Why it matters
Most token advice is prompt advice, and prompts decay as a session drifts. A hook sits underneath the model where it cannot be forgotten mid-session, which makes the tool call, not the system prompt, the cheapest place to enforce spend.
Tokenmaxxing read
The saving compounds: a blocked 2,000-line read is one you never re-send on any later turn. But his ROI script is an estimate rather than a meter, priced at $3.00 per million input tokens on Claude 3.5 Sonnet rates, so treat 75 to 80% as a claim, not a measurement.
Source takeaway
His rule for the gate: never just refuse. A bare denial sends the agent guessing and burns more than the original mistake, so the script returns a rewritten command along with the block.
