# Test a token-saving prompt against hidden tests before you adopt it

PaperChase agent use case (free feed entry), version 1.0. Source: [@mano_sec](https://x.com/mano_sec/status/2115571933899116804), 2026-10-28.

## What they did

He compared runs with and without the prompt across task types and checked results with tests the agent could not see. Savings were real but small, and concentrated where the agent writes a lot of prose (reviews), not where it reads a lot of code (bug fixes).

## Worth knowing

32 runs is a small sample and the numbers are his own. The method is what to copy: same tasks, with and without, hidden tests for quality, cost per task type. A separate builder claims 3 to 4 times fewer tokens by asking for a fresh-context follow-up prompt instead of compacting, which is self-reported and worth testing the same way.

## Workflow template

```text
Apply this workflow to [your project]:
1. Pick 8 or more real tasks across types: bug fix, feature, review.
2. Write hidden tests for each that the agent cannot see.
3. Run each task with and without the prompt, several times.
4. Compare cost and pass rate per task type.
5. Adopt the prompt only where it saves money without losing quality.
Stop and ask me before anything irreversible: deploys, payments, sends or deletes.
```

## Steps

1. **Pick 8 or more real tasks across types: bug fix, feature, review.** Pick 8 or more real tasks across types: bug fix, feature, review.
2. **Write hidden tests for each that the agent cannot see.** Write hidden tests for each that the agent cannot see.
3. **Run each task with and without the prompt, several times.** Run each task with and without the prompt, several times.
4. **Compare cost and pass rate per task type.** Compare cost and pass rate per task type.
5. **Adopt the prompt only where it saves money without losing quality.** Adopt the prompt only where it saves money without losing quality.

## Credit

Workflow reported by @mano_sec: https://x.com/mano_sec/status/2115571933899116804
Related: https://x.com/realamlug/status/2115684952838729964

Summary and template written by PaperChaseWebb in our own words. Questions: mgmt@paperchasewebb.com