What problem does it solve?
Tool description prompts for AI coding agents often restate information already inferable from the tool's JSON parameter schema, wasting context tokens. This Skill provides an evidence-based method to identify which prompt lines are redundant and which are load-bearing, plus authoring rules for what belongs in a tool prompt versus in code.
Core Features & Use Cases
- Schema-to-prompt inference probe: Runs a multi-model panel (Fireworks Kimi, Claude Opus, GPT) that reconstructs a tool prompt from only its wire schema and a blanked outline, flagging reliably recovered lines as prune candidates.
- Builtin tool shortcut: Resolves live builtin tools from the coding agent to extract the exact wire schema and real prompt, with a --show mode for inspection before spending API tokens.
- Authoring guidelines: Defines what belongs in a tool prompt (when to choose the tool, input grammar, agent-owned failures, anti-patterns) versus what stays in code (recovery logic, caching, internals).
- Use Case: Before deleting a line from a tool's .md prompt, run the probe across multiple samples and models, then git blame the line to confirm it is schema restatement rather than incident scar tissue.
Quick Start
Run the builtin probe on the read tool with --show to inspect its wire schema and derived outline, then launch the multi-model panel to find prune candidates in its prompt.
