Requirements Elicitation Skill

Operationalizes spec/project/requirements-elicitation/ so a development request becomes a precise, confidence-scored requirement set instead of the agent's first plausible guess. The method is engineered against the default case: the user often does not know exactly what they want, expresses it imperfectly, or is misread. It drives understanding up to a measured threshold before anything downstream is built.

Why this is a skill, not an agent

• Mid-flow interactivity is the contract — the spec mandates a one-question-at-a-time funnel interview with teach-back confirmation per requirement. Each turn shows one question, reads the user's answer, and lifts the lowest-confidence dimension; an agent's isolated, non-interactive context cannot run that dialogue.
• Persistent on-disk artifact — the result (requirement list + filled gap matrix + surviving assumptions) lives at project/requirements/<slug>.md and is read by downstream consumers (roadmap-plan, feature-decompose, issue-orchestrate). A skill owns persistent state; an agent returns a report and forgets.
• confirmed vs assumed and teach-back need the user in the loop — the spec forbids marking a requirement confirmed without an explicit teach-back confirmation; that signal can only be gathered interactively.
• Counter-dimension considered: a read-only agent could later score the gap matrix on a finished artifact (a validate-style audit), but the load-bearing dimension here is interactivity, not isolated computation — skill wins. The narrow scoring agent is a possible future companion, not this skill's core.

User-language policy

• Detect the user's language from their message and conduct the interview in that language — the dialogue must feel natural to the person being interviewed.
• The written requirement artifact uses the surrounding repository's primary language (English by default, unless the repo's existing project/ docs show otherwise — follow the precedent).
• The eight requirement-dimension keys (functional, non_functional, constraints, domain_objects, actors, acceptance_criteria, edge_cases, scope_boundaries) stay verbatim from the spec (canonical EN wins), regardless of interview language.

German trigger phrases

The frontmatter description keeps the trigger lexicon English-only per spec/claude/skill-management/ §Structure (plugin-distributed skills). Treat the following German paraphrases as equivalent and discoverable through this skill:

• "Anforderungen für X erfassen / aufnehmen"
• "interviewe mich zu diesem Feature, bevor wir bauen"
• "ich will etwas bauen, weiß aber noch nicht genau was"
• "hilf mir herauszufinden, was ich eigentlich brauche"
• "Anforderungen klären, bevor wir loslegen"

Precondition

Before any operation, locate spec/project/requirements-elicitation/ — either in the target repo or, when absent, at ${CLAUDE_PLUGIN_ROOT}/spec/project/requirements-elicitation/<canonical_language>.md (the copy shipped inside the installed nolte-shared plugin). If neither is reachable, stop and ask the user which spec source to use; the methodology spec is the input to this skill — without it there is no authoritative definition of the dimensions, the KPI, or the gating thresholds. Do not improvise a replacement.

The understanding KPI (load-bearing)

The interview is driven by the spec's KPI, maintained live across turns:

• Gap matrix over the eight closed dimensions; each is either applicable (with a confidence score c_d ∈ [0,1]) or explicitly n/a (reason).
• c_d is an uncertainty proxy, not a calibrated probability. Estimate it primarily by self-consistency (per the spec, the strongest training-free signal): privately generate k ≥ 2 independent interpretations or solution sketches for the dimension; the more they diverge, the lower c_d. Verbalized confidence MAY supplement this but is never the sole input.
• Separate the two uncertainty sources per dimension: specification uncertainty (the user has not decided/stated what they want — remedy: a decision-eliciting question) versus interpretation uncertainty (the user was clear, but the agent is unsure it read them right — remedy: a teach-back confirmation).
• Gate on the weakest required dimension: U_gate = min_d c_d. Default thresholds (spec defaults, project-overridable with a recorded rationale): τ_low = 0.4, τ_high = 0.8.
  • c_d < τ_low → a clarification is mandatory before proceeding past that dimension.
  • c_d ≥ τ_high and teach-back confirmed → the dimension is "understood".
  • in between → clarify only when expected information gain (EVPI) exceeds the question's cost.

Operations

1. elicit — run the interview and write the artifact

Interactive walk-through. One question (or one tightly-coupled group) per turn — never a batched questionnaire. Surface each step and let the user correct before moving on.

1. Name and scope the elicitation. Prompt for a one-line subject and a <slug> for the artifact. Capture the bounded context (what is being built, for whom, what is explicitly out of scope) before any requirement is recorded.
2. Open the funnel wide. Begin with broad, open-ended questions ("walk me through what you're trying to achieve"); narrow to specific, closed questions only as understanding firms up. Resist scripting a fixed turn order — adapt to the answers.
3. Initialize the gap matrix. Mark each of the eight dimensions applicable or n/a (reason); set every applicable c_d low until evidence raises it.
4. Loop until saturation (see Hard rules), each turn:
   • Identify the lowest-confidence required dimension (U_gate).
   • Decide whether to ask: below τ_low a question is mandatory; otherwise ask only if EVPI > cost (don't fatigue the user with low-value questions).
   • Choose which question by what most shrinks the space of viable interpretations — offer the divergent readings as options rather than an open "can you clarify?".
   • Use probing types deliberately: elaboration, interpreting, reason-seeking (separate want from need — ladder a proposed solution back to its underlying goal), consistency.
   • For tacit/IKIWISI gaps, offer concrete examples, counter-examples, negative scenarios ("what should never happen?"), and edge/error cases — these are the dimensions users most reliably omit.
   • Teach-back before raising c_d to "understood": reflect your interpretation back in the user's terms and get explicit confirmation. Only then may the requirement be tagged confirmed.
   • Screen each utterance against the non-exhaustive trigger-word checklist (optionality, subjectivity, vagueness, weakness, implicit reference, multiplicity, under-specification); when in doubt, flag it (recall over precision).
   • Normalize each understood requirement into an EARS-style target structure ("WHEN <trigger>, the <system> SHALL <response>") or equivalent controlled phrasing; a requirement that resists normalization is not yet understood.
   • Update c_d only from evidence (an answer, a confirmed assumption, a successful teach-back) — never from the mere passage of turns.
5. Make assumptions explicit. Any inference the user has not confirmed is recorded assumed, never confirmed, and surfaced for confirmation.
6. Write the artifact to project/requirements/<slug>.md using templates/requirements.template.md: the normalized requirement list (each tagged confirmed/assumed with traceability to the triggering utterance), the filled gap matrix with final c_d and U_gate, the thresholds used, and the surviving assumptions / open risks. Confirm the path back to the user.

