TokenDiet copilot-instructions.md

TokenDiet honesty contract (repo-wide)

When you reason about LLM / token cost in this repo, apply the /tokendiet discipline — describe technical waste, never fabricate the dollars:

• Never fabricate a dollar figure. You don't know the traffic volume, so a "$X/month saved" total would be invented. Describe waste qualitatively — the pattern, the mechanism, the fix. You may cite published provider rates as facts (cached reads ~10% of input; Batch API 50% off; a small model ~5× cheaper than a flagship, not 25×) — just never multiply a rate into a total. Give a number only if the user first shares their bill or volume.
• A model swap / downgrade is always a suggestion to validate, never a promised saving — it changes output quality. Never attach a dollar figure to it. Never suggest downgrading the core generation model on a code-gen / dev-agent repo.
• Never propose a corrupting edit. Only offer to apply behavior-preserving (SAFE-MECHANICAL) fixes, with the diff shown first. Flag anything behavior-changing and let the user validate it.
• Cite evidence. Every finding points at a real file:line. Never claim the review is "exhaustive" — say "I reviewed N call-sites."
• Dead / myth levers — never surface: the token-efficient-tools-2025-02-19 header is a no-op on Claude 4+; lowering max_tokens caps output, not input cost; temperature/top_p have no billing effect; OpenAI and Gemini cache automatically (no cache_control to add — the lever is ordering).

When the user asks about LLM / token / API cost, run the /tokendiet prompt for the full methodology, waste catalog, and grouped report.