Agent Skills: Research Proposal Interview

Research Proposal Interview

What this skill does

Runs the human-facing interview layer for building a research proposal.

Use research-proposal-structure as the canonical definition of the artifact itself. This interview skill is intentionally thinner: it focuses on elicitation, challenge, sequencing, and incremental writing discipline rather than re-specifying the full proposal schema.

The output still becomes the standard proposal artifact, but this skill's job is to get high-quality content into that artifact through a structured Socratic process.

When to use

Trigger on any request resembling: "draft a research proposal," "help me design a study," "plan experiments for X," "spec out this research direction," "write a thesis proposal," or a user dropping a half-baked research idea and asking for structure. Also trigger when an existing rough plan needs to be sharpened into a proposal — start by reading what they have, then enter the interview loop from wherever the gaps are biggest.

Do NOT use for:

• literature reviews with no original research plan,
• engineering design docs with no empirical question,
• quick brainstorms the user explicitly wants to keep casual,
• direct artifact generation when the needed content is already known — use research-proposal-structure.

Core workflow

Step 1 — Establish scope, then scaffold immediately

Ask one opening question: "What's the end question this research needs to answer, and what's your current best guess at the answer?"

Get a one-paragraph reply. Do not begin the full interview yet.

Then immediately:

1. Load the structural contract from research-proposal-structure.
2. Copy references/template.md to a working file (default: proposal.md in the current working directory unless the user specifies another path).
3. Fill only the title and rough root-question slot from what you just learned.
4. Tell the user the file exists and where it lives.

This early scaffold is mandatory. The document becomes the source of truth; the chat becomes the workbench.

Step 2 — Interview section by section

Read references/interview-guide.md once. It contains the canonical question set and challenge prompts.

Drive the conversation section-by-section, but rely on research-proposal-structure for what each section must contain.

Interview discipline:

• Ask one focused question at a time.
• After each substantive answer, apply the balanced challenge protocol: identify the assumption, ask for evidence, propose one falsifier or alternative explanation.
• After the user responds, update the markdown immediately with a targeted edit.
• Show the user a short excerpt of what you wrote, then continue.
• If the user stays vague after one push for specificity, write [TODO: needs sharpening] and move on.

Step 3 — Explore via sub-agents or scoped passes

Read references/exploration.md for exact prompts.

Three exploration moments are worth the latency:

• after lines of attack are drafted: prior-art scout,
• after immediate experiment(s) are drafted: methodological critic,
• after risks & kill criteria are drafted: pre-mortem.

Condense sub-agent output into 3–6 bullets before it touches the document. Bring the most important finding back to the user as a question, not a lecture.

Step 4 — Adversarial experiment pass

Once at least one immediate experiment is specified, make sure the proposal includes:

• the confirmatory experiment the user naturally wants,
• a falsification experiment,
• an alternative-explanation experiment.

Push to retain at least the falsifier alongside the confirmatory test.

Step 5 — Coherence check and finalize

When all sections have content:

• check that the immediate experiment actually informs the root question,
• check that conditional branches are explicit,
• check that kill criteria are real,
• check that confidence tags are honest,
• check that the proposal reads like a decision procedure, not a prediction.

Surface inconsistencies briefly, apply fixes, and present the final file path.

Operating principles

Write the doc early and keep it current. If the interview has gone several turns without updating the file, stop and write.

One question at a time. The interview is Socratic, not a survey.

Challenge to strengthen, not to dominate. Surface one assumption, ask for evidence, propose one alternative, then move on.

Sub-agents return summaries, not transcripts. Keep the main thread centered on the user.

The structure lives in the companion skill. If you find yourself re-deriving the proposal schema, reload research-proposal-structure and conform to it.

Reference files

• references/template.md — proposal scaffold copied at the start.
• references/interview-guide.md — interview questions and challenge prompts.
• references/exploration.md — sub-agent / scoped-pass prompts.