Offline, dependency-light measurement of Kagan's agent prompts — chars, words, lines, and approximate tokens per prompt across three arms:

• __baseline__ — empty prompt (control)
• __terse__ — "Answer concisely." (control: generic terseness)
• __current__ — the prompt actually shipped (orchestrator, review, attached worker, MCP prompts)

Pairs with evals/promptfooconfig.yaml (which measures quality). This script measures cost. Honest delta is __current__ vs __terse__ — beating baseline only proves the model needs some instructions, not that yours are well-engineered.

uv run poe prompt-tokens                                # print current table
uv run poe prompt-tokens --out evals/tokens/snapshots/now.json
uv run poe prompt-tokens --diff evals/tokens/snapshots/pre_compression.json

Uses tiktoken/cl100k_base if installed, else a chars/4 heuristic. For relative deltas the heuristic is stable (±1–2pp). Absolute numbers are ±15%. Anthropic's exact tokenizer is proprietary; cl100k_base is the standard public proxy.

Add to the collect() dict in measure.py. If the prompt is built dynamically (needs a Task etc.), add a stub builder above and feed it deterministic input — measurement must be reproducible.

• snapshots/pre_compression.json — baseline captured before the feat/prompt-compression-evals rewrite. Reference for the compression PR.

promptfoo runs the model and asserts on output. Token-cost questions (how many input tokens does this prompt cost?) need no model call. Running this is free and offline; running promptfoo costs API quota.