What’s actually left for evidence

A context window is a capacity limit. It is not a promise that every token gets used equally well. Before any of it holds the thing you’re working on, five other claims are already against it. This tool is that subtraction, made visible.

usable evidence budget = context capacity
                       − trusted instructions
                       − task and retained history
                       − exposed tool definitions
                       − reserved generation
                       − operational margin

This page has no server. What you paste or drop is read in your browser and never sent anywhere.

Usable evidence budget

810K

81.0% of the window, after 190K is claimed. Largest single draw: operational margin.

Start here

About the best case on a 1M model. The largest draw is the margin, which is a choice rather than a cost.

Where the window goes

Each draw floats down from what is left after the one before it. Hover or focus a bar to read it, and select a draw to jump to its control.

Capacity
Draws against it
Usable
Usable when nothing is left

Pin the scenario you have as A, change something, and the chart ghosts A behind the new bars while the table shows the difference.

Every figure in the chart with its share of the window.
Term	Tokens	Share of the window
Context capacity	1M	100.0%
Trusted instructions	18K	1.8%
Task and retained history	40K	4.0%
Exposed tool definitions	0	0.0%
Reserved generation	32K	3.2%
Operational margin	100K	10.0%
Usable evidence budget	810K	81.0%

Adjust the terms

Sliders step by 1K, and the boxes accept shorthand such as 25k, 1.2m, or 25,000.

Context capacity

The whole window, before anything claims part of it.

Trusted instructions 18K

System prompt, project instruction files, memory, and the skill listing.

Task and retained history 40K

Prompts, replies, file reads, and tool results.

Exposed tool definitions 0

Tool schemas loaded into the prefix.

Reserved generation 32K

Room held back for the reply itself.

Operational margin 100K

Slack before automatic compaction fires.

Or set it from the autocompact threshold

Autocompact at a number of tokens or a share of the window. The margin is whatever is left after it.

Autocompact at (tokens)

Autocompact at (%)

Will my evidence fit?

Drop the material you want the model to work with and see how much of the usable budget it takes. Token counts are estimates from each file’s length.

Loading the file check…

Getting your own numbers

Run /context in Claude Code to see where your window is going. This table maps each term to what that readout calls it.

What each term is called in the coding tool’s context readout, and how to move it.
Term	What /context calls it	How to move it
Context capacity	The total in the header, such as the 1M in 120k/1M tokens.	Choose a model with a different window.
Trusted instructions	System prompt, Memory files, Skills, and Custom agents. That covers the system prompt, environment info, user- and project-level instruction files, auto memory, and skill descriptions.	Trim instruction files and memory. Skills flagged to stay out of model invocation also stay out of the listing.
Task and retained history	Messages: prompts, replies, file reads, and tool results.	This is the term that grows on its own. Compacting replaces it with a summary, and subagents keep large reads out of it.
Exposed tool definitions	System tools and MCP tools. A row marked deferred means tool search is active, and the page shows it without counting it.	This is the most reducible term. Turn tool search on, or load fewer MCP servers.
Reserved generation	Not a readout row. It is the room the reply needs.	It is set by what you ask for: a longer answer needs more room.
Operational margin	Not a row in every version. Where the readout has an Autocompact buffer row, the paste box uses it. Otherwise it is the slack before automatic compaction.	Set it through the autocompact threshold, with the two boxes under the margin control.

Loading the readout box…