Mods

cells

Several edits and their test run in one step.

It comes with the pack and is on from the start.

The moment

You ask Claude to change two small settings and check that the tests still pass.

Without cells

Claude reads the file, makes the first edit, makes the second, then runs the tests: four steps, and each one sends the whole conversation back to be read again. Or it writes a quick script that changes the files through the shell. That is cheap, but you see no diff, and /rewind can’t undo it.

With cells

Claude writes a short program, a cell, whose lines are the real tools: read, edit, run. The two edits and the test run happen in one step. Each edit still goes through Claude Code’s own edit tool, so you see the diff, your permission prompts apply, and /rewind undoes it.

Flip the switch and watch four steps fold into one.

Show
  1. StepRead the fileawait read('src/app.py')
  2. StepFirst editawait edit('src/app.py', 'retries = 3', 'retries = 5')
  3. StepSecond editawait edit('src/app.py', 'timeout = 10', 'timeout = 30')
  4. StepRun the testsconst t = await sh('python -m pytest -x -q tests/test_app.py 2>&1')
One stepOne cell
await read('src/app.py')
await edit('src/app.py', 'retries = 3', 'retries = 5')
await edit('src/app.py', 'timeout = 10', 'timeout = 30')
const t = await sh('python -m pytest -x -q tests/test_app.py 2>&1')
return t.ok ? 'tests pass' : fold(t.out)

Four steps, and each one reads the whole conversation again.One step, one read of the conversation. Each edit still shows as a diff, and /rewind undoes it.

What we measured

Of all the mods, cells saves the most. Take it out of the pack and the same jobs cost 9% more8.9% more expensive, 95% range 2.1 to 16.4% more expensive8.9% more expensive
95% range 2.1 to 16.4% more expensive
4 per condition on multi-step tasks, 2 on small ones
Opus 5.5, 13 Oct 2026; the feature job took 14.2 steps instead of 9.814.2 against 9.814.2 against 9.8
4 per condition
Opus 5.5, 13 Oct 2026.

Every measurement for cells

The dot is the measured change, the bar around it the 95% range, the upright line no change. Left of the line is cheaper or fewer.

Tried for cells and dropped: Moving text before the cache boundary.

Where it doesn’t help

A cell only helps when Claude already knows its next few moves. Where each step depends on what the last one showed, it still takes them one at a time, and on a quick question there is nothing to fold.

Turn it on or off

/mods cells off turns it off, and mode: nudge keeps only its short notes, with no tool. Every option is in the README; the code is in cells/.

In a real session

A real Claude Code session: asked to add two docstrings and run the whole test suite, Claude runs one cell; a card shows the read, the two edits, the diff and the test run live, and the tests pass. Play the recording