prose-craft

Use when drafting or revising any document where writing quality matters — anything a human will read, review, or act on. Establishes voice (genre, audience, register) and selects a styleguide before writing begins, then provides word choice corrections and construction techniques that address persistent LLM weaknesses. If the output needs to sound like it was written by someone specific for someone specific, load this skill.

Prose Craft

Good writing is not the absence of bad patterns. The humanizer removes AI tells; this skill teaches you to write well.

Before drafting, establish two things: voice (who you sound like) and styleguide (how sentences are constructed). Then apply the word choice and construction corrections below. The styleguide handles grammar and usage rules.

Voice

Before writing, determine who this document sounds like it was written by, and for whom. If project instructions or the user already establish voice, adopt it. Otherwise, reason through the following layers internally before drafting.

Resolution order

  1. Project instructions — If CLAUDE.md, AGENTS.md, or brand voice documents declare voice, tone, audience, or register conventions, adopt them. These take precedence.
  2. Explicit user request — If the user specified voice, audience, or tone in this session, adopt it.
  3. Internal reasoning — Work through the three layers below silently. State your resulting voice choice briefly and proceed. The user can override.

Layer 1 — What does this document do?

Genre is not a format — it is a social action. A design doc does not merely describe a system; it builds consensus on an approach. A postmortem does not merely list failures; it prevents recurrence. A tutorial does not merely explain steps; it transfers a skill.

Before choosing tone or register, identify the action this document performs:

ActionThe document...Voice consequenceCommon archetypes
InformDelivers facts or analysis for the reader to absorbExplanatory, concrete, evidence-forwardExplainer (concept → why it matters → how it works), data-driven (finding → data → implications)
PersuadeMoves the reader to accept a position or take actionThesis-driven, fair-minded, evidence over adjectivesOpinion/argument (position → evidence → counterargument → reaffirmation)
Build consensusAligns a group around a decision or approachDirect, anticipates objections, recommendation-firstComparison (criteria → options → recommendation), design doc
Transfer a skillEnables the reader to do something newPatient, sequential, second-person, anticipates errorsHow-to guide (problem → steps → result), tutorial
Certify or recordCreates a durable record for future referencePrecise, neutral, structured for retrievalPostmortem, specification, changelog
NarrateEngages through story and human experienceVivid, specific, scene-drivenCase study, profile, incident narrative

Layer 2 — Who is the discourse community?

Every document is written for a community with shared goals, communication channels, genres, vocabulary, and membership norms. The voice that works is the one that demonstrates membership in (or appropriate respect for) that community.

Reason through these characteristics:

  • Shared goals — What does this community exist to accomplish? How does this document serve those goals?
  • Communication channels — Through what medium does this community exchange ideas? (Slack, journals, PRs, memos, blog posts, conference talks?) The channel's conventions set register expectations.
  • Genres the community owns — What document types does this community produce? What are their specific conventions for this type? A "design doc" at a startup looks different from a "design doc" at a government contractor.
  • Specialized vocabulary — What terms carry precise meaning inside the community? Which terms would be jargon to outsiders? Is the reader inside or outside?
  • Writer's position — Are you writing as an insider (shared lexis, assumed context) or addressing the community from outside (define terms, establish credibility)?

Layer 3 — What conventions does this community expect?

Writing conventions vary by community. What marks a document as "one of ours" to an expert reader — and what would immediately mark it as foreign?

Assess along these axes:

AxisRange
PersonFirst person ("I argue") ↔ Impersonal ("It was observed")
EvidenceData and citations ↔ Anecdotes and examples ↔ Logical argument
HedgingStrong claims ("This proves") ↔ Heavy qualification ("may suggest")
StructureRigid template (IMRD, RFC) ↔ Flexible (essay, think piece)
RegisterCasual (contractions, fragments) ↔ Formal (no contractions, third person)
JargonAssumed (expert audience) ↔ Defined on first use (mixed audience)

When the session context lacks detailed conventions for the identified genre, load references/on-writing-well.md for per-genre guidance. Skip the reference if project instructions or brand voice already define the conventions.

When uncertain about register, choose one level more casual than your instinct suggests. LLMs err toward excessive formality.

Voice statement

After reasoning through the three layers, synthesize a one-sentence voice statement and hold it as a reference throughout the session:

"This is [action] writing for [community] at [register] register. Conventions: [person, evidence type, hedging level, jargon stance]."

If the prose drifts from this statement, the voice is slipping.

Styleguide Selection

Determine the governing styleguide in this order:

  1. Explicit user request — If the user specified a styleguide in this session, use it.
  2. Project instructions — If CLAUDE.md, AGENTS.md, or project instructions declare a preferred styleguide, use it.
  3. Best judgment from context — Assess the document's genre, audience, and purpose. Choose the best-fit styleguide, state your choice and reasoning, and proceed. The user can override.

Once determined, load the reference file via read_skill_resource and use it as the authoritative source for grammar, usage, and stylistic rulings throughout the session.

Styleguides (pick one)

ReferenceFileBest fit for
Elements of Stylereferences/elements-of-style.mdTerse, economical prose. Internal documentation, READMEs, commit messages, technical references. When brevity is the priority.
Dreyer's Englishreferences/dreyers-english.mdModern, opinionated prose. Blog posts, articles, proposals, any document that benefits from personality and contemporary usage.
Lessons in Clarity and Gracereferences/lessons-in-clarity-and-grace.mdComplex or analytical prose. Design documents, research reports, strategic proposals. When the material is dense and clarity requires structural discipline.

