chayuto

Manages the flow of scraped fuel sign images from data/ingest/ into the labeling pipeline at data/tmp/. Handles dedup (by filename and content hash), manifest registration, and ingest cleanup. Use this skill when the user mentions "process ingest", "check ingest", "new images", "move scraped images", "run pipeline", "ingest to tmp", or asks about images waiting to be processed.

Stage 1 Labeler: Annotates sign_board bounding boxes on full camera photos for Finder training. Find the physical sign panel in the photo and draw ONE bbox per image. Use this skill when the user wants to label images for the Finder model.

Plans and executes scrape campaigns for Australian fuel station price sign images. Assesses current dataset state, brand gaps, and historical yield. Use this skill when the user wants to "scrape more" or "collect images".

Manages the flow of scraped fuel sign images from data/ingest/ into the labeling pipeline at data/tmp/. Handles dedup (by filename and content hash), manifest registration, and ingest cleanup. Use this skill whenever the user mentions "process ingest", "check ingest", "new images", "move scraped images", "run pipeline", "ingest to tmp", or asks about images waiting to be processed. Also trigger when the user wants to know what's in data/ingest/, or after a scraping session when images need to be moved into the labeling queue.

Stage 1 Labeler: Annotates sign_board bounding boxes on full camera photos for Finder training. This is the FIRST stage of a 2-stage labeling pipeline. Stage 1 (this skill): Find the sign in the photo → 1 bbox per image Stage 2 (sign-crop-labeler): Read the cropped sign → fuel rows, prices, brand Use this skill when the user wants to label images for the Finder model.

ML experiment runner following the scientific method. Handles hypothesis formation, dataset building, training, evaluation on canonical test sets, cross-experiment comparison, and experiment documentation in docs/experiments/EXP-NNN format. Use when the user says "train", "retrain", "experiment", "EXP-", "mAP", or wants to evaluate model performance.

Plans and executes scrape campaigns for Australian fuel station price sign images. Assesses current dataset state, brand gaps, and historical yield before recommending queries. Use this skill when the user wants to "scrape more", "collect images", "fill brand gaps", or "get more training data".

Stage 2 Labeler: Annotates fuel rows, prices, and brand within cropped sign images. This is the SECOND stage of a 2-stage labeling pipeline. Input: Cropped sign images produced by the Stage 1 Finder model. Output: Fuel row bounding boxes, price text, fuel types, brand classification. Use this skill ONLY after the Finder model reaches mAP50 > 0.70. DO NOT use this on full camera photos — use fuel-sign-labeler (Stage 1) instead.

Write, debug, or update Playwright E2E tests for the game — including fast, frame-exact deterministic tests using seeded RNG and fixed-timestep stepping.

Drive a headless Chromium browser to take screenshots of the running dev server so Claude can VISUALLY verify its own UI/canvas work via vision-inspection. Use whenever the user asks to "visually check", "take a screenshot", "see how it looks", "verify visually", or after touching code in src/ui/, src/rendering/, src/game/, or anything that produces visible output. This is for ad-hoc visual inspection, NOT for writing Playwright test specs (use the e2e skill for that).

Safely refactor a module with full test coverage verification

Run full verification suite (lint, test, build) and report results

Analyze raw Klondike Solitaire AI export logs from this repo (data/raw/solitaire-ai-log-*.json or files dropped in /Users/chayut/Downloads/) and deliver a kill-or-continue verdict. Use this whenever the user shares a solitaire-ai-log path, asks "should I kill this session?", asks whether a board is solvable, asks to ingest a new export, or wants to update data/DATASET_NOTES.md with findings from a new harvest. The user is running an active harvest pipeline and needs the early-termination signal — not a postmortem after the session stalls further.

Refresh and push the public Hugging Face dataset chayuto/klondike-llm-decisions after the corpus grows. Use this whenever the user wants to "publish the dataset", "update HF", "push to Hugging Face", "refresh the dataset", "re-publish klondike-llm-decisions", or bump the public corpus after a harvest has been ingested. It runs the regenerate, validate, push, verify, record flow end to end: regenerate the publish shards, gate them by id+field, upload only the changed Parquet shards, round-trip verify against the Hub, then append a Push-history row to the canonical log and update the recall memory. The cached HF token is usually invalid (temp tokens get revoked after each push), so the skill expects the user to supply a fresh WRITE token inline and reminds them to revoke it afterwards. This skill is for the PUBLISH step only; bringing new harvester exports into the corpus first is the solitaire-ingest skill's job. ALSO use it for a card-only change (edit the generator, re-push just data/publish/README.md) or to

Triage and ingest the Klondike harvester's terminal and snapshot export files (solitaire-win-*.json and solitaire-game-*.json, usually paired with a solitaire-ai-log-*.json) into the training corpus. Use this whenever the user drops one or more harvester export files and wants them ingested, cataloged, or added to the dataset / corpus / manifest, or says things like "ingest these wins", "ingest WINS", "add this harvest", "catalog this game", or pastes solitaire-win- / solitaire-game- paths. It triages FIRST (terminal-win vs terminal-loss vs mid-game pending-snapshot) before any deep analysis, reads model / build / prompt attribution per file because the harvester mixes models and builds across time, and routes each session to the right ingest and DATASET_NOTES action. ALSO use this skill when the user reports that pending sessions were killed (e.g. "all killed", "I killed those") after an ingest: the operator-kill flow turns the already-ingested ai-logs into terminal loss-by-kill records and adjudicates each

