菜单
Use when starting a research, eval, benchmark, or empirical comparison session - explains how to activate Epistemic skills, choose stage skills, and enforce research gates before any experiment work.
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
Use when a research idea, vague comparison, or surprising observation must be turned into a falsifiable, prereg-ready hypothesis before any benchmark, eval, training, or judge-backed work begins.
Use when starting or continuing empirical/ML research, evals, benchmarks, or any "is X better than Y?" claim — the umbrella mechanism that enforces hypothesis → preregistration → baseline → experiment → statistical rigor → falsification → kill-or-ship → verification, with gates you self-enforce.
Use when a claim, result, or draft compares your system against an external baseline, published paper number, or competitor result and you need a sourced local reproduction before quoting it.
Use when a preregistered hypothesis is ready to run and you must generate provisional evidence under the locked method without contaminating headline outputs.
Use when a hypothesis has reached a real decision point and you must choose KILL, PIVOT, RECOMMIT, REFINE, or SHIP from repository evidence.
Use when a hypothesis is concrete enough to freeze the experiment contract, lock the judge and compute environment, and scaffold the run before any experiment-shaped command executes.
| name | using-epistemic |
| description | Use when starting a research, eval, benchmark, or empirical comparison session - explains how to activate Epistemic skills, choose stage skills, and enforce research gates before any experiment work. |
Epistemic is a research-discipline layer for coding agents. It exists to stop empirical work from turning into cherry-picked runs, stale baselines, or headline claims with no paper trail.
If the task involves empirical research, evals, benchmarks, model comparisons,
ablation studies, "is X better than Y?", or publishing a result, invoke the
epistemic skill before answering or acting.
If a stage-specific skill applies, invoke it next:
research-questionpreregistrationbaseline-reproductionexperiment-executionstatistical-rigorfalsification-reviewsurprise-triagekill-or-shipverification-before-publicationDo not skip ahead because a run looks cheap, obvious, or exploratory. A cheap unregistered run can still contaminate the evidence chain.
Follow explicit user and repository instructions first. Epistemic skills define how to do research work, but they do not override direct user constraints, repository policy, or safety requirements.
Treat a repo as Epistemic-active when it contains any of:
HYPOTHESES.mdexperiments/RESULTS.mdIf none exists but the user asks to start empirical work, use
research-question to create the first falsifiable hypothesis before running
benchmarks, evals, training, or judge-backed scripts.
In the OMP agent, several gates are runtime-enforced. In portable harnesses such as Claude Code, you must enforce them yourself:
experiments/<id>/prereg.md.OVERRIDES.md entry explaining why.smokes/ number in README.md, RESULTS.md, a paper, or a claim summary.surprise-triage when a result diverges by more than 15%.kill-or-ship when cost exceeds 1.5x cap or a hypothesis goes stale.Overrides are allowed only when recorded with a concrete reason. If the reason would embarrass you in a methods section, do not override.
Claude Code should use the Skill tool for full skill content. Other harnesses
may expose skills differently; use their native skill mechanism when available.
The method is portable even when runtime gates, dashboards, or OMP-specific
widgets are not.
When Epistemic applies, start with a short announcement:
Using the epistemic mechanism.
Then load the umbrella epistemic skill and the current stage skill before
continuing. Ask one clarifying question at a time when defining a hypothesis.