second-draft

A Claude skill that revises prose until it reads like a careful human wrote it.

AI output is a first draft that performs writing: it announces significance instead of demonstrating it, hides actors behind passive voice, and falls into rhythms a reader can predict. Most tools attack this with banned-word lists, and the result is text that dodges "delve" while still saying nothing. second-draft works upstream. It diagnoses the five causes that produce AI-sounding text, then runs a four-pass revision that fixes them at the source.

Install

For Claude Code, clone this repo into your skills directory:

git clone https://github.com/dreulavelle/second-draft ~/.claude/skills/second-draft

On Windows:

git clone https://github.com/dreulavelle/second-draft "$env:USERPROFILE\.claude\skills\second-draft"

Restart Claude Code (or start a new session) and the skill registers automatically. Invoke it with /second-draft, or let it trigger on its own when Claude writes human-facing prose. To confirm it loaded, run /skills and look for second-draft in the list.

Elsewhere:

• Claude Projects: upload SKILL.md and the three files under references/ to project knowledge.
• API or system prompts: include SKILL.md directly; it links to the reference files, which you can attach as needed.

How it works

SKILL.md carries the model: every AI tell traces back to one of five diseases.

Disease	What it produces
No stakes	Hedges, filler, claims nobody could dispute
Performed emphasis	"Let that sink in," trailing "-ing" clauses, bold-everything formatting
Symmetry addiction	Triads everywhere, telegraphed reversals, metronomic sentence lengths
Missing actor	"The decision was made," "the data shows," "the culture shifted"
Abstraction	"Streamlines workflows" where "drops setup from 40 lines to 6" belongs

Three design choices separate this from a ban list:

Budgets instead of bans. Human writing contains em dashes, triads, and adverbs; AI text is marked by their frequency, and over-corrected AI text is marked by their total absence. Each feature gets a budget (em dashes: at most 2 per 500 words) rather than a prohibition. Zero is a fingerprint too.

Register awareness. Passive voice is a bug in an essay and a feature in API reference docs. Repeating the same word is a sin in prose and mandatory in technical writing. The skill classifies what's being written (reference, tutorial, commit message, essay, UI text, correspondence) before applying any rule, so it won't wreck your docs to satisfy essay rules. See references/registers.md.

Falsifiable checks instead of scores. Self-assigned quality ratings are noise with a number attached. The skill ships six tests with yes/no answers, including the disagreement test (could a reasonable person claim the opposite of this sentence? if not, it's filler) and the pointing test (can you point at the concrete thing this abstract noun stands for?).

Example

Before:

The team adopted trunk-based development in 2024, marking a pivotal shift in their engineering culture and underscoring the importance of continuous integration.

After:

The team switched to trunk-based development in 2024. Merge conflicts that used to eat the Friday before each release mostly disappeared, and CI time per change dropped by half.

The participle clauses asserted significance the sentence never earned. The revision replaces them with the observable consequences that would justify it. More before/after pairs, including one showing how ban-list over-correction produces a different robot voice, live in references/revision.md.

Files

SKILL.md                  Core model: diseases, workflow, budgets, checks
references/tells.md       Symptom catalog mapped to causes, graded by severity
references/registers.md   Per-register rules, including where prose rules invert
references/revision.md    Four-pass procedure with before/after examples

Grounded in corpus research

The vocabulary catalog draws on measured data rather than intuition:

• Kobak et al., Science Advances (2025) identified 319 "excess vocabulary" style words by measuring frequency jumps across 15 million PubMed abstracts after ChatGPT's release; at least 13.5% of 2024 abstracts carry LLM fingerprints.
• Juzek & Ward (2024) trace the overrepresentation of words like "delve" to RLHF rather than pretraining data.
• Wikipedia's signs-of-AI-writing guide, maintained by editors who clean up AI text at scale, supplied structural tells the word lists miss: copula avoidance, trailing participles, weasel attribution. It also documents the case for frequency over single hits — humans distinguish AI from human text at roughly chance, and one marker word is coincidence while four in a paragraph is a fingerprint.

What it won't do

It won't shorten at all costs; a sentence that gains the missing specifics gets longer, and that's the point. It won't chase individual words, since the diseases sit upstream of vocabulary. And it won't apply essay rules outside essays — check the register section before judging it for leaving passive voice in your API docs.

License

MIT