logbook-creator
Design and create a logbook — the meta-skill that produces other logbooks.
When to apply it
Reach for /logbook-creator when one of these five patterns describes the situation. Each one is a separate motivation; many real cases blend two.
Tracking across sessions
Debug attempts, design history, what's been ruled out — survives memory loss between sessions.
Drafting before commit
Agent proposes; human or another agent reviews before pushing to Jira / docs / report.
Human-in-the-loop review
Agent generates a structured set of items; reviewer needs to skim, filter, and act.
Multi-agent coordination
Several agents (or runs) read+write the same state; one source of truth, not chat-based handoff.
Collection for later analysis
Capture a noisy stream now; queryable structure for later.
How to apply it
Motivation
Identify the pattern that justifies this logbook (one of the five above).
Scope & lifecycle
Decide where it lives, how partitioned, when entries are appended/queried, expected lifetime, sunset rule.
Schema (entity-first)
Identify record types → derive per-type fields → set identity rule → set partial-row convention.
Storage + projections
Pick the authoritative store (CSV/JSONL/SQLite/spreadsheet) + optional projections (run-trace / export-only / mirror).
Create the two artifacts
Write the logbook instance + sibling spec.
Each step writes to the spec, not to a database. The two artifacts that get materialized at step 5 are the logbook instance and its sibling spec — the spec is what makes the logbook reviewable, queryable, and inheritable across sessions.
Sample output
The two artifacts side by side: a logbook instance (rows you append to) and the sibling spec that defines its schema, identity, and query patterns.
id,run_id,observed_at,severity,observation,supersedes,resolution_id,owner
1,r-12,2026-04-20,med,"agent re-tries same approach",,,@dy
2,r-12,2026-04-20,low,"too many sub-questions",,,@dy
3,r-13,2026-04-21,high,"agent ignores feedback when long",1,,@dy
# retro-observations.logbook.md
## Schema
| col | type | semantics |
|---|---|---|
| id | int | sequential, immutable |
| severity | enum | low / med / high |
| supersedes | int? | id of corrected row |
| resolution_id | int? | FK to resolution.csv |
## Identity rule
Sequential id, never reused.
## Correction rule
Append a new row referencing the original
via supersedes; never edit in place.
## Query patterns
- `severity = high AND resolution_id IS NULL`
- group by `pattern_tag` across runs