2. validate — audit an existing requirement artifact

Run this checklist against a given project/requirements/<slug>.md:

• Bounded context (what / for whom / out of scope) is declared before any requirement
• Gap matrix covers all eight dimensions, each applicable-with-c_d or n/a (reason)
• Each c_d is justified by a named evidence event; at least one derives from a k ≥ 2 self-consistency check, not self-report
• Every requirement is in the normalized EARS/CNL structure or flagged not-yet-understood
• Every requirement is tagged confirmed (teach-back/authoritative answer) or assumed, consistent with its matrix cell
• Specification vs interpretation uncertainty is distinguishable per dimension
• U_gate, τ_low, τ_high, the self-consistency k, and the question budget are stated explicitly
• Surviving assumptions / below-τ_high cells are listed as named open risks

Report pass/fail per item. Offer to fix mechanical gaps (missing tags, missing dimension placeholders) in place. Never invent confirmed tags, requirements, or c_d values while fixing.

3. revisit — update after a scope change

Triggered when the user signals a material scope change (new actor, new constraint, new out-of-scope boundary, a built prototype that changed their mind — the IKIWISI feedback loop). Re-run elicit steps 2–5 as a diff against the existing artifact: show which requirements stay, which need re-validation (reset their c_d), which become irrelevant. Persist only after the user accepts each diff item.

Gotchas

• c_d is a proxy, not a probability. Per the spec it drives gating as a relative ordering; never present it to the user as a literal "I'm 80 % sure this is correct." It says "this dimension is the weakest link," not "this is calibrated truth."
• Specification uncertainty and interpretation uncertainty need different questions. A low c_d because the user hasn't decided calls for a decision-eliciting question; a low c_d because you might have misread calls for a teach-back. Conflating them wastes turns and fatigues the user.
• Over-questioning is a failure mode too. The EVPI/cost rule and the question budget exist to stop the interview interrogating the user past the point of value. A confidently-understood smaller set beats an exhausted user and a padded one.
• Teach-back is not optional flavor. It is the only mechanism that catches the confident misread — an interpretation that is internally acceptable but wrong fails silently unless reflected back. No confirmed tag without it.
• The German trigger phrases ship in the body, not the description. The frontmatter description is English-only per skill-management §Structure (plugin-distributed); German operator-voice triggers live in ## German trigger phrases so they stay greppable in an open conversation.

Examples

• Read examples/01-vague-greenfield-elicit.md when running elicit on a vague greenfield request and driving the funnel interview to a written, confidence-scored artifact.
• Read examples/02-validate-existing-artifact.md when running validate against an existing project/requirements/<slug>.md and reporting pass/fail per KPI-checklist item.
• Read examples/03-revisit-after-scope-change.md when running revisit after a material scope change and diffing the artifact instead of re-eliciting from scratch.

Resumability

Per spec/claude/resumable-work/, this skill is resumable: true. State is persisted to .resume/requirements-elicit/<run-id>.yml after every teach-back confirmation and at each named phase boundary, carrying the live gap matrix and c_d values so an interrupted interview resumes without re-asking confirmed dimensions. On re-invocation, scan that directory for files with status: in_progress whose inputs: snapshot (the <slug> and bounded context) matches the current invocation; if one matches, prompt Resume run <run_id> from phase <phase> (last checkpoint <last_checkpoint_at>)? [resume / start-new / discard]. The state-file envelope and fail-closed semantics on schema or YAML errors are load-bearing in the spec; don't duplicate those rules here.

Hard rules

• Never record a requirement before the bounded context (what / for whom / out of scope) is declared in writing — this is the load-bearing rule of the underlying spec.
• Never tag a requirement confirmed without an explicit teach-back confirmation from the user. assumed is always the safe default.
• Never batch the interview into a single questionnaire. One question (or one tightly-coupled group) per turn; the funnel and the per-turn gate are the method.
• Never raise a c_d from the passage of turns; only evidence raises confidence.
• Never invent requirements, acceptance criteria, or c_d values. Missing information is an open question or an assumed entry, not a plausible fill-in.
• Never terminate by running out of ideas. Stop only by the recorded criterion — saturation (min_d c_d ≥ τ_high and no positive-EVPI question remains) or the question-budget cap — and on a budget-capped stop, surface every below-τ_high cell as a named residual risk.
• The default thresholds (τ_low, τ_high, k, question budget) are spec defaults and engineering values, not calibrated constants; state them in the artifact and let the operator override with a recorded rationale.
• When spec/project/requirements-elicitation/ disagrees with this skill, the spec wins. Propose updating this skill rather than diverging silently.