lockedin-capture

lockedin-capture

Dedicated capture skill for in-session work moments. Promotes capture from an implicit sub-role of the main lockedin skill into a first-class calibrated flow with its own writer/reviewer contract, matching the rigor of lockedin-render-jaso and lockedin-render-resume-en.

Capture quality determines vault quality. Vault quality determines render quality. This skill owns that first link.

Use this skill when

• The user says "save this", "log this", "track this", "absorb this", "remember this", "capture this", or "add this to my vault".
• The user describes a thing they just did: shipped a feature, attended a meeting, made a decision, learned something, completed a project milestone.
• The user pastes notes, a git log excerpt, a PR description, a Slack summary, or any short text and says to save it.
• The main lockedin skill detects capture intent during a session and routes here.

Do NOT use this skill when

• The user is doing coding, debugging, or writing without any signal that they want to persist the work moment to the vault.
• The user asks to render an artifact (resume, jaso, ideas) — route to the appropriate render skill instead.
• The user wants to query their vault without adding anything — use the main lockedin skill's query flow.
• The user wants to audit a document — use lockedin-audit.

Execution model

User input
    │
    ▼
1. Writer turn (prompt-writer.md)
   - Read user input + any provided context
   - Propose entity types, fields, edge structure
   - Run _infer_edges() mentally from EDGE_SCHEMAS domain/range
   - Quote source phrases from user input
   - Output: structured proposal (markdown table)
    │
    ▼
2. ValidatorDeterm (lockedin validate --dry-run)
   - Deterministic field-type and required-field check
   - Catches schema errors before LLM reviewer sees the proposal
    │
    ▼
3. Reviewer turn (prompt-reviewer.md) — SEPARATE Claude context
   - Re-load RUBRIC.md fresh (not from writer-turn memory)
   - Score 5 dimensions independently
   - Query vault for duplicate candidates (slug similarity, alias match)
   - Output: JSON score + candidates + suggested revisions
    │
    ▼
4. ReconcileNegotiator (if duplicates surfaced)
   - Present candidates to user as a numbered list
   - Ask one question: merge / update / create new
   - Never auto-merge, never auto-create-duplicate
    │
    ▼
5. write-before-confirm
   - State what will be written: "Saving 2 entries — meeting X, decision Y. OK?"
   - On confirmation: call write_entity for each proposed entity
   - Edges written atomically with entities