ntent

Find what your client hasn't told you, and get the answer before you build.

Case study · Product designSeptember 20264 min read

Software delivery teams build from documents they did not write: RFPs, scope workbooks, and call transcripts. Every brief contains unstated assumptions and contradictions. When those gaps reach implementation unresolved, both human developers and coding agents fill them with guesses.

ntent runs before implementation begins. It parses client source documents into a structured model, isolates undecided rules and conflicts as multiple-choice questions, collects stakeholder answers via a single web link, and exports a verified brief to coding agents and issue trackers.

52 of 55
Stakeholder questions answered in 45 minutes, with no meeting or login
9
Answers that changed v1 scope before code was written
20 min
Time to complete a question set, down from 42 minutes with open prose
Client view on mobile: enter your name once, then tap through structured multiple-choice questions.Client view on mobile: enter your name once, then tap through structured multiple-choice questions.
The client question link

Origin

ntent started as an internal set of repository validation scripts for coding agents. While applying those scripts to client delivery projects, we found that most defects originated upstream in the requirements documents rather than in code generation.

On a client portal build, we extracted 55 open questions from the RFP and workshop notes and sent them to the client on a single web link. Three stakeholders answered 52 questions asynchronously in 45 minutes. Nine answers directly changed project scope: one screen moved out of v1, authentication shifted to a later phase, and offline storage and data retention rules received explicit numerical limits. Two answers surfaced direct contradictions between the commercial workbook and the workshop transcript.

We packaged that workflow into ntent. Rather than competing with coding agents like Claude Code, Cursor, or Lovable, ntent supplies them with a verified specification.


The six-stage workspace

Each project progresses left to right across six tabs: Sources, Model, Questions, Sets, Review, and Versions. Live badge counts on each tab show how many items await ingestion, stakeholder replies, or team sign-off.

The model extracted from an RFP and call transcript. Every item is marked confirmed, proposed, or open alongside its verbatim source quotes.The model extracted from an RFP and call transcript. Every item is marked confirmed, proposed, or open alongside its verbatim source quotes.
Model view

When source documents are uploaded, Claude extracts a structured five-part model: actors and roles, user journeys, business rules, data entities, and screens. Every item in the model carries 1 to 3 supporting quotes from the uploaded documents. The backend verifies each quote character-for-character against the source text and records its exact character offsets. Any quote that fails substring verification is flagged as unverified, preventing fabricated rules from being marked as confirmed facts.

Verified passages highlighted inside the source document, each linked to the model items it supports.Verified passages highlighted inside the source document, each linked to the model items it supports.
Source verification

Structured questions for fast stakeholder replies

Every open item or cross-document conflict in the model generates a structured question. Each question specifies the target stakeholder role, a plain-English prompt capped at 20 words (enforced by an automated length check), a context line quoting the source document, 2 to 5 closed multiple-choice options, and an explicit “We do not know” option.

In controlled trials comparing this format against traditional open-ended prose questions, completion time dropped from 42 minutes to 20 minutes, and total words typed by stakeholders dropped from 1,407 to 177.

The question bank. Unverified source quotes are flagged before publishing, and questions raised by coding agents via MCP land in the same queue.The question bank. Unverified source quotes are flagged before publishing, and questions raised by coding agents via MCP land in the same queue.
Question bank

One link, no account required

The delivery team groups questions into a shareable set and sends a single URL or QR code to the client. Stakeholders enter their name once and tap through the options; every selection saves immediately. Responses stream back to the delivery team’s workspace within a minute, grouped by question and respondent.

Stakeholder responses grouped by question, showing selected options and optional notes.Stakeholder responses grouped by question, showing selected options and optional notes.
Incoming answers

Human-gated review

Stakeholder answers never mutate the specification automatically. In the Review tab, each incoming response is paired with the model rule it affects and sorted by priority:

  1. Conflicts: Two stakeholders selected different answers for the same question.
  2. Needs a reply: A stakeholder selected “We do not know” and the question must be re-routed.
  3. Model updates: Agreed answers ready to be merged into the specification.

Claude drafts the proposed rule update from the stakeholder’s answer, and the delivery lead commits one of three actions: Apply to model, Keep current rule, or Send follow-up on the same client link.

Resolving a stakeholder conflict: two client leads picked different day-one permit types, shown beside the rule their answers affect.Resolving a stakeholder conflict: two client leads picked different day-one permit types, shown beside the rule their answers affect.
Review and conflict resolution

Version snapshots, exports, and MCP integration

Locking a version records an immutable snapshot with a SHA-256 content hash and attribution for the client stakeholder who signed off. Each locked version exports to six formats:

Version 2 snapshot with SHA-256 verification and six export formats.Version 2 snapshot with SHA-256 verification and six export formats.
Versions and exports

Any question that has not yet been answered by the client is exported with an explicit status: "open" marker. In a 90-run benchmark comparing prose briefs against exported ntent models across 20 stated facts and 10 unstated gaps, coding agents reading prose silently invented behavior that contradicted the client in 13 of 30 runs (43%). When reading the ntent export where those same gaps were marked open, agents contradicted the client in 0 of 30 runs.

Coding agents also connect directly to ntent over the Model Context Protocol (MCP). Agents read the latest locked version and, when they encounter an unstated edge case during implementation, post a structured question back into the delivery team’s Question queue rather than guessing.

Connecting Claude Code and Cursor via MCP with a scoped project token.Connecting Claude Code and Cursor via MCP with a scoped project token.
MCP agent integration

Core product rules

ntent is live at ntent.app, built with Next.js, Supabase with row-level security, and the @ntent/tokens design system across light and dark modes.

All work