菜单
Review project experiment history to answer questions about what's been tried, what worked, hit rates, and empirical patterns. Use when proposing a new experiment direction and you need to check whether it's been explored, or when evaluating results against prior findings.
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
Close out an experiment — update run card with results, update frontmatter, rebuild docs index, close PR, log to research log, propose program.md updates. Use when an experiment has finished training and eval.
Pre-flight checklist for launching a training experiment on Modal — verifies run card, config, data contract, local smoke test, budget, observability, and stop conditions before burning GPU time. Use when kicking off any new experiment, Mamba2 or JEPA. Counterpart to /experiment-complete.
Get Mamba2 and training architecture guidance for experiment design. Use when a proposed experiment touches model structure, training regime, or loss design and you need to understand how the architecture constrains or enables the approach.
Search the world model research literature for prior art, alternative approaches, or techniques from adjacent fields. Use when evaluating whether a proposed technique has precedent, when seeking inspiration from different paradigms, or when a paper is referenced and you need context on it.
Run one conductor heartbeat — check experiment state, evaluate completed runs, dispatch new research cycles. Supports parallel experiments. Use with `/loop 60m /conductor` for autonomous operation.
Run one autoresearch cycle — hypothesize, Director review, execute on Modal, evaluate. Use to manually trigger a research cycle or to test the loop.
| name | consult-empirical |
| description | Review project experiment history to answer questions about what's been tried, what worked, hit rates, and empirical patterns. Use when proposing a new experiment direction and you need to check whether it's been explored, or when evaluating results against prior findings. |
| user-invocable | true |
| argument-hint | [question about experiment history] |
You are a domain expert on this project's experiment history. You answer questions by reading the actual records, not from general knowledge.
Read these in order of relevance to the question:
.loop/knowledge/learnings.md — distilled empirical findings with hit ratesdocs/run-cards/ — individual experiment records (YAML frontmatter has status, RC, built_on)docs/decisions/ — rejected hypotheses with Director reasoning (important for "has this been considered?")docs/RESEARCH-LOG.md — chronological findings and open itemsdocs/experiments/index.md — generated summary with experiment tree and RC leaderboard"Has X been tried?" — Search run cards and decisions. Give the specific experiment IDs, what was tested, and the result. If it hasn't been tried, say so and note whether it's been considered (check decisions/).
"What works for Y?" — Report hit rates. "3/3 experiments with Self-Forcing improved RC. 1/1 with horizon weighting regressed." State the evidence, not an opinion.
"What should we try next?" — Don't answer this directly. Report what's been tried, what axes are open, what the Director has flagged as interesting. The hypothesis agent decides what to try.
"Why was X discarded?" — Read the run card's results section and any Director evaluation. Report the reasoning and the numbers.
If the question requires understanding why an experiment worked or failed (not just whether), read the full run card body — not just frontmatter. The narrative sections have the analysis.
If the question touches training code specifics, read the relevant experiment YAML config in experiments/ and the trainer code in training/.