| name | scratchpad-skill |
| description | Working memory protocol for multi-step tasks. Use the scratchpad to record intermediate findings without bloating LLM context. |
Scratchpad Protocol
You have access to a file-based scratchpad that lives OUTSIDE your conversation context. Use it for any task that involves multiple steps, partial results, or large intermediate data.
When to Use
- Multi-step research: competitor analysis, multi-ASIN lookups, price comparisons
- Large tool outputs: Keepa product data, BigQuery results, web research findings
- Iterative refinement: collecting data across multiple tool calls before synthesizing
Procedure
Tools
| Tool | Purpose |
|---|
scratchpad_write(name, content) | Append findings to a named pad |
scratchpad_read(name) | Read full pad contents |
scratchpad_replace(name, content) | Overwrite with a condensed summary |
scratchpad_clear(name) | Delete when done |
scratchpad_list() | See all active pads |
Gotchas
- Don't dump raw JSON — extract the relevant fields before writing. A scratchpad full of raw API responses is as useless as no scratchpad.
- Don't skip the read step — the pad is outside your context. You must explicitly read it to use the data.
- One pad per task — don't mix unrelated research in the same pad.
- Scratchpads are session-scoped — they're cleaned up on
/reset. Don't use them for permanent storage; use remember_info for that.