菜单
ML/LLM operations discipline — versioned model/prompt artifacts, eval gates, drift and cost monitoring. Minimal MVP scope (expands in F4).
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
| name | model-lifecycle |
| description | ML/LLM operations discipline — versioned model/prompt artifacts, eval gates, drift and cost monitoring. Minimal MVP scope (expands in F4). |
MLOps/LLMOps discipline, deliberately minimal (D6). Platform tooling from memory-long §stack.
Models, prompts, datasets, eval sets — all versioned, all reproducible: a prompt in production has a version and a changelog exactly like code (a prompt edit is a deploy) · training/fine-tune runs record data version + params + code sha — "we can't reproduce the model" is a SEV-grade statement · model/prompt + config pinned together (temperature changes behavior as much as wording).
No model/prompt reaches production without passing its eval set — same philosophy as the harness's own evals: behavioral checks + anti-checks (does it refuse what it should refuse?) · eval sets are versioned and grow from production failures (a bad output becomes a test case — the ICL anti-example pattern) · A/B or shadow comparisons for meaningful changes; "it looks better" is not a promotion criterion.
Quality drift: output quality sampled and scored on a schedule, not just at launch · data drift: input distribution shifts silently — monitor feature/prompt-length/language distributions · cost & latency: per-inference cost and p95 latency are budgets (coordinate with cost-discipline — LLM spend is the fastest-growing waste category); caching and model-tier routing are the levers.
PII in training/prompt data follows data-safety classification · user inputs are hostile by default — prompt injection surface documented per integration · model outputs that trigger actions (tool calls, writes) go through the same guardrails as human-authored actions, no exemptions.
Rollback for models/prompts = redeploy previous version (artifacts make it one command — see release-engineering) · report includes eval delta vs current production · incidents follow incident-response with the model version in the timeline.
Context engineering — memory navigation, textual graph traversal, progressive disclosure, recovery. Load when assembling context for a task or when lost.
Implementation-side security discipline — headers, dependencies, secrets, authz. Load for hardening work (sec-ops domain).
Commit conventions, branching, anchors in messages, and the git safety rules. Load for any git operation beyond status/diff.
Anti-generic-AI design discipline — bans, tells, and the slop test. Load for ANY new UI surface or visual redesign. Inspired by Impeccable (Apache 2.0).
Motion & animation discipline — purpose, performance, reduced-motion. Load when work involves animation, transitions or scroll effects.
Microcopy discipline — buttons, errors, empty/loading/success states, alt text. Voice and tone come from the project's memory, not from this skill.
