# Pre-registration: what does the trending Karpathy CLAUDE.md actually cost?

WRITTEN BEFORE ANY MEASUREMENT RAN. Timestamp from `date` at the moment of writing, below.

## The subject

`andrej-karpathy-skills`, trending on GitHub at 491 stars in a day on 2026-08-24, described as
"A single CLAUDE.md file to improve Claude Code behavior, derived from Andrej Karpathy's
observations on LLM coding pitfalls."

Attribution, checked rather than assumed: the trending listing is `multica-ai/andrej-karpathy-skills`,
the plugin manifest names `forrestchang` as author, and the raw URL inside the repo points at
`forrestchang/andrej-karpathy-skills`. Treat multica-ai as a re-host. Commit measured:
2c606141936f1eeef17fa3043a72095b4765b9c2.

The file is **2,357 bytes, 65 lines**. It is NOT plain prose: markdown headers, bold runs, bullet
lists, a fenced code block and arrow characters.

## Why this is worth measuring rather than describing

This site already publishes a cost model for CLAUDE.md files, and it makes a prediction here. So
does an older corpus, and **the two disagree**:

- **Prediction A**, from the `/cost/` calculator's curve (`overhead 106, perByte 0.26`, stated to
  hold from 2KB to 28KB): **106 + 0.26 x 2,357 = 718.8 tokens**
- **Prediction B**, from `/claudemd-per-turn-2-1-233.json`, which measured 18,000 bytes of
  ordinary English prose at 0.2759 tokens/byte: **0.2759 x 2,357 = 650.3 tokens**

They differ by 68.5 tokens. Both are this site's own published figures.

The older corpus carries a limit that bears directly on this file: *"This site has measured the
byte-to-token rate varying from 0.26 to 0.37 with content, so a file of tables or code will not
cost the same per byte."* The Karpathy file is markdown-structured, not flowing prose, so it sits
in exactly the cell that limit warns about.

## What is measured

The startup context of a trivial run in an isolated sandbox, with and without this exact file at
the project root. The difference is what the file costs.

## Declared BEFORE the runs

- Both predictions above are recorded now and will NOT be revised after seeing the result.
- If the measured cost falls outside BOTH, that is a finding about the transferability of this
  site's own curve, and it must be reported as such rather than quietly re-fitting.
- If it falls inside one, say which and by how much.
- A single number is not a result: report the observed range across runs, and prefer stating that
  ranges do or do not overlap over quoting a percentage.

## Controls, all published per run

- `--setting-sources project` asserted onto the command line.
- The file's byte size read back off disk on the with-file arm; its ABSENCE asserted on the other.
- ONE constant working directory for every run, because this site measured that the startup floor
  moves with the working-directory path. The harness this is adapted from used a different
  directory name per arm, which would put a path term of up to roughly 80 tokens inside a signal
  of roughly 700.
- `modelUsage`, not `.usage`. Input, cache creation and cache read kept separate.
- A 429 or 529 is recorded as `not_a_measurement`, never as a zero.
- Four runs per arm minimum.
Mon Aug 24 18:21:00 IST 2026