append-heavy · multi-agent

Ideation logbook

Multi-agent brainstorming run where many ideas accumulate, get scored, and converge toward a short list. A flat CSV is the substrate; the CLI is where the queries live.

Shape

Single flat table

Storage

CSV in git

Owner

Facilitator

Lifetime

One brainstorming run

Identity

Sequential id

Corrections

Patch in place

Schema

Column	Meaning
id	Sequential identifier, stable across patches and exports.
name	Short title — 3 to 8 words.
description	Prose payload. One paragraph. Free-form goes here, not in extra columns.
source_agent	Which specialist generated this seed (or `human`).
phase	`seed`, `scamper`, `six-hats`, `reverse`, `synectics`.
tag	Cluster label — filled by the tagging pass, not the generator.
impact	ICE: expected value if it works (1–10).
confidence	ICE: how sure we are it'll work (1–10).
ease	ICE: how cheap it is to try (1–10).
ice_score	Computed `(impact × confidence × ease) / 10`. Re-computed on patch.

Sample data

id,name,source_agent,phase,tag,impact,confidence,ease,ice_score 001,Inline comment threading on sections,explorer-ux,seed,collaboration,7,8,5,28.0 002,Annotate risk per section header,critic,seed,risk,6,7,6,25.2 003,Auto-dedupe near-identical seeds,generalist-1,scamper,ops,8,5,7,28.0 004,Render sections as cards with scores,explorer-viz,reverse,ui,5,6,8,24.0 005,One-click export of top-20 to Miro,generalist-2,synectics,export,8,7,9,50.4

Common queries

$ logbook filter phase=synectics
$ logbook sort ice_score --desc | head -20
$ logbook top 10 ice_score --group tag
$ logbook compute 'ice_score = (impact*confidence*ease)/10'

Humans read the converge output — not the CSV directly — but the CSV is what makes the queries cheap. A single idea's full prose lives in one cell; the interesting questions ("top 10 by tag") never need to load that prose into context.

Actions

Apply to Miro. Group top-N rows by tag; export each as a sticky with provenance. The workshop board is the next stage, not a mirror.

Generate report. Render the top 20 rows with full descriptions into a converge memo for the facilitator. Snapshot, regenerated on the next pass.

Visualize. Scatter of ease vs impact with ice_score bubble size, colored by tag.

Why this earns a schema

Hundreds of ideas, many agents appending in parallel, humans reviewing through filters and clusters. Rereading the whole file every time a new agent joins would be prohibitive. Patch-in-place for score updates; no supersession rows.

← All examples Examples directory Next example Backlog shaping →