Draft a Research Plan and Bias-Checked Discussion Guide

Workflow area: Research & Synthesis · Tier: Foundation · Hands-on time: ~25 minutes

The slowest part of running a study was never the sessions — it was the week the plan sat half-written while sprint work shouted louder. A model collapses the drafting to twenty minutes. But drafting speed makes a new failure cheap too: a biased guide, executed efficiently, produces confident wrong answers at scale. So this recipe has two halves that never get separated — generate, then attack what you generated.

When to use — and when not to

• Use it at the start of every planned studydiscovery interviews, concept tests, usability rounds; the structure is the same, only the questions change.
• Use it when stakeholders hand you a loaded brief"validate that users love the new dashboard" is exactly the brief that needs a bias pass most.
• Use the red-team pass on guides you wrote by handthe attack half of this recipe works on any guide, whatever drafted it.
• Skip the AI draft for studies in sensitive domainshealth, finances, grief, layoffs: draft by hand with an experienced researcher, and use AI only for the bias check.
• Skip it for a two-question hallway intercepta full plan for a coffee-corner chat is process cosplay.

The recipe

1. 1.

   Write the one-paragraph study brief yourself. What decision will this research inform, who you need to hear from, and what you already believe. The "already believe" line matters — it is what the bias check will hunt.

2. 2.

   Generate the plan, screener and guide in one pass. In Claude or ChatGPT:

   Copy

   You are an experienced UX researcher. Draft a study plan from this
   brief.
   
   Brief: [PASTE YOUR ONE-PARAGRAPH BRIEF]
   Decision this informs: [e.g. "whether to build approval delegation
   in Q3"]
   What we currently believe: [e.g. "we think approvers want mobile
   approvals"]
   
   Produce:
   1. RESEARCH QUESTIONS — 3-5, answerable by talking to users, none
      answerable by analytics alone.
   2. METHOD & SAMPLE — method, participant count, and a screener of
      5-7 questions that filters for the right people without
      revealing what we hope to hear.
   3. DISCUSSION GUIDE — opening, 8-12 main questions ordered
      general-to-specific, follow-up probes, closing. Every question
      open-ended. No question may presuppose a problem exists or that
      our product is the solution.
   

3. 3.

   Red-team your own guide in a fresh conversation. Do not do this in the same chat — a model grades its own homework gently. Open a new conversation:

   Copy

   You are a methods reviewer whose job is to find bias in this
   discussion guide before it reaches participants. Be adversarial.
   
   Guide: [PASTE THE FULL GUIDE]
   What the team believes: [SAME BELIEF LINE AS BEFORE]
   
   For each question, flag:
   - LEADING — presupposes an answer or plants a preference
   - LOADED — embeds an assumption the participant must accept to answer
   - DOUBLE-BARRELED — asks two things at once
   - CONFIRMATION-SHAPED — only elicits evidence FOR the team's belief
   - SOCIAL-DESIRABILITY — invites the polite answer over the true one
   For every flag, propose a neutral rewrite. Then list what the guide
   FAILS to ask that could disconfirm the team's belief.
   

4. 4.

   Adjudicate every flag yourself. Good red-team output looks like this — accept, reject, or rewrite each line deliberately:

   Copy

   Q4 "How frustrating is the current approval flow?" — LOADED:
   presupposes frustration. Rewrite: "Walk me through the last time
   you approved an expense. What happened?"
   
   MISSING: nothing asks when mobile approval would be WRONG —
   the disconfirming case for the team's belief.
   

5. 5.

   Pilot the guide on one colleague before any participant. A ten-minute dry run catches the questions that read fine and land clumsy. Then finalize the plan with your session logistics — and if you run unmoderated studies in Maze, paste the cleaned questions into the study builder rather than letting the tool's autogenerated phrasing back in.

Make it yours

• Claude Project as your research-ops template: store your standard consent script, screener boilerplate, and your product's belief inventory in a Project. Every new brief inherits the house method, and the red-team prompt can attack against your accumulated beliefs, not just this study's.
• ChatGPT for stakeholder-proofing: paste the final plan and ask for the three hardest questions a skeptical PM will ask about sample size and method. Walking in with answers beats improvising them.
• Gemini + Docs for the collaborative pass: draft in a shared doc and let teammates comment; run the red-team prompt again after human edits — well-meaning stakeholder additions are where leading questions sneak back in.

Quality check

• Every question is open-endedanything answerable with yes/no gets a rewrite before the pilot.
• The disconfirming questions existthe guide must be able to prove your team wrong; if every question can only confirm, the study is a ritual.
• The screener doesn't telegraph the studya participant who can guess what you want to hear will supply it.
• Every red-team flag has a decisionaccepted, rejected with a reason, or rewritten; unadjudicated flags mean the bias pass was theater.
• The plan names the decision it informsresearch without a decision attached becomes a deck nobody opens.