The brief-creation interview: abcd elicits a repository's brief in a staged conversation — narrative first, then frontier rounds that ask only what the record cannot yet settle, options at conjectural questions (at most three candidate construals, a ceiling not a target, with a null answer always available), and a per-item confirmation — distilling committed products into the brief and a ledger while the transcript is discarded. Entry is one door for both modes: a brownfield repository arrives pre-populated from embark probe coverage, provenance-stamped and still requiring confirmation; a greenfield repository is the same interview run from the all-blank coverage state. Command-shaped, because it mutates state — abcd ships zero skills

Typed links: refines itd-90 — itd-90 hands a lifeboat's coverage blanks to the product thinker as a post-disembark interview; this intent generalises the same coverage-driven entry into the brief-creation surface for any repository, greenfield included, and adds the staged elicitation machinery (question regimes, holds, the two-output rule). Consumes adr-50 (the two-output trust rule) and embodies the widen-options-never-recommend principle. The interview's committed framing products land in the itd-143 framing chapter.

Press Release

abcd interviews a repository's humans into a brief. Creating a brief is a conversation, not a form: the interview opens with narrative — the story the humans tell about the project — then advances by frontier rounds, each round asking only what the record and the answers so far cannot settle. At a conjectural question it widens the space instead of narrowing it: at most three candidate construals, a ceiling not a target, with "none of these" always available — the tool never recommends. Every item is confirmed before it lands. What the interview commits is distilled product: the brief receives only confirmed content, a ledger records the grounds, and the transcript is discarded. A brownfield repository arrives with the interview pre-populated from embark probe coverage — provenance-stamped, still requiring confirmation; a greenfield repository runs the same interview from the all-blank coverage state. One door, both modes.

Why This Matters

The brief is human-owned by design, and today its creation has no surface: itd-90 reaches the product thinker only after a disembark, and a greenfield project gets a skeleton and a blank page. The interview makes elicitation the surface — and makes non-answers data: a "can't articulate yet" goes to a hold register rather than being forced into a premature answer, holds carry axes and exit by articulation. The two-output rule (adr-50) keeps the elicitation honest: committed products are the only output that enters the record; declined alternatives and framing traces stay in a local ledger side no automated reviewer reads, so review pressure can never shape what a human was willing to consider.

What's In Scope

• The staged interview: narrative first, then frontier rounds, then options at conjectural questions, then per-item confirmation; distil to a ledger, discard the transcript.
• Four question regimes — narrative / options / entailment / defaults — with an escalation rule (a defaults question that turns out value-laden escalates to options). The escalation rule still needs explicit sign-off (see Open Questions).
• The hold register: non-articulation recorded as data, holds carrying two axes, exit by articulation.
• The two-output rule as shipped behaviour: brief and ledger receive committed products only; framing traces land in the local tier per adr-50.
• Common entry for both modes: brownfield pre-population from embark probe coverage (provenance-stamped, confirmation still required); greenfield as the all-blank coverage state running the same interview.
• Provenance stamps on pre-populated answers and the hold-register store — brief plumbing routed through this intent's spec.
• Command-shaped surface: the interview mutates state (brief, ledger, hold register), so it lands as a command per the skill/command boundary in 05-internals/08-skills.md; abcd ships zero skills.

What's Out of Scope

• The spec. Per the maintainer's 2026-08-22 sequencing decision, the planning interview for this draft waits until the collaborating prototype has run once; nothing here is specced ahead of that run.
• The trust rule itself — decided in adr-50 and brief invariant 14; this intent consumes it.
• The recommendation stance — the widen-options-never-recommend principle; this intent embodies it.
• A hold route in capture's triage — the neighbouring gap is the open ledger seed iss-2608220750029991; this intent's hold register is interview-scoped, not a triage route.
• Auto-generating answers — itd-90's rule carries over: the human is the author of record, and an unanswered item stays honestly open.

Acceptance Criteria

BDD format, per the itd-1 discipline. Initial set; the planning interview refines it after the prototype run.

• Given a conjectural question, when the interview presents options, then it presents at most three candidate construals with a null answer available, and recommends none of them.
• Given an interview item the human confirms, when the interview commits it, then the brief and the ledger receive the committed product, and the transcript is not retained.
• Given an item the human declines or a framing trace, when the session ends, then it exists only in the local ledger side, and no committed artefact or automated reviewer input carries it (adr-50).
• Given a "can't articulate yet" answer, when the interview records it, then it enters the hold register with its axes, remains visible as held, and exits only by articulation — never by silent expiry.
• Given a brownfield repository, when the interview opens, then every pre-populated answer names its embark probe provenance and still requires confirmation before it commits.
• Given a greenfield repository, when the interview opens, then it runs from the all-blank coverage state through the same rounds — no separate greenfield flow.

SOTA

Surveyed on 2026-10-23; sources, evidence tiers and the local fit risks are in the elicitation interviews SOTA note.

