跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
搜索与研究

experiment-planner

Use when exploring, continuing, or accepting a deep-learning or computer-science research idea. Converts claims into pilot-first experiment matrices covering diagnostics, failure analysis, authorized execution, evaluation, and paper-story viability.

sidiangongyuan · 已有静态检查记录

安装与使用

运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。

npx skills add sidiangongyuan/codex-skills-library --skill experiment-planner

技能说明


name: experiment-planner description: Use when exploring, continuing, or accepting a deep-learning or computer-science research idea. Converts claims into pilot-first experiment matrices covering diagnostics, failure analysis, authorized execution, evaluation, and paper-story viability. license: MIT

Experiment Planner

Overview

Use this skill when the user needs to explore a research idea, continue an authorized experiment workflow, or accept and interpret results. It turns claims into a testable story and experiment plan. It is an adapter over existing research-agent ideas, not a replacement for the user's writing, review, rebuttal, figure, evidence, or GitHub release skills.

Core Boundaries

  • Default domain: general deep learning and computer science research. Adapt to collaborative perception, 3D perception, or autonomous driving only when the task context calls for it.
  • For discussion-only planning, keep output in chat unless the user asks for a saved artifact. When the user asks to implement or run experiments in a repository, persist the pre-run result contract: update the paper's final LaTeX tables when a manuscript is in scope; otherwise update the existing experiment-planning document or create experiment-plan.md at the repository root.
  • Do not launch long experiments, deploy GPU jobs, modify code, or retry failed runs unless the user explicitly asks for execution or a still-valid approval covers the ordinary continuation. A bounded same-config recovery stays within that approval; do not silently expand its scope.
  • Do not replace paper-writing, paper-review-panel, rebuttal-response-skills, paper-visual-craft, or github-project-release; hand off to them only after the research plan or results are ready.
  • Treat external projects as references, not installed dependencies. Read references/source-map.md before discussing provenance or upgrading this skill from upstream sources.
  • During approved execution, a still-valid earlier user authorization covers ordinary implementation, local bug fixes, queueing in-scope runs, and evaluation. It does not authorize a new experiment family, dataset, seed policy, or paper-only TODO; those remain explicit opt-ins.

Entry Modes

Choose the mode that matches the current request:

  • Exploration: turn a new research direction into a falsifiable claim. Use the complete matrix for a substantial new study; a small pilot can use the minimal question/comparison/protocol-plus-budget/endpoint/decision contract.
  • Approved execution: continue within the frozen contract and the scope of a still-valid user authorization. Do not silently expand the experiment family, dataset, seed policy, or resource budget.
  • Acceptance: compare observed artifacts with the frozen contract and record whether the relevant gate passes, fails, or is inconclusive.

For a continuation, load the latest plan, results, and pass receipt when available. Reuse still-valid decisions about the claim, baseline, metrics, budget, seed, and protocol. Do not restart the initial grill, literature pass, or full matrix unless a material change or unresolved blocker makes it necessary; ask only questions that affect the next decision.

Default Workflow

  1. Grill consensus: for a substantial new exploration, use $grill-me style interaction to clarify problem, motivation, proposed claim, baseline/control, compute budget, success criteria, and unacceptable shortcuts. For a small pilot, record only the minimal contract. For a continuation, reuse the settled contract and ask only changed or blocking questions.
  2. Literature inspiration: for a substantial new exploration or a changed novelty claim, use $research-evidence for related papers, novelty risk, prior experiment patterns, and unsupported claims. Use $search-first when the task may need existing code, datasets, tools, or implementations. A small pilot may defer this pass when it does not change the decision. Reuse prior literature decisions when the claim and scope are unchanged.
  3. Story viability check: decide whether the idea can support a clean paper story: important problem, credible gap, specific method difference, feasible validation, and claims that will not outrun the evidence.
  4. Claim freeze: freeze the smallest verifiable claim before planning runs. Avoid changing the story repeatedly while experiments are running.
  5. Contract freeze: before scheduling a substantial study or full runs, define the final main-result, ablation, and necessary diagnostic tables. For every metric, record its plain-language definition, unit, direction, aggregation, and any delta reference. For a small pilot, freeze only the endpoint, comparison, protocol-plus-budget, and decision gate. Use the in-scope project's explicit missing-value schema and do not write claims from placeholder cells.
  6. Idea validation first: design the smallest pilot/smoke/sanity experiment that can falsify or support the core hypothesis. If multiple GPUs are idle, parallelize only independent exploration runs with clear ownership.
  7. Minimum sufficient matrix: only after the pilot passes, add the main result and claim-critical ablations. Add robustness, diagnostics, efficiency, qualitative results, or failure analysis only when they support a paper claim or answer a credible reviewer question; do not add them for symmetry.
  8. Subagent coordination: keep the main session responsible for planning, task decomposition, and final result acceptance. Use explorer for read-only repo/config/protocol investigation. Use worker for implementation with explicit file or module ownership. Do not manually override subagent model or reasoning settings unless the user explicitly requests it.
  9. Impact-based validation: prioritize checks by their ability to block the next d
阅读完整来源 ↗