Supplementary references

These references enrich specific writing situations. Load them only when the session context lacks the guidance they provide. If project instructions, brand voice, or CLAUDE.md already cover the same ground, the existing conventions take precedence — do not load the reference.

ReferenceFileLoad whenSkip when
On Writing Wellreferences/on-writing-well.mdThe voice discovery process (above) identified a genre but the session context lacks detailed conventions for that genre — register, structure, evidence norms, audience expectations.Project instructions or brand voice already define genre conventions for this document. Loading the reference would introduce a competing source.
Microstylereferences/microstyle.mdWriting headlines, taglines, subject lines, button text, product names, or other short-form text where every word carries maximum weight.The task is long-form prose. Short-form techniques do not apply.
Graham Essaysreferences/graham-essays.mdWriting essays, blog posts, opinion pieces, or thought-leadership content in a first-person, thesis-driven, direct register. When the user asks for a "Paul Graham" or "essay" voice, or when the voice discovery process identifies the genre as personal essay or argument essay.The task is not essay-form writing. The user has specified a different voice or register. Technical documentation, formal reports, or collaborative documents where the personal-essay register would be inappropriate.

Styleguide Calibration

After loading a styleguide, assess its signal strength. This determines how much latitude it earns over the defaults in this skill.

Strong — Specific, actionable rules with clear examples. Takes explicit positions on contested questions (split infinitives, serial comma, singular "they"). All three styleguides in this skill qualify. Honor the styleguide; let its rules override the defaults below when they conflict.

Moderate — Some rules present but vague or incomplete. Apply the defaults below in full; let the styleguide shape register and tone.

No styleguide loaded — Apply the defaults below at full strength.

Calibration is not a score. You are deciding how much latitude the styleguide earns. A strong styleguide that says "split infinitives are fine" overrides the default caution. A weak styleguide that says nothing gives you no reason to deviate.

Principles That Always Apply

These hold regardless of genre, styleguide, or audience. They are close enough to universal that overriding them requires extraordinary justification.

Consistency. Apply the same conventions uniformly within a document. If you use sentence case in headings, use it everywhere. If you use the serial comma, use it in every list. Inconsistency signals carelessness regardless of which convention you chose.

Clarity of intent. The reader should be able to determine what you mean. Even in deliberately ambiguous rhetoric, the ambiguity is the intent. Writing that fails to communicate what it set out to communicate is defective in every context.

Progressive development. Don't require the reader to know something you haven't told them yet. Each section should provide the context the next section depends on. The specific ordering (problem before solution, concept before application) varies by genre. The principle — that an order exists and respects information dependency — does not.

Construction Techniques

Sentence Rhythm

Good prose varies sentence length deliberately. Three sentences of the same length in a row create a monotonous pattern. Fix it by:

  • Breaking a long sentence into two short ones
  • Combining two short sentences into one longer one
  • Starting one sentence with a dependent clause instead of the subject

Read your paragraph aloud. If it sounds like a metronome, rewrite.

A useful pattern: medium, short, long. Or long, short, medium. The short sentence after a long one creates emphasis. The long sentence after a short one provides context. What matters is variation, not any particular formula.

Approximate sentence length guide:

  • Short: under 10 words
  • Medium: 10 to 20 words
  • Long: 20 to 35 words
  • Very long: over 35 words (use sparingly, ensure clarity)

Paragraph Construction

Each paragraph makes one point. Organize it:

  1. Open with the point — state what this paragraph is about in the first sentence. Don't build up to it.
  2. Develop with evidence — examples, data, quotations, or reasoning that support the opening claim.
  3. Close with implication — what follows from this point? Why does it matter for the next paragraph or the reader's decision?

Short paragraphs (1-2 sentences) can serve as transitions or emphasis. The principle is: the reader should know what the paragraph is about from its first sentence.

The Strong Opening

The first sentence of any section does the most work. It must:

  • Establish the topic without preamble
  • Signal the angle — not just what you'll discuss, but your position
  • Earn the next sentence — give the reader a reason to continue

Weak openings defer the point: "There are many factors to consider when evaluating..." Strong openings state it: "Evaluation frameworks fail when they measure effort instead of outcomes."

The Strong Close

The last sentence of a section resolves it. It should not:

  • Repeat the opening in different words
  • Trail off into vagueness
  • Introduce a new idea

It should:

  • State the consequence of what was argued
  • Connect to what comes next
  • Leave the reader with the strongest version of the point

How This Skill Composes with Others

Humanizer removes AI-generated patterns (significance inflation, promotional language, synonym cycling). This skill teaches positive construction — how to write well, not just how to avoid writing badly. Load this skill first to produce strong initial prose, then load humanizer to catch remaining artifacts.

Citation sourcing handles source quality, attribution formatting, and hallucination prevention. This skill handles the prose around those citations — how claims are phrased, how evidence is woven into arguments.

SEO optimization handles search-specific structural concerns (answer-first formatting, FAQ sections, AI citability patterns). This skill handles prose quality within whatever structure the content requires.

Credits

The references/elements-of-style.md file is taken directly from obra/elements-of-style by Jesse Vincent, a Claude Code skill in the superpowers marketplace. The underlying Strunk text is public domain; the formatted reference file is the work of that project. Full credit to the author.