ifp

Garden Patch Home · Models

Collaborative Meeting as Composable Knowledge Pipeline

A meeting with a knowledge collaborator is not an event to document — it is a pipeline input that produces garden nodes, tool notes, person notes, and patch updates. The pipeline has seven stages. Each is independently useful, composable with others, and produces outputs that feed downstream stages.

The Seven Stages

1. Pre-Work

Input: Prior session outputs (garden patch, research, citations) Output: Updated agenda reflecting current state

Update the meeting agenda with work done since it was written. The agenda is a living document — it should reflect the current state of the garden, not the state when originally drafted. Add new items, note what’s changed, prepare questions that reference specific nodes.

Composable parts: Agenda template, workstream state review

2. Meeting

Input: Agenda, shared context, collaborator’s pre-loaded topics Output: Recording, live notes (collaborator’s edits to agenda)

The meeting itself. Accept that the agenda will diverge — the collaborator’s topics often reveal more than planned questions. Pete’s Section 0 dominated 60 of 90 minutes and produced more garden-relevant content than the planned walkthrough.

Composable parts: Recording setup, Chatham House protocol, live collaborative editing (HackMD or equivalent)

3. Compound Capture

Input: Recording, live notes Output: Compound meeting folder (lead file, raw transcript, audio in Archives/, recording sidecar)

Set up the compound folder structure immediately after the call. Copy audio from Zoom to Archives/. Create the whisper transcript placeholder. Create the recording sidecar with artifact predicates. Touch the raw transcript file for whisper output.

Composable parts: Compound folder template, recording sidecar template, whisper pipeline

4. Post-Processing

Input: Raw transcript, agenda with live edits Output: Cleaned transcript, meeting notes (lead file), person notes, daily note, spelling reference updates

Four sub-stages, each with its own skill:

Composable parts: /transcript-cleanup skill, /meeting-notes skill (with AI Summaries conditional), person note status tracking pattern, spelling reference as cumulative asset

5. Garden Integration

Input: Meeting notes, call observations, resources shared, consequences triage Output: Garden nodes, tool notes, clippings, garden-foundation tasks, workstream BACKLOG updates

The stage where meeting content feeds the garden. Five sub-stages:

Composable parts: Consequence triage protocol, resource routing rules (Tools/ vs Clippings/ vs Garden/), three-source validation signal, link conversion scripts, repatriation checklist

6. Report-Out

Input: Updated meeting notes, updated patch, compound folder Output: Cover letter, zip (without Archives/), merged PR, next-meeting agenda items

Package the work for the collaborator:

Composable parts: Cover letter template (raw voice), compound zip script, PR merge checklist, next-meeting seeding

7. Close

Input: All prior outputs, workstream state Output: Updated workstream state, deep learning captures, session log

Update the workstream BACKLOG (check off completed, defer blocked, move items to next call). Run deep learning to catch patterns and cascade improvements. Capture the session in the session log. Update MEMORY.md where session insights change persistent state.

Composable parts: Deep learning skill, session capture skill, workstream state machine

Filtering Ratios

Each stage filters. The 2026-03-19 Peter Kaminski call produced:

90 min call
  → 13K words raw transcript
    → 12K words cleaned
      → 9 topic sections in meeting notes
        → 18 consequence items triaged
          → 8 garden nodes seeded (2 inquiries, 2 glosses, 2 patterns, 1 model, 1 gloss)
          → 8 garden-foundation tasks
          → 9 tool notes
          → 4 person notes (2 created, 2 updated)
          → 120+ file garden patch (from 118+)
          → 31 URLs triaged from queue

The pipeline is a funnel at each stage and a cycle overall — garden outputs become next meeting’s pre-work inputs.

What This Model Does Not Cover

First Instance

2026-03-19 Peter Kaminski - Garden Patch Review — the meeting that produced this model. Compound folder with cleaned transcript, exemplar meeting notes structure, cover letter, and recording sidecar. All seven stages exercised in a single Claude Code session.

Relations