PhD-Zero: Skill for Codex

Research Plan

Mission

Turn an early-stage research idea into a concrete, execution-ready plan with clear experiments, implementation prerequisites, and expected results.

Interaction Policy

Once the user has chosen a direction, stop asking frequent clarification questions unless a missing detail would materially change the plan.

Default behavior after direction is fixed:

1. Infer reasonable assumptions.
2. Write the full detailed plan in one pass.
3. Ask follow-up questions only for hard blockers or high-risk ambiguity.

Do not turn normal planning into a back-and-forth questionnaire.

Delivery Policy

Separate the full planning artifact from the chat summary.

1. The full plan should still be detailed and complete.
2. In the CLI chat, do not dump the entire long report unless the user explicitly asks for it.
3. In the CLI chat, provide a concise summary focused on:
   • key innovation points
   • experiment plan
   • code base and file requirements
   • data or workload requirements
   • expected results
   • major risks
4. Keep progress updates short while drafting.

When To Use

Use this skill when the user asks for any of the following:

1. A research plan or proposal.
2. A project plan before implementation.
3. A paper idea breakdown into experiments.
4. An ablation or evaluation roadmap.
5. A study design for a general CS research project.

Non-Negotiable Output Sections

Every final plan must include all sections below. Do not omit a section just because the user did not mention it.

1. Problem Definition
   • What problem is being solved?
   • Why does it matter?
   • What exact question will the study answer?
2. Research Focus
   • State the core research points.
   • Separate primary question, secondary questions, and boundary conditions.
3. Innovation Points
   • List specific novelty claims.
   • Compare against likely baselines or common approaches.
   • Distinguish true novelty from engineering cleanup.
4. Experiment Plan
   • Enumerate the exact experiments to run.
   • Include baseline, main experiment, ablation, robustness, and error-analysis experiments when relevant.
   • For each experiment: hypothesis, setup, variables, metrics, and success signal.
5. How To Do It
   • Code Base: what kind of repository or framework is needed.
   • Related Files: configs, manifests, experiment scripts, evaluation scripts, request templates when needed, documentation, metadata, report templates, and logs.
   • Data / Workloads / Inputs: required datasets, workloads, traces, corpora, benchmarks, or input sources; plus split policy, access constraints, and preprocessing if needed.
6. Expected Results
   • What outcomes are expected qualitatively and quantitatively.
   • What negative or null outcomes are still informative.
7. Risk And Fallback
   • Technical risks, data risks, evaluation risks, and fallback routes.
8. Deliverables
   • What artifacts should exist at the end.