Token costs

The workbench has knobs that change how much of your model's context window the output consumes. Some moves cost tokens rather than save them. This page tells you which is which, with measured numbers where they exist.

Show line numbers

The Show line numbers toggle in the FilterRail Advanced section prefixes every line of every file with its 1-based line number. The toggle's own help text says it the plainest way possible: "Increases token count by roughly 40 to 50 percent."

The transform itself lives in packages/core/src/file-processing/transform.ts as addLineNumbers. Measured against 8 representative repo files on 2026-06-07, the average increase was 44.83 percent. The cost is real on every file.

Show line numbers earns its place when you want the model to refer to a specific line in a follow-up question. It is a navigation aid, not a saving. Leave it off unless line references will appear in your prompts.

Output format

The Format group in the ActionBar Options popover offers two choices:

• Single file packs everything into one blob, ready to paste into a single prompt.
• Multi-part splits the output into chunks aimed at fitting your model's per-turn input limit. The popover notes that the recommendation switches above roughly 200K tokens.

Multi-part adds a per-chunk header so the model knows the chunks belong to one project. The overhead per chunk is small relative to the file contents, and the tradeoff buys you the ability to feed projects that would otherwise overflow your context window in one shot.

Output style

The Style group offers XML and Markdown. The XML style wraps each file in named tags and tends to read better for Claude. The Markdown style uses fenced code blocks and tends to read better for chat-style assistants that emit Markdown themselves.

There is no measured token delta between the two styles worth quoting here. The difference in practice is small. Pick whichever your assistant handles better.

Chunk size (multi-part only)

When Multi-part is the selected format, a Chunk size input appears. Smaller chunks produce more parts, which means more headers and slightly more total tokens but also more flexibility in feeding the output across multiple turns. The default sits at the value configured for the per-file size cap.

Cross-link

For the tokenizer itself, the 1 MB self-cap, and the model selector, see Token estimation.