gap-scan

name	gap-scan
description	Read what the store already holds and work out what is missing, thin, stale or contradictory — then ask targeted questions instead of generic ones. Use after ingestion, periodically, or when the user says "what's missing", "what else do you need to know".

personal-context: gap-scan

Phase two. Generic intake questions are hard to answer; specific ones are easy. "Who else was around during the Berlin years?" gets a real answer where "tell me about your life" gets a shrug. That difference is what this skill exists to produce.

Procedure

1. Read the whole store. This is one of the two skills permitted to (the other is context-review), because gap analysis is meaningless on a slice. Say that you're doing it.

2. Score coverage by type and by period:

   Check	Gap when
   Types	A type has no entries at all — commonly preference and pattern
   Chronology	A stretch of years with no period or event
   People	A person referenced in another entry's body who has no entry
   Relations	related: links that point at nothing, or point one way only
   Scope balance	Almost everything is general — the store has no scoping
   Staleness	review_after in the past; situation older than six months
   Contradiction	Two entries that cannot both be true
   Thinness	Entries of one sentence, or with no history line

3. Produce the report — grouped, shortest section first, with a count so it's obviously finite:

   ## Coverage
   34 entries · people 9 · period 4 · event 6 · pattern 7 · care 3 · preference 5 · fact 0
   
   ## Gaps
   - No entries for 2016–2019. The only reference is "before Berlin" in period-berlin.
   - `person-r` is referenced by care-cbt-2022 but has no entry.
   - No `preference` entries. Every workspace is guessing at how to talk to you.
   
   ## Stale
   - `situation-current` — updated 2025-11, review_after passed 2026-05.
   
   ## Contradictions
   - period-berlin says 2019–2023; event-fathers-death (2021-03) says "while I was in Tel Aviv".
   
   ## Over-broad
   - 14 of 34 entries are scoped `general`. Suggested re-scopes below.
   

4. Ask, at most five questions, one at a time. Pick the gaps that would change the most downstream: missing preference entries and missing people beat a thin decade. Every question names what prompted it — "care-cbt-2022 mentions R; who is R?" — so it doesn't read as interrogation.

5. Write confirmed answers through the write protocol. Contradictions are resolved by asking, never by picking the newer entry.

What is not a gap

• Something they declined to answer during intake. Skipped is a decision, not a hole. Do not re-ask; you may list it once under a Deliberately empty heading if they asked for a full report.
• A period they have said they don't want in the store.
• Anything the schema allows to be absent. facts/ being empty is fine.

Rules

• Five questions maximum per run. A gap report that generates thirty questions gets closed.
• Never ask for sensitive material because the schema has a slot for it. An empty care/ directory is not a gap unless they say it is.
• Report contradictions neutrally. Two accounts of the same period usually mean an imprecise date, not a lie.
• Do not fill gaps by inference. A gap stays a gap until answered.
• Offer to stop after the report. Reading the analysis without answering anything is a legitimate outcome, and often the person just wanted to see what the system thinks it knows.