Operator Use — See where your Claude limit goes

You hit the limit and you can't say why.

The work got done. The limit ran out anyway. Maybe it ran out on you before lunch, or one person on your team keeps hitting it while the others sail through the week.

The usual answers are guesses: "use it less," "use a smaller model," "that task must be heavy." Nothing tells you what to actually change, because nothing shows where the usage went.

Watching people work is slow, and people change what they do when they're watched. Asking them to remember is unreliable. The numbers are already sitting inside the session. They just need to be read.

Same job, same tools, same week. One person's limit lasts twice as long. The difference is almost always habits, and habits can be seen.

What it is

Two skills. One report. One comparison.

Operator Use runs inside Claude. Each person runs it for their own session. A manager, or you alone, reads the results. Nothing to install on anyone's computer beyond two skills.

1

Page one

The usage report

Measures the real session. Total usage, the share of your limit it used, how many deliverables you finished, and the one number that matters: usage per deliverable. A breakdown shows where the tokens went, and up to three flags point to likely causes.

Run by each person: run my usage report

2

Page two

The habits guide

Seven habits that drain a limit, each with why it matters and what to do instead. The ones that showed up in your data are marked, and every habit carries an example written for your own job, whether you recruit, sell, keep books or run a company.

Adapts to the person. Nothing is hard-wired to one role.

3

The comparison

The team diagnosis

Drop everyone's reports into one folder. The second skill lines the operators up, flags who is above the team on what, names the likely cause, and gives each person up to three plain changes. A week later it shows whether the gap closed.

Run by the manager: compare these operators' usage reports

It works in Cowork, on any plan. It does not work in Claude Code or in Claude Chat.

If you work alone

A coach for your own habits.

You don't need a team to get value. The report shows you what you could or should be doing differently, using your own session as the evidence.

What you get

Feedback from your own work

One session in, specific changes out

A baseline. Your first run is the number to beat. Later runs show whether a change helped.
Where it went. Instructions Claude re-reads, the conversation itself, file work, skills, web research, memory. Each as a share of the whole.
Likely causes, not verdicts. Flags are labeled "Informed, not yet Diagnosed" because they are leads worth checking.
Habits matched to your job. The guide marks the ones that showed up in your data and says which to start with.

What it asks of you

About a minute at the end of a session

Light enough to actually do

The first time, Claude asks a few setup questions: your name, your role, one sentence about your work, and what counts as a unit of work in your job. It builds a role profile and shows it to you before using it.

After that, you type the sentence, answer one short set of questions (the usage bar now, how many of each task you did, an optional note), and the report is saved to your folder. Claude replies with the file name and one line.

Real numbers

Two reports from one founder's day.

These are David Van Winkle's own Operator Use reports from a single Sunday of building, run as a Baseline and then again after the next stretch of work. The figures are straight from the reports.

Run 1 · 1:21–2:37 AM

Baseline

2.73M

Total usage

7

Deliverables

390K

Per deliverable

17 messages, 75 working steps, 76 minutes. Claude's standing instructions were 42% of usage, re-read on every step.

Run 2 · 2:37–10:14 AM

Build session 2

6.60M

Total usage

11

Deliverables

600K

Per deliverable

18 messages, 78 working steps, 457 minutes. File work was 33% of usage, the largest share.

Where the tokens went

Share of total usage by group, as Claude's usage report names them

Baseline (top bar) Build session 2 (bottom bar) Largest share in that run

Claude's instructions

42%

20%

Our conversation

25%

14%

File work

16%

33%

Skill review cards

7%

22%

Skill text read in

6%

Web research

4%

3%

Your computer's folder

1%

2%

Source: the two Operator Use reports dated October 18, 2026. Usage is weighted for cost from Claude's metered counts, so the split is approximate; read the percentages, not the bar widths. Memory was under 1% in both runs and is left off the chart.

