gaxia-skills

Use when the user wants to benchmark, test, score, evaluate, or measure how well another agent skill performs: how an agent follows its workflow, makes decisions at branch points, and holds up under pressure. Also use for evidence-backed tuning guidance when a target SKILL.md path is given.

Use when creating a skill that chains or orchestrates other skills, tools, or agent roles in a fixed order. Use for workflow, pipeline, and orchestrator skills that need explicit gates, hand-offs, branch handling, or multi-phase execution from one invocation.

Use when the user wants to minimize token spend or context use for a stretch of work: "low token mode", "save tokens", "work cheaply", "ประหยัด token", "โหมดประหยัด", "อย่าเผา token", or /dont-burn-my-tokens. Run a persistent economy mode that keeps context lean, delegates only when worthwhile, and stays concise without sacrificing correctness.

Use when the user wants to LEARN or be taught a topic (not get work done) — "teach me X", "explain X so I really understand it", "I want to learn/understand X", "สอน X หน่อย", "อยากเรียนรู้/อยากเข้าใจเรื่อง X", or invokes /lemme-teach-you. Picks a learning mode (conceptual vs hands-on), builds a tailored, verified, sequenced lesson, teaches it, optionally quizzes (on the companion page or in plain text), then points at real practice. NOT for quick factual lookups, and not for doing work (use start-work for work).

Use when an attempt fails after real effort, or a never-seen-before problem gets solved with a generalizable fix, and the lesson should outlive the session: "จำไว้ว่าวิธีนี้พัง", "รอบหน้าอย่าทำแบบนี้", "remember this failure", or any dead end worth not repeating. Captures it as a reusable rule. Do not use when the user only reports routine success with no new problem or method (e.g. "วิธีนี้เวิร์คดีมาก" alone).

Use when the user asks which skill fits a job ("ไม่รู้ใช้ skill ไหน", "เลือก skill ให้หน่อย", "which skill should I use"), or when two or more installed skills plausibly own the same task. Picks one owner from names and descriptions, or declares skill-less work instead of forcing one. Not the default entry point for new work — start-work routes ordinary requests.

Use when the user delegates work to run UNATTENDED and steps away — going to sleep or AFK — wanting it carried to the given scope and the machine powered off, with a thorough resume-ready report left behind for whatever didn't pass. Triggers — "จะไปนอนแล้ว ฝากทำให้เสร็จ แล้วปิดคอมให้ด้วย", "ทำให้เสร็จข้ามคืนแล้วปิดเครื่อง", "เสร็จแล้วฝากปิดคอม/ปิดเครื่องให้ด้วย", "merge เลยไม่ต้องรอรีวิว", "finish it overnight and shut down", "ship it and power off", "shut down when it's done", or /ship-it-and-shutdown.

Use when standing up or recovering a persistent, self-hosted AI assistant on a server or container host (VPS, Railway, Render) — one that keeps memory files, a messaging channel, and an allow-list of users. Use for fresh installs, restores from a backup or export bundle, storage rescues, and wiring the agent to external services. Not for deploying a stateless app, API, or bot without persistent memory.

Use when the user starts work without specifying a workflow: a feature request, bug report, refactor, mechanical change, or vague goal. Route the work through explicit capabilities and evidence gates. Triggers on "เริ่มงาน", "ทำ X ให้หน่อย", "ช่วยแก้/ช่วยทำ", or /start-work when no more specific skill owns the request.

Use before any coding task — implementation, fix, refactor, or edit — whenever the user asks Claude to write or change code. This includes bug reports, performance issues, and feature requests that imply code changes, even when the user doesn't explicitly say "fix" or "implement". Applies four discipline rules as a pre-coding filter: state assumptions before touching anything, write only the minimum that solves the problem, touch only what was asked, and define a verifiable goal before looping. Also triggers on /think-first, or when these caution phrases appear in a coding request: "ทำให้หน่อย", "อย่าทำมั่ว", "อย่าแก้มั่ว", "don't mess it up", "be careful". These keywords only fire when the task involves writing or changing code — they do not trigger for read-only tasks (code review, explanation, analysis, translation).

Use when any skill or task needs an interactive browser screen the user can see and click — picking from visual options, reviewing mockups or lesson pages, or any choice better made by seeing than reading. Needs Node.js and a local browser. Not for text-only questions or headless setups with no browser.