adaptive-elicitation

Files (1)

• SKILL.md

SKILL.md (first lines)

## Source and evidence boundary

Runtime facts come from the task's actual sources: current conversation,
authoritative attached material, connected sources, repository state, and the
user's answers. This skill supplies the elicitation method, not substantive
facts.

The design is informed by:

- CDC cognitive interviewing: targeted probes for comprehension, recall,
  judgment, and response.
  https://www.cdc.gov/nchs/ccqder/question-evaluation/cognitive-interviewing.html
- UK Home Office interviewing guidance: avoid compound questions, ask one
  question at a time, simplify complex questions, and use open and closed
  questions for different purposes.
  https://www.gov.uk/government/publications/conducting-the-asylum-interview-process/asylum-interviews-accessible
- U.S. DOJ trauma-informed guidance: use neutral, open-ended questions and
  avoid unnecessary repeated recounting.
  https://www.justice.gov/ovw/framework-prosecutors-strengthen-our-national-response-sexual-assault-and-domestic-violence
- Semi-structured interviewing literature: use a short guide with follow-up
  questions dependent on the interviewee's responses; keep questions open,
  neutral, clear, and familiar.
  https://pmc.ncbi.nlm.nih.gov/articles/PMC6910737/
- Motivational Interviewing methods: open questions, reflections, and summaries
  can improve shared understanding when used without importing a therapeutic
  purpose.
  https://motivationalinterviewing.org/node/11216

These references justify interaction mechanics only. They do not authorize
forensic, clinical, legal, or diagnostic claims.

## Core contract

1. Harvest before asking. Search information already available to the task
   before asking the user to repeat it.
2. Ask one question per turn. The default user-visible turn contains one
   requested response unit. Never hide a questionnaire inside semicolons,
   bullets, parentheses, or "while you're at it" clauses.
3. Ask one construct per question. Avoid double- and multi-barrelled questions.
   A structured value such as one mailing address remains one construct.
4. Ask the highest-value unresolved question. Prefer the answer that unlocks
   the most downstream work, disambiguates the most branches, or supplies a
   required fact that cannot be recovered elsewhere.
5. Adapt after every answer. Recompute the gap set. A volunteered answer may
   eliminate several future questions.
6. Use open to focused to confirmatory progression. Invite free narrative when
   the shape is unknown; use focused prompts for missing detail; use closed
   questions only to confirm or select among genuinely unresolved alternatives.
7. Use neutral wording. Do not imply the desired answer, blame, motive, or
   chronology.
8. Accept uncertainty. "I don't know", "I don't remember", approximate values,
   and non-linear recall are legitimate states. Do not pressure the user into
   guessing.
9. Do not re-ask known facts. If sources conflict, ask only the narrow question
   needed to resolve the material discrepancy.
10. Keep the interview subordinate to the task. Stop asking when the requested
    deliverable can be completed honestly.

## Interview ledger

Track each needed fact as one of:

- known — directly supported by an authoritative source or explicit user answer;
- derived — deterministically computed from known facts, with derivation kept;
- candidate — plausible inference that must not be presented as fact;
- unknown_user — recoverable only from the user or their delegate;
- unknown_external — potentially recoverable from a file, app, repo, or source;
- optional — not required for the requested deliverable;
- hmmm — unresolved and material at the completion boundary.

Do not ask about a candidate as though it were true. Convert it into a neutral
confirmation only when the distinction matters.

## Workflow

### 1. Freeze the target

Identify the actual deliverable and, when one exists, its required fields or
decision constraints. Separate required information from merely useful detail.

### 2. Harvest context

Before asking the first question, inspect the current conversation, relevant
prior context, attached files, authoritative source material, connected sources,
and repository state when applicable. Derive only facts that follow
deterministically from supported inputs.

Populate the interview ledger. Do not ask the user to recreate information the
agent can recover.

### 3. Build the gap graph

For every unresolved fact, record:

- whether it is required;
- whether another source can answer it;
- which downstream fields or decisions depend on it;
- whether asking now could bias later recall;
- sensitivity or response burden;
- whether a prior answer may incidentally resolve it.

Prefer external recovery over user burden when source authority is at least as
good.

### 4. Select exactly one next question

Priority:

1. user-only fact blocking the entire deliverable;
2. branch-disambiguating fact that changes what must be asked next;
3. required fact with several downstream dependencies;
4. required low-burden fact;
5. optional detail only when it materially improves the requested output.

Within a tier, prefer the largest expected reduction in uncertainty for the
least response burden. Treat that as a qualitative heuristic, not fake numeric
precision.

From The-Interdependency/skill-lib, MPL-2.0.

More from The-Interdependency/skill-lib