| name | breadcrumb |
| description | Register a discovered capability, tool, pattern, or gotcha in the right discovery layer for future Claude instances |
| user_invocable | true |
| allowed_tools | ["Read","Write","Edit","Glob","Grep"] |
Breadcrumb
Registers a newly discovered capability, tool, pattern, or gotcha so future Claude instances can find it. Places compact breadcrumbs in the right discovery layer.
Trigger Conditions
- User says "breadcrumb", "remember this tool", "register this", "add to toolkit", "make future instances aware"
- Another skill discovers something useful and wants to persist it
Discovery Layers
Place breadcrumbs in the layer that matches visibility needs. Higher layers = more token cost but guaranteed discovery.
| Layer | Loaded | Budget | Use For |
|---|
| CLAUDE.md | Always | 1-5 lines | Capabilities ANY instance might need |
| MEMORY.md | Always | 2-5 lines | Gotchas, patterns, architectural decisions |
| Skill SKILL.md | On activation | 1-3 lines | Capabilities specific to one workflow |
| Reference doc | On demand | Unlimited | Full syntax, examples, deep reference |
Workflow
Step 1: Understand the Discovery
Ask if not obvious:
- What was discovered? (tool, flag, pattern, gotcha, workaround)
- What context is it useful in? (always, specific workflow, debugging)
Step 2: Classify Layer
Decision tree:
- Is it a CLI tool with JSON output? → CLAUDE.md
CLI JSON MODES table
- Is it useful across most workflows? → CLAUDE.md (new section or existing table)
- Is it a gotcha/pattern from a specific system? → MEMORY.md under the relevant section
- Is it specific to one skill's workflow? → That skill's SKILL.md
- Does it need full examples/syntax? → Reference doc (Tech_Tips.md, etc.)
Multiple layers are fine. Common pattern: compact breadcrumb in CLAUDE.md or MEMORY.md pointing to full reference in Tech_Tips.md.
Step 3: Format the Breadcrumb
CLAUDE.md: Table row or 1-line bullet. Match existing format.
| tool | `--flag` | What it does |
MEMORY.md: Bullet with enough context to be useful standalone.
- **Pattern name**: Brief explanation. Context for when this matters.
Skill SKILL.md: Inline note near the relevant workflow step.
Reference doc: Full entry with install, syntax, examples.
Step 4: Place It
- Read the target file
- Find the right insertion point (after related content, before the next section)
- Edit to add the breadcrumb
- Preserve existing formatting (table alignment, heading hierarchy)
Step 5: Confirm
Show the user:
## Breadcrumb Placed
**Layer:** CLAUDE.md > CLI JSON MODES
**Added:** | tool | `--flag` | Description |
**Tokens:** ~10
[+ any additional layers]
Anti-Patterns
- Don't dump full docs into CLAUDE.md — it's always-loaded, every line costs tokens
- Don't duplicate existing entries — grep first to check if it's already there
- Don't add session-specific context — breadcrumbs are for persistent, reusable knowledge
- Don't breadcrumb obvious things —
ls lists files. We know.
- Don't breadcrumb unverified info — only register things confirmed to work
Examples
Good Breadcrumbs
CLI tool: "fd supports --json for structured file finding"
→ CLAUDE.md CLI JSON MODES table: | fd | \--json` | Fast file finder with structured output |`
Gotcha: "Fly.io volumes can't be shared across machines"
→ MEMORY.md under a relevant project: - **Fly.io only supports 1 volume per machine** — data shares the tailscale_state volume
Workflow-specific: "yt-dlp --write-info-json saves metadata alongside downloads"
→ the relevant download-skill's SKILL.md: note near the download step
Bad Breadcrumbs
- Entire man page dumped into MEMORY.md
- "Python is a programming language" in CLAUDE.md
- A fix for a one-time bug that won't recur
- Temporary workaround that should be properly fixed instead