菜单
QA the Hebrew Codenames engine — run the clue-legality regression, the clue-quality benchmark against the feedback set, and optionally verify the live HF Space. Use before committing engine changes, after tuning scoring or legality, or when asked to "run QA", "run the tests", "run the benchmark", or "check the engine still works".
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Test-driven development. Use when the user wants to build features or fix bugs test-first, mentions "red-green-refactor", or wants integration tests.
Apply vertical slice architecture to organize code by feature instead of by technical layer. Use when adding a new feature, endpoint, command, query, or use case. Use when deciding whether to create new files or edit existing ones. Use when implementing CQRS, mediator pattern, plugin architecture, event-driven patterns, or auto-discovery registration. Use when the user asks about additive architecture, feature folders, or minimizing edit surface. Also trigger when you notice yourself about to add a new method to an existing controller or service — that is the moment this skill is most needed.
Use when the user wants a hands-free loop where grabbing UI elements in the browser with React Grab feeds tasks to the agent automatically, with no copy-paste or manual handoff. Triggers: "watch react grab", "monitor my grabs", "auto-process react grab", "watch my clipboard for grabs". Not for a one-off paste of a single grab; this is the continuous, always-on loop.
Deploy the Hebrew Codenames co-pilot to its HF Space (shmulc/hebrew-codenames-copilot) by uploading changed files from hf_space/, and report the build stage. Use when asked to "deploy", "push to HF", "ship the Space", or "release the co-pilot".
Pull and summarize user feedback (👍/👎 on clues) for the Hebrew Codenames co-pilot from the private HF dataset shmulc/codenames-feedback. Use when asked to "check feedback", "any new feedback", "grab feedback", review clue thumbs-up/down, or see what users complained about.
Use this skill for playing, teaching, or moderating the Hebrew version of Codenames (שם-קוד). Triggers include requests to explain rules, generate clues, validate guesses, simulate games, or clarify legal/illegal actions. Especially for generating high-quality Spymaster clues that avoid opponent and assassin words.
| name | codenames-qa |
| description | QA the Hebrew Codenames engine — run the clue-legality regression, the clue-quality benchmark against the feedback set, and optionally verify the live HF Space. Use before committing engine changes, after tuning scoring or legality, or when asked to "run QA", "run the tests", "run the benchmark", or "check the engine still works". |
Three checks, cheapest first. All run from the repo root. The offline ones need the compressed fastText model the app uses:
export FASTTEXT_COMPRESSED=data/cc.he.300.fp16.bin
python tests/test_legality.py
Locks in the shared-root + transparency-gate behaviour (the מדע/מדען bug, the opaque
cognate מלחמה/לחם staying legal) using calibrated fastText cosines. Must print
All 18 legality cases pass. If you change ROOT_TRANSPARENCY_THETA or the root lexicon and
a case flips, that's the intended trip-wire — re-judge, don't just edit the expectation.
python -m research.bench_feedback
Reconstructs every 👍/👎 spymaster board (from the HF feedback dataset) and reports:
Health metrics are the robust signal. Learning/retention are directional (small N, prior-engine clues) — read them as trends, not pass/fail.
python .agents/skills/codenames-qa/scripts/verify_live.py
POSTs to the deployed endpoints and asserts: מדען illegal next to מדע, מלחמה legal next to
לחם, a spymaster clue generates for each risk profile, and the input guards return clean 400s
(unknown role, empty board). Pass --base <url> to target a different host (default is the public
Space). Only meaningful once the Space build stage is RUNNING (see the codenames-deploy skill).
probe.py → 1 + 2 before committing.codenames-deploy → 3 once the Space is RUNNING.