Built to reduce wasted tokens.
Winskel reuses relevant context and avoids repeated work when it is safe. Here is what it does, and what one matched test measured.
Measured on the next beta
13.9% fewer tokens than a clean Claude Code install
9 coding tasks with Sonnet 5.5, two runs each, all passing the same hidden checks. Codex is still being measured.
What we measured
In a matched test of the harness and pipeline in testing for the next beta, Winskel and a clean 
Claude Code install ran the same 9 coding tasks with the same model, Sonnet 5.5, two runs each: small tasks, harder tasks, and trap tasks built around time zones, emoji and cycles. All 36 runs passed the same hidden checks. Through Winskel the tasks used 13.9% fewer tokens overall.
| Task set | Difference through Winskel |
|---|---|
| Small tasks | 2.7% fewer tokens |
| Harder tasks | 14.1% fewer tokens |
| Trap tasks | 20.6% fewer tokens |
Codex was not measured yet, and results on your own projects will differ. Against a heavily customized Claude Code setup the difference can be larger; we do not count that.
What avoided work looks like
In a finished mission on the next beta, Claude Code asked for the same Git status and the same passing tests again on unchanged code. Winskel answered from its record instead of running them again, and still ran every check on the result.
In the current beta
- One model under Send
- No separate planning or evaluating model runs beside the one doing the work, and choosing the model does not call a model.
- A brief with a size limit
- Each run gets a short project brief. When it is full, older observations go first and your own words stay.
- Answers resume the session
- When an agent waits for you, your answer goes back to the same session. The memory is not sent again.
- Checks without a model
- Resuming verification of kept work checks the same code again without a model call.
In testing for the next beta
- Shorter instructions to each agent, with the same rules.
- Project facts Winskel already found, such as the package manager and the declared checks, so the agent does not look them up again.
- A check that passed is not run again while the code has not changed.
Three kinds of numbers
Winskel keeps them apart, and says which one it is showing.
- Recorded usage
- What Claude Code and
Codex report about your plans, shown under each agent in the app. Winskel does not estimate your quota. - Estimates
- In testing for the next beta: Home places the tokens Winskel recorded beside an estimate of the same work without Winskel, labeled as an estimate.
- Measured comparisons
- The same coding tasks, run with and without Winskel and checked the same way, published with their tasks, agents and method, like the test on this page.