菜单
Defines a cheap-to-evaluate score and one mutable workspace so agents iterate from numbers; separates immutable prep from editable experiment files. Use when designing overnight experiment loops, autoresearch-style harnesses, or metric-gated agent iteration.
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
Top-level entry for the AK.skill pack: routes agent-era public-idea work to the right sub-skill (paradigm surface, proof, idea file, hype+discipline, irreducible core, scoreboard loop, year review). Use when the user invokes /ak, says AK.skill, or needs a single checklist for shipping a thread, README, talk, teaching repo, or annual stack essay without picking modules manually.
Welds bold forecasts to plain limits, one real verification habit, paradox framing (capable yet brittle), and honest benchmark skepticism. Use when writing about agents, autonomy, or LLMs and preserving trust in the same breath as excitement.
Turns attention spikes into durable agent-readable idea files: invariants, policy vs mechanism, paste blocks for tools, and discussion-friendly hosting. Use when a post goes viral but the value is the pattern, not a frozen app or private dotfiles dump.
Ships a terminal simplification claim with a single-file or gist core, section-by-section teaching, efficiency vs essence split, and optional commit-granular video or repo twin. Use when demystifying a stack or teaching from minimal runnable code.
Surfaces paradigm shifts in public technical writing: open with lived workflow change, name one abstraction (optional metaphor or era label), and list the real stack honestly. Use when posting threads, READMEs, or memos about how tools or the profession are changing.
Pairs strong claims with the smallest falsifiable proof (script, gist, notebook, or runbook) and labels bare-bones baselines clearly. Use when arguing that a capability crossed a threshold or when credibility must fit in one link.
| name | scoreboard-loop |
| description | Defines a cheap-to-evaluate score and one mutable workspace so agents iterate from numbers; separates immutable prep from editable experiment files. Use when designing overnight experiment loops, autoresearch-style harnesses, or metric-gated agent iteration. |
Distilled pattern: give agents (or humans) a cheap-to-score objective and a single mutable workspace; let iteration be driven by numbers, not vibes.
train / eval, read score. Everything else is optional. When you need hygiene, split immutable prep (data, constants) from mutable experiment surface so diffs stay reviewable.