seed

name	seed
description	Scan prompt files, CLAUDE.md docs, skill definitions, and other documentation to seed trigger phrases — bootstrapping total-recall from existing project knowledge. Supports --models flag for model-scoped triggers.

Seed — Bootstrap Triggers from Docs and Prompts

Overview

Automatically discover and scan documentation, prompt files, skill definitions, and configuration docs in the project. This seeds the trigger index from existing project knowledge so total-recall is useful immediately without manually scanning every file.

Process

1. Parse Arguments

• --models <list>: Comma-separated list of target models (e.g., --models sonnet,opus). Default: sonnet
• --force or --rescan: Re-scan files that already have triggers
• Optional path: Limit seeding to a specific directory

2. Discover Seedable Files

Glob for documentation and prompt files across the project:

• **/CLAUDE.md — project instructions
• **/.claude/**/*.md — claude config docs
• **/skills/**/SKILL.md — skill definitions
• **/agents/*.md — agent definitions
• **/commands/*.md — command definitions
• **/docs/**/*.md — documentation directories
• **/README.md — readmes
• **/*.prompt.md, **/*-prompt.md — prompt files
• **/ARCHITECTURE.md, **/CONTRIBUTING.md, **/CHANGELOG.md

Exclude: node_modules/, dist/, build/, .git/

If the user provided a specific path, scan only that path.

3. Filter Already-Scanned

Read .claude/triggers.json. Skip any files that already have trigger entries for ALL requested models, unless the user passed --force or --rescan. Files with triggers for some but not all requested models will be scanned for the missing models only.

4. Invoke Researcher

For each batch of unscanned files (up to 20 per batch), spawn the researcher agent (use the Agent tool with subagent_type set to total-recall:researcher):

Generate trigger phrases for these files.
Target models: <comma-separated model list>

Files:
<list absolute file paths, one per line>

5. Store Results

Merge results into .claude/triggers.json using the model-scoped format (version 2). Tag seeded entries with "source": "seed":

{
  "src/auth/middleware.ts": {
    "models": {
      "sonnet": {
        "phrase": "jwt route authentication guard",
        "samples": ["...", "..."],
        "convergence": ["jwt", "route", "auth"],
        "confidence": 0.85,
        "phases": 1
      }
    },
    "crossModelTerms": ["jwt", "route"],
    "source": "seed",
    "scannedAt": "2026-04-01T10:00:00.000Z"
  }
}

6. Rebuild Index

After storing all results, invoke the total-recall:index skill to rebuild the master word index.

7. Report

Seeded N new triggers from M files (models: sonnet, opus):

  File                              | Model  | Trigger Phrase                    | Confidence
  --------------------------------- | ------ | --------------------------------- | ----------
  CLAUDE.md                         | sonnet | project rules and conventions     | 0.88
  CLAUDE.md                         | opus   | project conventions enforcement   | 0.95

Skipped K files (already scanned for all target models). Use --rescan to regenerate.
Rebuilt recall index: X words → Y files across Z models.