Use this skill for ANY situation involving LLM prompts, response schemas, or LLM misbehavior diagnosis. Trigger eagerly — do not wait for the user to say "audit". USE IMMEDIATELY when (1) the user pastes a prompt, system prompt, response schema, few-shot example, or JSON template for any reason; (2) the user describes ANY LLM misbehavior — model loops, repeats itself, ignores instructions, keeps apologizing, drifts, picks wrong from a numbered list, hallucinates JSON keys, gives identical answers to different inputs, infinite loops, parroting; (3) the user is writing a new system prompt or response schema for production; (4) the user mentions "review/audit/critique/improve/evaluate/check/look at" anywhere near a prompt; (5) the user asks about structured outputs, JSON mode, function calling, few-shot examples, reasoning trail, history accumulation, chain-of-thought, or schema design; (6) the user names a model (Claude/GPT/Gemma/Gemini/DeepSeek/o-series/Llama) alongside a behavior problem. Produces a severity-

Run gas analysis on contract functions and report costs. Use when the user asks about gas usage, optimization, or contract efficiency.

Comprehensive security audit of SaThuCoin contract, dependencies, deployment scripts, and scraper code. Only invoke manually — never auto-trigger.

Compile Solidity contracts and run the full test suite. Use when the user asks to test, run tests, or verify changes work.

Fine-tune Thai NER transformer models (PhayaThaiBERT, WangchanBERTa, or new checkpoints). Covers full pipeline — data prep, config generation, hyperparameter tuning, training, evaluation, comparison, and deployment. Use when the user wants to train, retrain, tune, or experiment with NER models.

Deep web research on ML/NLP techniques, academic papers, HuggingFace models, and implementation best practices. Use whenever the user asks about a technique ("what is LoRA?", "how does CRF work?"), wants to explore new approaches ("what's the latest in NER?"), needs paper references ("find papers on Thai NER"), wants model/method comparisons, or is considering a new experiment and needs evidence before committing. Also trigger when the user mentions researching, benchmarking against SOTA, looking for implementations, or finding alternatives to their current approach. This skill does research only — it does NOT modify code or run training.

Analyze stored shot telemetry like a golf coach, using TrackMan metric semantics, tour and amateur benchmarks, and motor-learning practice principles

Back up or restore the local Swing-Stack database using timestamped pg_dump snapshots

Comprehensive SEO audit methodology for websites, biased toward local service and catering businesses like The Pig Baron (Sydney spit-roast catering). Use this skill whenever you audit, check, or review a site or page for SEO and need prioritised findings plus an action plan. It covers technical and infrastructure SEO, indexability (sitemap, robots.txt, noindex, canonicals, HTTPS), on-page SEO (meta titles and descriptions, H1s, alt text), content and freshness, local SEO and Google Business Profile, structured data and schema (LocalBusiness, Caterer, FAQPage JSON-LD), Core Web Vitals and PageSpeed, analytics, and internal linking. Trigger it to diagnose a ranking decline (a page 1 to page 3 drift), rewrite meta titles or descriptions, add LocalBusiness or FAQ schema, fix sitemap, robots, canonical or noindex problems, optimise a Google Business Profile for a catering or local service business, run a technical, on-page, or local SEO review, or verify the SEO of the new Pig Baron Astro rebuild. Use it even whe

Use this skill to visually verify the pig-baron-site (Astro) in a REAL browser - it drives Chromium headless to screenshot routes, check responsive layouts at mobile and desktop widths, detect horizontal overflow, catch console/page errors, confirm every route returns 200 and images load, and smoke-test the interactions (FAQ accordion, gallery lightbox, quote-form submit→success swap, mobile nav). ALWAYS reach for it - a clean `astro build` does NOT prove the pages look right - whenever the user wants to screenshot, smoke-test, "check it in a browser", verify pages render, or look at the site; whenever they ask if a page is responsive, "looks right", "looks off", or "is squished on mobile"; whenever they worry an edit broke the layout or want to "check my work"; and proactively after you touch any page, component, global.css, or styling, and before pushing or deploying - even if they never say "Playwright", "screenshot", or "visual". Prefer it over the generic local-visual-test skill for this repo. It never o

Music theory copilot for the ESP32 nursery rhyme player (project 15_nursery_rhymes). Handles melody transcription, rhythm validation, song data format, and the Kconfig build toggle system. Use this skill whenever modifying files under projects/15_nursery_rhymes/, when the user mentions nursery rhyme songs, melody notes, music transcription, song accuracy, rhythm checking, adding new songs, or fixing note data. Also trigger when the user references song_data.c, note durations, MIDI notes, or music theory in the context of this project.

High-fidelity frontend testing and visual debugging for the panic-on-rails train-track planner/simulator (React 19 + React-Konva + Zustand). Use whenever writing or debugging Playwright/E2E tests, verifying UI changes visually, capturing screenshots, doing visual regression, testing the requestAnimationFrame simulation deterministically, or using the debug bridge, store bridge, Playwright MCP, or Playwright Test Agents. Trigger on: e2e test, Playwright, visual test, screenshot, visual regression, snapshot, simulation test, deterministic test, Clock API, flaky test, debug bridge, store bridge, agentic testing, headless browser test, "smoke test", "verify in a browser", "check the canvas", consistency check.

High-fidelity frontend testing and visual debugging for the Solitaire app — always headless. Use whenever working on end-to-end tests, browser testing, the Playwright suite, the window.__solitaire test bridge, deterministic seeded games, data-testid selectors, visual verification, screenshots, driving the UI with an agent (Playwright MCP), or smoke-testing the AI advisor against the live Gemini API. Trigger on: "e2e test", "Playwright", "browser test", "smoke test", "visually verify", "screenshot the page", "test bridge", "seed the game", "is the app functional", "MCP", "agentic testing", "ask the advisor live", or any change to packages/app that needs UI-level verification.