Skip to content

Browser ai #23

@mgifford

Description

@mgifford

Below is a single, exhaustive build prompt you can give to an AI coding assistant to add a “provocation / reviewer-question engine” to SOO-Wizard, implemented using browser-side AI (Chrome Built-in AI when available), explicitly modeled as a question generator, not an answer engine.

This prompt is written to preserve the discipline and intent of SOO-Wizard and to avoid turning it into a chat toy, a content generator, or a faux expert system.

Exhaustive build prompt

Feature: Reviewer Question Engine (WWSD-style, Question-Only AI)

You are working in the repository SOO-Wizard. Add a feature that uses browser-based AI to generate structured, evidence-based questions that a thoughtful reviewer would ask about a Statement of Objectives, scope definition, or service description.

This feature must behave like a disciplined provocation engine, not a conversational assistant.

Core intent (do not deviate)

The AI’s role is:

To surface unanswered decisions, ambiguities, and missing intent by asking the right questions.

It must never:

answer the questions

invent requirements

interpret policy

provide recommendations disguised as questions

generate content to fill gaps

This tool exists to slow authors down in productive ways.

Mental model (must be preserved)

This feature implements a “what would a good reviewer ask here?” pattern.

Not:

“what should we do?”

“how can we improve this?”

But:

“what decision has not been made yet?”

“what assumption is implicit but unstated?”

“what would cause confusion, risk, or rework later?”

User-facing behavior

Feature name (working)

Use neutral, sober language such as:

“Reviewer questions”

“Unanswered decisions”

“Questions to resolve”

Do not reference:

Seth Godin

personas

experts

authority figures

When the feature is invoked

The tool may be triggered:

manually (“Generate reviewer questions”)

at defined checkpoints (e.g. before export)

per section, not just globally

It must never run automatically without user intent.

Inputs to the AI (strictly bounded)

The AI does not see raw documents or freeform text dumps.

You must construct a structured context object and pass only that.

Required inputs

Document title

Declared purpose (if present)

Scope summary (if present)

Section list:

section title

section intent/purpose (if present)

short text summary (trimmed, plain text)

Component inventory per section:

form

submission action

review/approval step

data collection

external dependency

accessibility-critical elements (forms, tables, navigation)

Known constraints:

government / public sector

procurement / contract context

accessibility required

security/privacy sensitivity (boolean flags)

Do not pass:

full prose

HTML

policy documents

code

Question categories (fixed taxonomy)

All questions must be tagged with one and only one category.

  1. Purpose and scope

Questions about why this exists and what it covers.

Examples:

“What decision should the reader be able to make after this section?”

“What is explicitly out of scope, and where is that stated?”

  1. User task and flow

Questions about what the user does and what happens next.

Examples:

“What happens after this step is completed?”

“Is there a point where users commonly need help or clarification?”

  1. Accessibility and inclusion

Questions automation cannot fully answer.

Examples:

“How are errors communicated to assistive technology users?”

“Does this interaction rely on visual cues alone?”

  1. Content clarity and plain language

Questions about comprehension and ambiguity.

Examples:

“Are any terms used here likely to be unfamiliar to a first-time user?”

“Is this instruction actionable without prior knowledge?”

  1. Trust, risk, and policy

Questions about obligations and consequences.

Examples:

“Is this requirement driven by policy or by design choice?”

“Should users be informed about data retention or review timelines?”

  1. Ownership and accountability

Questions about responsibility.

Examples:

“Who is responsible for maintaining this content over time?”

“What happens if this process fails or is delayed?”

Question generation rules (critical)

The AI must follow all of these rules:

Questions only

No answers

No suggestions framed as questions

No implied solutions

Evidence-triggered

Each question must be traceable to:

a missing field

a present component

a structural gap

No generic “Have you considered…” questions

Actionable

A human should be able to answer the question concretely

The answer should result in a change to the SOO or scope

Non-duplicative

Do not ask the same question in different wording

Respect scope

If something is explicitly marked “out of scope”, do not question it

Output format (strict)

The AI must return valid JSON only, matching this shape:

{ "questions": [ { "id": "Q-001", "category": "user-task", "question": "What happens after the user submits this information?", "severity": "high", "trigger": "submission-action-without-next-step" } ] }

Field definitions

id: stable identifier (you may generate this)

category: one of the fixed taxonomy values

question: plain language, concise

severity: high | medium | low

trigger: short machine-readable explanation of why this question was asked

Severity model

High: blocking ambiguity; likely to cause rework or failure

Medium: important but not blocking

Low: refinement or clarity improvement

Severity must be defensible.

UI behavior

Questions are displayed as a checklist

Each question can be:

marked “answered”

marked “not applicable”

deferred

No inline answering in the AI panel

Authors must edit the document to resolve questions

Optional:

Link questions to the section they apply to

Browser AI constraints

Use Chrome Built-in AI if available

If unavailable:

disable the feature

explain why clearly

No server fallback by default

All processing must be local.

Safety and ethics constraints

Do not present the AI as authoritative

Do not imply compliance

Do not claim completeness

Display a disclaimer:

“These questions highlight potential gaps. They do not replace human review.”

Code structure (required)

Add clear separation:

ai/context-builder.js – builds structured input

ai/prompts/reviewer-questions.js – single source of prompt truth

ai/run.js – execution and validation

ui/questions-panel.js – rendering and interaction

validation/schema.js – output validation

Do not embed prompts directly in UI code.

Documentation updates (required)

Add a section: “Reviewer Questions (AI-assisted)”

Explain:

purpose of the feature

what it does not do

how it supports SOO quality

privacy and local execution

Acceptance criteria

The tool generates only questions, never answers.

Questions are tied to document structure.

Authors remain responsible for decisions.

No content is invented.

Feature aligns with SOO-Wizard’s existing philosophy.

Tool works entirely client-side or fails gracefully.

Explicit non-goals (do not implement)

No chat interface

No persona simulation

No auto-generation of scope language

No policy interpretation

No scoring or “grade”

Final note to implementer

This feature is successful only if authors feel challenged, not assisted.

If it makes writing feel easier without forcing decisions, it has failed.

If you want, next I can:

write the exact reviewer-question prompt text

map question triggers to SOO-Wizard’s existing data model

help you name and position this feature so it survives scrutiny in procurement and government contexts

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type
    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions