Anthropic removed the temperature parameter. Pictionary puts it back, in Gaussian blur.
The newest models (Fable 5, Opus 4.8/4.7) return a 400 if you send temperature;
the official guidance is to "prompt for it." Pictionary disagrees: it renders your
prompt to an image and physically smudges it until the model misreads its way
into being creative. The misreads are the sampling noise.
I measured it. On a question Opus normally answers identically, turning the knob up made the output 3x more varied:
| temperature | regime | output diversity |
|---|---|---|
| 0.0 | pristine scan | 0.25 |
| 0.4 | office photocopier | 0.74 |
| 0.8 | fax machine | 0.75 |
claude-opus-4-8, 8 runs/level, balanced density. Diversity = mean pairwise edit
distance across runs (0 = identical). Reproduce: pictionary bench --runs 8.
A meme, built petite, not a serious sampling system. The knob is, annoyingly, real.
npm i -g pictionary
pictionary install-skill
pictionary pack prompt.txt --temperature 0.8
Then, in Claude Code, Read the PNG instead of the text file, the only stochastic
decoding left on the frontier.
| t | regime |
|---|---|
0 |
pristine scan (deterministic-ish) |
0.4 |
office photocopier |
0.8 |
fax machine |
1.0 |
photocopy of a fax of a photocopy |
Billing is unchanged (same pixels). Fidelity is not. That's the point. The smudge scales with font size, so the knob bites at any density, not just the tiny one.
Same four lines, four temperatures, pristine scan to photocopy-of-a-fax:
Claude bills images at ~(w × h) / 750 tokens, capped near 4,784/image; text at
~4 chars/token. Pack enough characters under that cap and a page of text bills like
a fraction of itself, up to ~3.8x cheaper at max density. Real files rarely fill
every line, though, so run pictionary pack --dry-run first, it tells you whether
packing your actual file wins.
| density | font | best case | fidelity |
|---|---|---|---|
conservative (default) |
~14px | ~1.05x (often break-even) | near-lossless |
balanced |
~12px | ~1.6x | good |
max |
~8px | ~3.8x | lossy risk |
Pages land in the billing sweet spot (≤2576px long edge, ≤3.75MP) so nothing silently
downscales; long inputs paginate to file.p1.png, .p2.png, …
pack <file> [--density …] [--temperature 0..1] [--out dir] [--dry-run]: render the PNG(s) + a savings report.--dry-runprints the token math without rendering.bench [--prompt-file f] [--runs N] [--density …]: measure whether the smudge actually moves the output, viaclaude -pover your OAuth login (no API key, no SDK).install-skill, teach Claude Code when to reach for this, and when not to.
- OCR fidelity: the model reads pixels; misreads happen (the feature at high temperature, the bug at low). Never pack anything byte-exact: IDs, hashes, code you'll edit.
- No prompt caching: image reads skip the cached-prefix discount. Best for read-once bulk.
- The extra turn: in an agent, Reading a PNG is another tool call that re-sends
the conversation; on a short chat that can erase the packing savings. It wins when
the file is big and the surrounding context is small or already cached.
benchmeasures it end-to-end, not just the static math.
Keep load-bearing text as text; pack the boring bulk. The bundled skill encodes exactly that.
pxpipe does the token savings for real: a proxy that auto-images your API traffic with profitability gates and production benchmarks (~3.1x on dense content, independently confirming the math). Pictionary is the temperature knob wearing a party hat. Want a lower bill? Use pxpipe. Want your sampling parameter back? Right repo.
No proxy, no MCP server, no hooks, no auto-interception, no telemetry, no config files.
One runtime dep (sharp). Petite by design.
MIT.