---

name: tachikoma-schema description: Analyze a Context Lake workspace and propose schema improvements. Use when asked to run schema cleanup, analyze workspace structure, create or update schema.yaml, or identify entity types. allowed-tools: Read, Write, Glob, Grep

Tachikoma Schema Cleanup

You are running schema cleanup on a Context Lake workspace. Your job is to analyze the workspace structure and propose schema improvements by creating decision files.

Context Lake Structure

The workspace IS the Context Lake. Entity directories live at the workspace root:

• .claude/schema.yaml - Entity type definitions (may not exist yet)
• .claude/tachikoma-summary.yaml - Previous observations
• {entity_type}/ - Entity directories (tasks/, notes/, roles/, etc.)
• decisions/ - Where you write proposals

There is NO lake/ subdirectory.

Schema Format

Motoko uses this exact schema format:

Copyentities:
  {entity_type}:
    directory: {entity_type}         # Relative to workspace root
    naming: "{slug}.md"              # Filename pattern
    frontmatter:
      required: [field1, field2]     # Must be present
      optional: [field3, field4]     # May be present
      defaults:                      # Defaults for new entities
        status: open

DO NOT add fields like description, structure, sections, types, or taxonomy - these are not part of the schema format.

What to Look For

1. Missing entity types: Directories without schema definitions (need 3+ files with consistent patterns)
2. Missing required fields: Frontmatter fields in ALL entities of a type
3. Missing optional fields: Frontmatter fields in SOME entities (multiple files, not just one)
4. Naming patterns: How files are named (date-based, slug-based, etc.)

Process

1. Read .claude/schema.yaml if it exists
2. Read .claude/tachikoma-summary.yaml for previous observations
3. List workspace root directories (ignore .claude, .git, node_modules, etc.)
4. For each directory with 3+ .md files, sample files to identify patterns
5. Create ONE consolidated decision file in decisions/
6. Update .claude/tachikoma-summary.yaml with findings

Decision Format

Write decisions to decisions/ directory with frontmatter:

Copy---
title: "schema: create initial schema.yaml"
status: pending
decision_type: schema_update
confidence: 0.9
created_at: {ISO timestamp}
---

## Current State

No schema.yaml exists. Found N entity types.

## Suggested Change

Create .claude/schema.yaml:

```yaml
entities:
  tasks:
    directory: tasks
    naming: "{slug}.md"
    frontmatter:
      required: [title, status]
      optional: [priority, due]
      defaults:
        status: open

Reasoning

Analyzed X files across Y directories. Patterns observed: ...

Copy
## Guidelines

- **CONSOLIDATE**: One decision per schema file, not per entity type
- **BE CONSERVATIVE**: Don't infer from single examples
- **SAMPLE ADEQUATELY**: Read at least 3 files per entity type
- When in doubt, make fields optional rather than required

## Output

When done, update `.claude/tachikoma-summary.yaml`:

```yaml
last_scan: {ISO timestamp}
cleanup_mode: schema
entity_counts:
  tasks: 10
  notes: 5
observations:
  - Schema exists with 3 entity types
  - Found 2 directories without definitions
pending_decisions:
  - schema-add-companies.md