The second run cost more per deliverable, not less. That is the point of the tool. Usage per deliverable rose from 390K to 600K, up 54%, and the report said why: a long build session carried through many rounds of fixes (78 working steps for 18 messages), with report pages re-checked after every round and skill text printed into the chat to edit. Those are specific, changeable habits, and the next run shows whether changing them worked. The report is a measuring stick, not a promise of savings.

If you lead a team

See why one person runs out first.

Everyone on a team can have the same plan and the same tools and still use their limit at very different speeds. Comparing reports side by side turns "she hits the limit more" into a specific, fixable difference.

1

Pick the window

Choose a normal working block

One day, or a five-hour stretch, when everyone is doing their usual job. Label it Baseline. Hold back which habits you're watching until after the Baseline, or people will change their behavior for the test.

2

Each person

Run the report at the end of each session

Each operator types run my usage report. If someone works in several sessions during the window, later runs add to the same day's report instead of starting a new file.

3

The manager

Put the reports in one folder and compare

Ask compare these operators' usage reports. The diagnosis lines everyone up, flags who is above the team on what, names the likely cause, and gives each person up to three plain changes.

4

A week later

Measure again, same window

Repeat with the run label Follow-up and ask run the follow-up comparison. It shows each person's usage per deliverable before and after, and whether the gap closed.

Fair by design

Comparing people honestly

Different jobs, different plans, one fair number

Usage per unit of work compares fairly across any plans and any jobs. A recruiter's candidates and a bookkeeper's reconciliations each become that person's own unit.
The usage bar is a percentage of each person's plan, so it is compared only between people on the same plan.
Sources are marked. Every number is tagged as metered, reported by the operator, or estimated by Claude.

Safe to roll out

Habits, not a performance rating

Present it to the team that way

Counts only. Reports never include names or content belonging to customers, applicants or prospects.
No made-up numbers. If a figure can't be measured, the report says "Not measured."
Flags are leads, not verdicts. The comparison names a likely cause for each, so the conversation starts with a specific habit to try.

Where it works

Cowork, on any plan.

Operator Use is built for Cowork sessions. It measures the session you just worked in, so nobody has to type in numbers from memory.

Works

Cowork

On any plan, for one person or for every seat on a team. Skills need code execution and file creation turned on; on a Team plan, the owner controls that setting.

Does not work

Claude Code

The skills are not designed for Claude Code sessions, so run the report in Cowork instead.

Does not work

Claude Chat

Chat conversations aren't measured by these skills, so run the report on the Cowork sessions where the heavy work happens.

Work outside Claude, such as phone calls or typing straight into a CRM, doesn't show up in the report. It doesn't use Claude's limit either.

What it checks

Seven habits that use up a limit.

Page two of every report. Claude's limit refills every five hours and every week, and some habits drain it much faster than others even when the work gets done.

1Keeping one chat going all day

2Dropping whole files into the chat

3Leaving connected apps on that you don't use

4Asking for long write-ups when a short answer will do

5Using the most powerful model for routine work

6Re-pasting the same reference material every day

7Piling up instructions and skills you no longer use

The kit

Everything in one download.

One zip file with both skills ready to upload, the plain-language guide, and two real example reports.

Skills

The operator skill each person runs, and the token-usage-expert skill the manager uses to compare reports. Each comes as a ready-to-upload skill file and as readable text.

Docs

A README with setup and how to run a measurement, an architecture document explaining how and why it was built, and a short guide to token practices and models.

Examples

The two real reports from this page, as HTML and PDF, plus a blank report template so you can see the layout before you run your own.

Start here

A one-page file that tells you which thing to open first.

Installing a skill, one step at a time

In Claude, open Settings, then Capabilities, and make sure Code execution and file creation is on.
Go to Customize, then Skills.
Click the + button, then Create skill, then Upload a skill.
Pick the skill file from the Skills folder in the kit. It appears in your skill list.
Make sure its switch is on.

On a Team or Enterprise plan, the owner can add a skill once under Organization settings so it appears for everyone. Do one practice run first to confirm the report saves where you expect.

Find out where your Claude limit actually goes.