Requirements, One Question

A guided Q&A that ships a stakeholder-ready doc

The amplifier-qna-bundle

It gathers requirements one question at a time, then emits a stakeholder-ready doc

A guided, stateful Q&A tool that interviews you instead of handing you a wall-of-text questionnaire — and turns your answers into a formatted requirements document.

So what has it actually built? Start with the evidence.

The interview

One question (or two) at a time, tracked in a stateful session.

The output

A polished markdown or JSON requirements document — not a scattered thread.

Compact and real

~1,375 lines already ship 4 templates — 62 questions across 15 sections

The whole tool-qna module is about 1,375 lines across 5 files, and it encodes the four recurring requirements scenarios: software requirements, user stories, bug reports, and project kickoffs.

Real scenarios, already structured. But there's a trap in shipping all of them at once.

1,375

lines across 5 files

4

built-in templates

62

structured questions

15

sections

The failure mode

Dumping all 62 questions at once is the wall of text nobody finishes

A raw dump of every question is itself the questionnaire people abandon. The bundle's own instructions name the enemy it is fighting.

The tool's answer: never present the wall in the first place.

“Do NOT present the user with a wall of questions.” — context/qna-instructions.md, line 35

Mechanism — batching

A stateful Session hands the model one small batch at a time

Batch size defaults to 5 (questions_per_batch), and the instructions go further: ask one — at most two related — at a time, so the conversation stays human-sized.

Batching splits the wall. Statefulness makes the path smart.

questions_per_batch = 5 default batch size
Ask one, at most two per the instructions
Session tracks answers & notes never re-asks the answered

Mechanism — statefulness

The Session unlocks conditional questions only when a prior answer matches

The interview adapts instead of marching straight through: active_questions skips what's answered and gates follow-ups on earlier answers, while progress_pct tracks completion.

A smart path still needs a human-facing driver — meet the agent.

active_questions gate follow-ups on prior answers
progress_pct answered ÷ applicable, as a percent
is_complete tracks when the interview is done

The driver

A proactive requirements-analyst agent frames each question naturally

The 77-line requirements-analyst agent, included via the Q&A behavior, asks questions conversationally and uses record_note to capture off-template requirements the user reveals in passing.

Once the interview is done, the agent has one job left: produce the document.

Frames the questions

Asks them naturally, one at a time — not as a form.

record_note

Captures requirements that emerge organically but fit no predefined question.

The payoff

generate_report turns the answers into a stakeholder-ready document

generate_report converts the accumulated answers and notes into a sectioned, dated, completion-stamped markdown or JSON document — per the agent's Output Contract, “ready for stakeholder review and distribution.”

And the same pattern reaches past the four scenarios it ships with.

# The generated requirements document
# Topic
Template · Date · Session · Completion
  per-section answers + notes
Generated by Amplifier Q&A

# format: markdown (default) or json

What you keep

Bring your own YAML templates and the pattern generalizes

Custom templates in a scanned directory merge with the built-ins and override same-id ones — the repo even ships api-design-review.yaml (4 sections, 17 questions). A guided interview beats a questionnaire whenever structured requirements must become a shareable doc.

Built on the Microsoft Amplifier foundation, as a composable module.

YAML

custom templates merge with built-ins

17

questions in the shipped api-design-review example

7

operations span the full Q&A lifecycle

Sources

Research Methodology

Data as of: 2026-03-05 (last push) · Working · v0.1.0

Repo: manojp99/amplifier-bundle-qna — single-author bundle by Manoj Prabhakar Paidiparthy (Microsoft), 10 commits (7 mpaidiparthy@microsoft.com + 3 manojp99 merge account). No automated tests are present in the repo.

Commands run:

git rev-list --count HEAD & git shortlog -sne --all — 10 commits, author split
find modules/tool-qna -name '*.py' -exec wc -l {} + — 1,375 lines across 5 files
grep -cE 'Question\(' templates.py => 62; grep -cE 'Section\(' templates.py => 15
sed -n '560,577p' templates.py — 4 built-in templates (software_requirements, user_story, bug_report, project_kickoff)
sed -n '68,119p' models.py — Session state, progress_pct, active_questions gating
sed -n '29,79p' reports.py — generate_markdown_report / generate_json_report
sed -n '34,36p' context/qna-instructions.md — “Do NOT present the user with a wall of questions” (line 35)
wc -l agents/requirements-analyst.md => 77; gh api repos/manojp99/amplifier-bundle-qna

Gaps: No adoption/usage metrics exist in the repo or via gh; no example generated report is committed; correctness is grounded in reading the implementation, not a test suite (none present). The pain-point framing is the topic's framing, not a repo-measured baseline.

Primary contributor: Manoj Prabhakar Paidiparthy (mpaidiparthy@microsoft.com), 7 of 10 commits; the manojp99 account (3 commits) is the same person's GitHub merge account.