Targeted questions

"Tell me more about your food preferences." The household groans. They already said four people, busy evenings, spinach to use before Wednesday. The question feels like the system did not listen, because it did not. It fired a generic prompt template invented at runtime to perform care. A targeted question works differently. It is the smallest query whose answer materially changes the plan. It belongs to policy and named cognitive units like AskTargetedQuestion, not improvised each run for appearance of diligence. Restraint prepares the decision describes when to hold. Targeted questions describe what to ask when holding is correct.

Helpful context: Collaboration contract assigns question ownership across agents. One bounded judgment keeps asks tied to named work. Cognitive posture includes targeted questions as a conduct dimension.

The materiality test

Every candidate question passes one test: if both answers (yes and no, or either plausible branch) produce the same plan structure, the question adds cost without value. Materiality concerns plan architecture, not curiosity. Leftover acceptability is material because yes anchors busy days on reuse and no requires fresh composition on those evenings. Cuisine philosophy is usually not material when allergies, effort caps, and inventory deadlines are already specified.

Generic (bad)	Targeted (good)
"Any dietary restrictions?"	"Leena's cashew allergy is on file. Still confirmed?" (when stale risk)
"What do you like?"	"Are leftovers acceptable once this week? Changes Thursday anchor."
"Tell me about your schedule"	"Monday shows busy. Is that under 25 min active effort or early meetings?"

The bad questions feel helpful. They increase chat length without changing outcomes. Users learn to skim or abandon. The good questions respect attention because each one carries a visible consequence. The same materiality test applies outside meal planning. In invoice intake, asking "tell me about your vendor relationship" is generic, while asking "Is this PO number authoritative for accrual?" is targeted because the answer changes posting treatment. In support triage, "what happened?" is generic, while "Did the outage start before or after the deploy?" is targeted because routing changes. In compliance review, "any concerns?" is generic, while "Does this clause override the approved data retention policy?" is targeted because the escalation path changes.

The smallest useful question is the one whose answer changes the decision.

Introduction to Thoughtware Ch. 13

Owned by architecture, not prompt luck

Targeted questions belong in libraries as organisational cognitive units with tests: the question fires when a gap is present, stays silent when the gap is absent, maintains wording stable enough for regression, and routes the answer to the correct working-state field. Prompt-only clarification drifts with model version, temperature, and whoever edited the system message last week. Designed questions travel with substitution rules, evaluation suites, and behavioural standards across teams.

This matters because answers need to land in working state with provenance, not in chat scrollback alone. AskTargetedQuestion writes leftover acceptability to a field that downstream composition reads. Regression tests assert field population, not message text alone. Chat-only storage reintroduces least trustworthy input wins on the next run when retrieval misreads prior answers. Even material questions do not repeat every session if the answer persists in working state or endorsed knowledge. AskTargetedQuestion regression includes: question silent when field populated, question fires when field stale by policy. Field staleness policy depends on the kind of fact being stored, because a leftover preference may expire weekly while an allergy confirmation persists until a household member explicitly revokes it. Repeated identical asks train users that the system forgets on purpose.

The originating request might look like: busy Monday, Tuesday, Thursday, spinach before Wednesday, leftovers maybe. The system identifies leftover acceptability as material to composition strategy. AskTargetedQuestion fires once: "Are leftovers acceptable once this week?" Not a twenty-question intake. If a medical frontier is detected, the material response may be deferral rather than interrogation: "Shall I wait for your clinician's constraints before planning Wednesday through Friday?" That differs from "describe your medical history." If the calendar is ambiguous on which evenings are busy, the ask is: "Is Monday under 25 minutes active effort or early meetings?" Material to AssessMealPracticality. The anti-pattern is a questionnaire covering diet philosophy, cuisine preferences, and childhood food memories before reading the user's actual request.

Targeted questions versus challenge and deferral

Challenge as conduct surfaces conflicting constraints the user already supplied. Targeted questions resolve missing facts. Mixing them confuses users. "Fresh-only greens all week and use fridge spinach by Tuesday" is a challenge moment: name the conflict, offer tradeoffs. "Are leftovers acceptable?" is a targeted question: missing fact with material plan consequence. Both require restraint. They differ in whether the gap is missing information or incompatible instructions.

Some gaps cannot be closed by a user answer within product authority. Medical interpretation, purchase approval, and policy override require deferral, not another targeted question. The materiality test still applies: the correct conduct is defer with safe partial scope rather than "describe your medical history." Targeted questions resolve missing household facts within the system's role. Deferral returns leadership when facts require authority the system lacks.

UX placement matters. Targeted questions appear where the user understands why they matter. Pair the ask with the plan element it affects: "This changes Thursday's anchor." Generic chatbot clarification patterns that feel like the system reset the conversation erode trust regardless of answer quality. Record answers in working state with provenance. Chat text alone is weak storage. Endorsed answers may require promotion paths per knowledge is endorsed, not retrieved.

Questions as library assets

Organisations catalogue targeted questions the way they catalogue cognitive units: owner, applicability, evaluation cases, substitution rules. A question without an owner becomes prompt debris. A question with an owner becomes conduct infrastructure shared across agents. Agent policies reference named question cognitive units rather than embedding question text in orchestration code. Substitution rules then cover question behaviour when models or wording change. Conduct travels with libraries the same way judgment does.

Generative surfaces may vary phrasing if materiality and field routing stay fixed. Regression owns materiality and routing, not exact string match. Product copy guidelines constrain tone without breaking architectural ownership of the ask. Localised products preserve materiality when wording changes, because localisation teams need the materiality spec alongside translated strings. A longer translated question that asks about immaterial preferences breaks the architectural contract even if grammar is perfect.

During repair loops, targeted questions fire only when new material gaps appear, not because the loop restarted. Loop design passes working state forward so answered fields suppress repeat asks. Repeat asks during repair feel like the system lost the thread. Long sessions accumulate answered fields. Session design expires stale answers on policy, not on chat length. Targeted question cognitive units read field staleness before firing. Users tolerate one material ask per decision. They abandon products that repeat asks they already answered.

On-call runbooks name which targeted question cognitive units fire for which gaps so incidents are debugged against policy, not against chat logs alone. UI highlights the plan region each question unlocks, because users answer faster when they see why the ask matters.

What to do next

Targeted questions connect the restraint gate to the specific ask that unlocks composition. The materiality test prevents over-questioning, and library ownership prevents drift across model versions. When questions live in named cognitive units with regression suites, they become part of the architecture rather than prompt luck that varies with the next model update.

Read restraint prepares the decision, challenge as conduct, and collaboration contract.

Read next: Challenge as conduct covers what happens when the missing piece is a conflict, not a fact.