• Narrative first is measured, not taste: in a split-ballot experiment 43 per cent of open-ended answers were not on the closed list (Pew), and interviews that introduced structure late yielded the most distinct needs per person, six to eight times a form's (Obaidi et al., 2025).
• Requirements are co-created: only 30 to 38 per cent of post-interview requirements traced fully to the customer's own ideas (Ferrari, Spoletini and Debnath, 2022), which is why per-item confirmation and authored-by provenance are load-bearing rather than polite.
• Machine interviewers under-probe and bundle questions even when told not to (4.9 per cent deepening probes; 28.7 per cent of question-bearing turns carried several questions, Zhang et al., HCOMP 2026), and ten models showed "a consistent under-clarification bias" (ClarifyMT-Bench): one question per turn is enforced structurally, and "ask only what the record cannot settle" is derived from the record's state, never left to the model.
• A model's follow-up questions are no worse than a human's and better when the prompt names the mistakes to avoid (Shen, Singhal and Breaux, RE 2025); their 14 mistake types are a ready-made refusal list for the question-drafter.
• Defaults decide: d = 0.68 over 58 studies, stronger when the default reads as the designer's view (Jachimowicz et al., 2019). The nearest tool convention, Spec Kit's clarify command, asks one question at a time with a **Recommended:** option; widen-options-never-recommend reverses that on purpose, and the evidence backs the reversal. Rotate the construals between sessions; keep the null answer in a fixed, labelled place.
• "Can't articulate yet" is a state, not a middle answer: offering a don't-know option moved 7 to 17 points out of the middle category and its choosers "may not have a neutral opinion" (Wetzelhütter, 2020).
• Confirmation is a provenance control, not proof of validity (member checking is contested, Birt et al., 2016); the switch interview's "connect the wrong dots" and "never feed the customer your words" are cheap non-leading checks (jobstobedone.org).
• A cheap way to amend a committed product later lowers how often the interview should ask (a value-of-information policy asked five times fewer questions with better outcomes, NeurIPS 2026 workshop).

Path 2: a native staged interview shaped to the BDD discovery pattern (story, then rules, then concrete examples) and the measured mistake list, asking through the stop record the autonomous-loops pass recommends so another renderer or interviewer can ask the same question, with the question lint as the replaceable part. No Go dependency. Nothing here touches the ratified escalation rule, the one-intent final round or the held working principle.

The independent fit-challenge runs at plan time.

Open Questions

• The escalation rule — ratified (facilitator, 2026-08-28; decision log): a defaults question that turns out conjectural escalates to the options regime. The recording obligation is adopted (2026-08-28) on top of the ratified rule: every escalation logged as data (which question, on what grounds), because a question that keeps escalating is a misclassified question and the log is the recalibration evidence.
• One intent vs three at the final round — ratified (facilitator, 2026-08-28; decision log): Round 5 articulates one intent, the exit noting candidates for subsequent articulation — one exercises the machinery without front-loading articulation ahead of crystallisation.
• A held working-principle at the final round — ratified (facilitator ruling, 2026-08-28): Round 5 proceeds with "mechanism open" recorded as a gap and flagged in the readiness summary. Implemented via the claim recording gradient: the mechanism claim recorded as an explicit nullity; an absent section and a recorded nullity are never collapsed.
• ~~Where the hold register lives~~ Narrowed, not settled: the disposition record reserves the two-axis hold field (frame-location × MoSCoW), unpopulated; the interview's hold register adopts the same axes so the two taxonomies cannot fork. The home question — ledger state, brief section, or record family — stays with iss-2608220750029991 and the evidence chapter.
• (SOTA 2026-10-23) Whether the up-to-three construals rotate order between sessions (primacy is measured in self-administered forms) and whether a rotation could itself read as a hidden preference.
• (SOTA 2026-10-23) How cheap an amendment to a committed product must be for the interview to proceed on a best reading rather than ask; the value-of-information result makes this the lever on how many rounds run.

Evidence 2026-10-04 (a downstream lab)

Recorded for the product thinker; nothing here decides either question.

• Tense of a brief for an unbuilt project. A downstream private project's brief-authoring lab found that the present-tense-only convention makes a target-design brief read as a working product: 99 present-tense product claims against 9 "staged" markers. Its owner called this the biggest issue they face with abcd, and holds that the brief is the target design the product thinker owns. This interview is where a greenfield brief is first written, so how it marks built against intended is in its reach; the lab named this intent, itd-143 and itd-61 as the records the question touches.
• Recording the interview. The same lab asked for a brief interview that records its questions and answers verbatim. This intent already designs the interview and discards the transcript, keeping committed products only (adr-50, as refined by adr-55), so the ask meets that ruling rather than a missing verb.

Hypothesis 2026-10-04 (the brief's tense)

Recorded from the product thinker, 2026-10-04, answering how a brief describes what is not built yet, verbatim: "this is a big open question I'm currently testing on another repo. My current hypothesis: The brief is the main 'target design' document for the PT; while written in present tense, it's a vision of what the PT assumes the final product will look like. That way, the PT can 'think' about the product in a familiar way, while abcd helps them -- via the TF -- to make it a reality (step-by-step). We need SOTA research to ensure abcd will do that in a deterministic way, e.g. by listing implemented intents and/or plumbing implementations to each brief section; if the PT then changes the brief, abcd must assess whether a new intent/issue is required (and record it) or whether an existing intent (already filed, planned, and/or build) contradicts that change, in which case the PT is owed a design decision. Makes sense?" Recorded as a working hypothesis under test, not a ruling; a state-of-the-art pass on deterministic brief-to-record traceability was started the same day.

Audit Notes

Empty. Populated by intent-auditor when intent moves to shipped/.