create prd, start ml, run experiments.

Generate an ML-centric PRD for ML-Ralph. Use when planning an ML project, experiment plan, or when asked to create an ML PRD. Triggers on: create a prd, write prd for, plan this ML feature, requirements for, spec out.

Convert an ML PRD into prd.json for ML-Ralph. Use when you have an ML PRD and need prd.json. Triggers on: convert this prd, turn this into ml-ralph format, create prd.json from this, ralph json.

Generate an ML-centric PRD for ML-Ralph. Use when planning an ML project, experiment plan, or when asked to create an ML PRD. Triggers on: create a prd, write prd for, plan this ML feature, requirements for, spec out.

Convert an ML PRD into prd.json for ML-Ralph. Use when you have an ML PRD and need prd.json. Triggers on: convert this prd, turn this into ml-ralph format, create prd.json from this, ralph json.

Use when a program crashes, a test fails, or code produces wrong results and reading the source isn't enough to see why. Lets you pause execution at any line and inspect the actual runtime state, variable values, types, call stacks, to find what went wrong. Can attach to running servers by PID — no restart or code changes needed.

Use when the user wants two AI agents to debate or discuss something together, or presents a hard "X vs Y" architecture tradeoff. Default pair is Claude vs Codex; supports claude-codex, codex-claude, claude-claude, codex-codex. Triggers: "have them debate", "ask Codex too", "discuss with another agent", "two Claudes debating". Skip simple lookups and code edits.

Create truthful, polished product demos through manual browser exploration and the agent-demo capture, render, verification, inspection, comparison, and acceptance workflow.

Use when auditing or improving a Claude Code skill. Triggers on: "audit skill", "evaluate skill", "review skill", "skill score", "skill health check", "improve skill", "does this skill follow best practices".