burnratelisted

# burnrate Most token-saving advice optimizes the wrong thing. The agent's *output* — the part everyone tries to shorten — is a small minority of what you are billed for. The bill is dominated by **context: the conversation re-sent on every single turn**. Shortening replies attacks the visible number, not the expensive one. This skill measures the actual split on **this user's machine**, names the specific files and tools driving it, cuts those, and then measures again so the improvement is a fact rather than a claim. ## The prime rule > **Never state a saving you did not measure.** No "~30% cheaper". No "cuts 65% of tokens". If you have not run the before/after on this user's own transcripts, you do not have a number, and you say so. Every competing tool in this space leads with an average from someone else's sessions. That average is not the user's, and pretending otherwise is the failure this skill exists to correct. Corollary: never trade correctness, test coverage, or verbatim accuracy of code/commands/errors for tokens. A cheaper wrong answer costs more than an expensive right one — you pay for it again in the next turn. --- ## Step 1 — Measure first, always Run the analyzer before recommending anything: ```bash python scripts/burnrate.py --days 30 ``` Zero dependencies, stdlib Python 3.8+, no network, no keys. It reads `~/.claude/projects/**/*.jsonl` (override with `--root`). Useful flags: `--all`, `--days N`, `--project <name>`, `--json`. Then check what