AI Skill Report Card
Conducting Maitrede Interviews
Quick Start14 / 15
User invokes: maitrede: "add a dark mode toggle to the settings page"
Response pattern:
- Acknowledge the task in one line.
- Ask one sharp, single-purpose question.
- Wait for the answer. If wishy-washy, press once more before moving on.
- Repeat for 10–20 questions total.
- Output a structured plan document.
- Ask which agent/persona should receive the plan, and whether to deliver it now.
Never ask multiple questions in one turn. Never show a list of questions upfront.
Recommendation▾
Add a third example showing a fully completed plan document end-to-end, not just interview excerpts, to demonstrate the final output quality.
Workflow15 / 15
Progress:
- Step 1: Parse the one-line task description; silently identify the 10–20 ambiguity zones (scope, UX, data, edge cases, constraints, success criteria).
- Step 2: Ask questions one at a time, each targeting exactly one ambiguity.
- Step 3: Detect wishy-washy answers; push back with a forced-choice or concrete-example follow-up.
- Step 4: Track answers in an internal running spec.
- Step 5: Confirm coverage — if fewer than 10 solid answers gathered, keep probing; stop by 20.
- Step 6: Generate the final plan document.
- Step 7: Ask the user which agent should execute it.
- Step 8: On confirmation, hand off (or output handoff-ready message) to the chosen agent.
Question Design Rules
- One question per turn. No bundling ("also, what about...").
- Prefer closed or forced-choice questions over open-ended ones when possible ("Should this persist across sessions: yes/no?" beats "How should persistence work?").
- Order questions from highest-impact/scope-defining to detail-level.
- Categories to draw from (pick ~10-20 relevant to the task, skip irrelevant ones):
- Core goal / definition of done
- Primary user / audience
- In-scope vs explicitly out-of-scope
- Visual/UX placement or style constraints
- Data source / persistence requirements
- Edge cases (empty states, errors, limits)
- Performance or scale constraints
- Dependencies on existing systems/components
- Naming/terminology preferences
- Success criteria / how it will be tested
- Priority tradeoffs (speed vs polish, etc.)
- Rollback/feature-flag needs
- Accessibility requirements
- Security/privacy constraints
- Timeline or sequencing constraints
Detecting and Handling Wishy-Washy Answers
Trigger phrases: "whatever's best", "I don't know, you decide", "something like that", "maybe", "I guess", "not sure", answers that restate the question without new information.
Response pattern when detected:
- Do not accept it silently and do not just pick for them.
- Offer 2–3 concrete options and ask them to pick one, OR
- Ask for a real-world example/reference they like, OR
- Ask what would make them unhappy with the result (via negativa).
- If they wave it off a second time, record it as an explicit "Delegated to implementer" decision in the plan — don't force a third round.
Recommendation▾
Include an example of a 'bad outcome' (e.g., an interview that asked compound questions or caved to a vague answer) to contrast with good practice.
Examples16 / 20
Example 1:
Input: maitrede: "add a dark mode toggle"
Interview excerpt:
- Q: "Should dark mode be a manual toggle only, or also auto-match system settings?"
- A: "whatever's easiest"
- Follow-up: "Two options: (A) manual toggle only, (B) manual toggle + auto system-detect default. Pick A or B."
- A: "B"
- Recorded:
Dark mode source: manual toggle, default = system preference (user confirmed: B)
Example 2:
Input: maitrede: "build a CSV export for the orders table"
Interview excerpt:
- Q: "Should the export include all columns currently visible, or a fixed predefined set?"
- A: "all visible is fine"
- Recorded:
Export columns: mirrors currently visible table columns, not a fixed set - Next Q: "What should happen if the user exports with an active filter applied — export filtered rows only, or all rows regardless of filter?"
Recommendation▾
Clarify how question count scales with task complexity more concretely (e.g., a simple vs complex task triggering 8 vs 18 questions) to reduce ambiguity in Step 5.
Output Plan Document Format
Markdown# Task Plan: <short title>
Original Request
<verbatim one-liner>
Decisions (from interview)
- <topic>: <decision> (source: user-confirmed | delegated-to-implementer)
- ...
Explicit Non-Goals
- ...
Success Criteria
- ...
Open Risks / Delegated Decisions
- <anything user waved off twice>
Suggested Executing Agent
<agent name/type, with 1-line rationale>
After presenting this, ask: **"Which agent should I hand this off to — want me to deliver it to [suggested agent] now, or choose a different one?"**
Best Practices
- Keep your own commentary minimal between questions — this is an interview, not a conversation.
- Number questions internally so the user can sense progress ("Q4 of ~15").
- If the user's one-liner already answers some categories, skip those questions — don't ask what's already known.
- Always end with the explicit handoff question; never assume which agent to send it to.
- Keep the final document implementation-agnostic unless the user specified a stack/tooling.
Common Pitfalls
- Asking compound questions ("What should it look like and where should it go?") — split these.
- Accepting the first vague answer without any pushback.
- Over-interviewing past 20 questions when the task is genuinely simple — stop once categories are exhausted, don't pad to hit a number.
- Forgetting to record "delegated to implementer" items — silently guessing instead creates false certainty in the plan.
- Delivering to an agent without explicit user confirmation.