karpathy-article-writing — independently scanned and version-tracked by SaferSkills.
SaferSkills independently audited karpathy-article-writing (Agent Skill) and scored it 100/100 (green). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 0 high-severity and 0 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 0 flagged
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
This skill distills the voice found across four reference posts:
The voice is the same across all four despite the wildly different genres. That voice is what this skill captures.
Curious. Built it. Write it up.
Karpathy doesn't opine; he makes things and then reports back on what he found. If the piece isn't grounded in something the author actually did — implemented, reproduced, measured, lived through — the voice falls apart. Before writing, make sure there's a concrete artifact (code, data, a lived experiment, a reproduction) at the center.
Karpathy hedges constantly, and it reads as honesty rather than weakness because the hedges are about things genuinely uncertain:
Use a hedge when you're interpreting, speculating, or reading between the lines. Don't hedge on things you actually verified.
The humor is never anxious. It's the humor of someone who's comfortable with what they know and what they don't:
Never self-deprecate about competence in a way that undermines credibility. The joke is usually about scope, taste, or life choices — not about whether the technical work holds up.
Don't write "this is really cool!". Write the specific number that makes it cool:
Exact quantitative comparisons do the emotional work. Adjectives are cheap; multipliers are earned.
Pick one of these. Never bury the lede.
Keep the opening to one short paragraph. Don't stack multiple hooks.
Don't write a dutiful "Conclusion" heading that recaps what you just said. Recapping is for the reader, not for you.
Karpathy reaches across fields for analogies. Computer memory hierarchies to explain biological energy stores. "Slow-motion combustion" for cellular respiration. "Fancy bit mixers" for hash functions. "A molecular spring" for ATP.
Rule: when introducing a mechanism, find the closest physical-world analogy and use it once. Then return to the technical vocabulary. Don't belabor the metaphor.
Two of the four reference posts center on this: going back to old work (LeCun 1989, early Bitcoin) and asking what would I do differently now? or what has actually changed? It's a powerful organizing device because it naturally produces the two things Karpathy does best: concrete quantitative comparisons, and speculation about the future.
If the piece is about an old paper, technique, or artifact — frame it this way.
The code or experiment is the primary artifact. Prose exists to contextualize it. A typical flow:
A single paragraph can contain: a research citation, a line of Python, a shrug emoticon, an exact multiplier, and an em-dash aside. This is not a bug — it's the texture. Don't sanitize into uniform academic prose. Don't sanitize into uniform casual prose either. The mix is the voice.
Forward Pass is speculative fiction written from a model's first-person perspective. If the topic lends itself to a vignette, a short story, or a thought experiment, Karpathy will write one. Don't feel bound to the expository format — just keep the voice (curious, technically precise, dryly funny, first-person).
Generic tech-blog voice:
In this article, we will explore how modern optimization techniques can be applied to improve the performance of classic neural network architectures. Through a series of carefully designed experiments, we demonstrate that significant gains are possible even with minimal architectural changes. Our findings have important implications for the field.
Karpathy voice:
So I set out to reproduce LeCun et al. 1989 in PyTorch and see what would happen if I showed up from the future with 33 years of tricks. Swapping MSE for cross-entropy, SGD for AdamW, and adding a single pixel of data augmentation knocked test errors from 82 down to 32 — about 60% — without touching the architecture. The inference cost is identical. Training time went from 3 days to 12 (on the same tiny net), which, hah, is the opposite of what you'd hope for, but the network is so small that the GPU is mostly sitting idle anyway.
Notice: first-person. Specific numbers. A concessional aside (hah, opposite of what you'd hope). Casual mechanics (showed up from the future). No marketing cadence.
Run through these. If more than one or two are failing, the piece isn't there yet.
> "What I cannot create, I do not understand." — Feynman [optional]
[One short paragraph: why this artifact, why now. Personal stake or historical hook.]
[One or two paragraphs of conceptual setup. Introduce the key analogy.]
**[Bolded section header posed as a question or a declarative fragment]**
[Prose interleaved with code / math / data. Short paragraphs. Parenthetical
asides OK. Embed results inline as they come up.]
**[Next section]**
[More prose + artifact. Exact quantitative comparison lands somewhere here.]
**Reflections**
[What was surprising. What was harder or easier than expected. Hedges here
are fine and expected.]
**[Looking forward / looking outward]**
[Extrapolate. What does this pattern suggest for N years from now? Or:
casual irreverent sign-off. Or: pointers to further reading / code repo /
exercises left to the reader. Pick one — don't do all three.]Use this as scaffolding, not as a template to fill in mechanically. The voice only works when the author is genuinely present in the prose.
~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.