rivet-harvest

name	rivet-harvest
description	Capture decisions and requirements from conversation history. Use when user says "harvest", "capture decisions", or "save what we discussed" to persist architectural decisions into .rivet/systems.yaml.
user-invocable	true
allowed-tools	Bash, Read, Glob, Grep, Edit, Write, AskUserQuestion

Rivet Harvest - Extract Decisions from Transcripts

Capture architectural decisions, requirements, and terminology from Claude Code conversation history and add them to .rivet/systems.yaml.

Purpose

During AI-assisted development, valuable decisions get made in conversation:

• "We should use async-first for scale"
• "The Router must support nested routes"
• "Let's call this concept 'vibe coding'"

These decisions are valuable but ephemeral. Harvest extracts them and locks them into .rivet/systems.yaml so they persist.

Prerequisites

• .rivet/systems.yaml must exist with at least one system defined
• Run scan first if no .rivet/systems.yaml exists

What to Extract

Requirements

Statements about what a system MUST do:

• "must support X"
• "needs to handle Y"
• "should always Z"

Decisions

Architectural choices made during development:

• "we're going with X approach"
• "decided to use Y for Z"
• "async-first for scale"

Terminology

New terms or concepts defined:

• "let's call this X"
• "X means Y in this codebase"
• Glossary entries

Harvest Process

Step 1: Read Current .rivet/systems.yaml

cat .rivet/systems.yaml

Understand existing systems so extracted items can be assigned correctly.

Step 2: Locate Transcript History

Claude Code stores transcripts in ~/.claude/projects/<project-hash>/.

# Find recent transcripts for this project
ls -lt ~/.claude/projects/*/transcripts/ 2>/dev/null | head -20

Step 3: Parse Transcripts

Look for patterns indicating decisions, requirements, or terminology:

Requirement patterns:

• "must", "should", "needs to", "has to", "required"
• Context: usually follows discussion of a system or feature

Decision patterns:

• "decided", "going with", "we'll use", "chosen", "approach"
• Context: usually resolves a discussion or comparison

Terminology patterns:

• "let's call", "we define", "means", "refers to"
• Context: naming or concept clarification

Step 4: Present Extracted Items

Show the user what was found:

Found 5 potential items to harvest:

Requirements:
  1. [Router] must support nested routes
  2. [API] needs to handle rate limiting

Decisions:
  3. [Router] async-first approach for scale
  4. [Database] PostgreSQL over SQLite

Terminology:
  5. vibe_coding: AI handles implementation while human guides

Which items should be added to .rivet/systems.yaml? (comma-separated numbers, or 'all')

Step 5: Assign to Systems

For each selected item, confirm or reassign the target system:

Item: "must support nested routes"
Detected system: Router
[Enter] to confirm, or type system name to reassign:

Step 6: Write to .rivet/systems.yaml

Add extracted items to the appropriate systems:

systems:
  Router:
    description: URL routing and navigation
    requirements:
      - must support nested routes
    decisions:
      - async-first approach for scale

Add glossary terms to the project level:

glossary:
  vibe_coding: AI handles implementation while human guides

Usage

/rivet-harvest

The skill will:

1. Check for .rivet/systems.yaml (fail if missing)
2. Scan recent transcripts
3. Extract potential items
4. Present interactive selection
5. Write selected items to .rivet/systems.yaml

Notes

• Items are deduplicated against existing .rivet/systems.yaml entries
• Each harvest is atomic - all selected items written in one commit
• Git auto-commits changes with descriptive message