| name | writing-voice |
| description | The user's full writing system for prose they author — README and docs text, landing copy, PR and commit descriptions, Linear proposals, design docs, Slack and email messages. Use when drafting or rewriting any human-facing prose, when a draft reads like AI wrote it, when stripping AI slop from documentation, or when relaying research output into a human-facing surface. CLAUDE.md carries the compressed rules; this carries the reasoning, the before/after pairs, and the per-format shapes. |
Writing voice
CLAUDE.md holds the always-on version: Orwell's six rules, the corrective
juxtaposition ban, and the mechanical rules. Those are the floor and they apply
whether or not this skill is loaded. What follows is the reasoning behind them and
the parts that only matter for a specific format.
What this governs
Prose the user authors themselves: messages, replies, docs, comments in their own words.
It does not mean rewriting AI-generated analysis into something that sounds like
them — that misrepresents authorship. When relaying your own research or analysis
into a human-facing surface (a PR comment, a Slack message, a doc): write a short
framing sentence in their actual voice (why they are including this, what they want the
reader to do with it), then paste the analysis verbatim in a fenced code block or a
clearly attributed section, unedited. Do not reformat it to match the bullet,
paragraph, or dash rules below — those apply to their words, not to a quoted block.
Reference pattern:
https://github.com/coverbasedev/irm/pull/13294#discussion_r3619540311
The six rules
- Never use a metaphor, simile, or figure of speech you are used to seeing in print.
- Never use a long word where a short one will do.
- If it is possible to cut a word out, cut it out.
- Never use the passive where you can use the active.
- Never use a foreign phrase, a scientific word, or a jargon word if an everyday
English equivalent exists.
- Break any of these rules sooner than write something outright barbarous.
They never touch code, identifiers, or precise technical terms. Swap in everyday
words only where precision survives. Review every prose output against them before
delivering.
When rewriting existing text, first name each violation (stale phrase, long word
with its short replacement, cuttable word, passive construction), then give the
rewrite with every fact, number, and name unchanged.
Worked before/after pairs, the full tic catalog, and ready-to-run rewrite prompts:
voice-examples.md.
Rejecting a draft
When a draft reads wrong, do not just delete and retry. Name the exact reason it
failed — which rule it broke, which tic it used — so the failure mode is fixed
rather than banned one word at a time. "Sounds like AI" is not a reason; "opened
with 'it's not just X'" is. Hold that reason across the session so the same tic does
not come back in a new form.
Banning "delve", then "robust", then em dashes, one at a time, never converges. The
six rules catch the class, not the instance.
Proposals and longer docs (Linear, design docs)
Order: problem first, then options considered, then the decision. Lead the reader to
the conclusion rather than opening with it.
Trim to essentials and link or fold the rest; reviewers can ask for more.
Reference code by GitHub permalink only, if at all. Avoid inline file:line
references.
At most one or two collapsible <details> sections, so the post stays scannable.
Tables stay compact (under ~120 characters wide) and high-level so they are easy to
hand-edit. Push detail into prose, not cells.
Validate any external links before including them. Cite docs, blog posts, or SDK
references where they back a claim.
Name concepts with common architecture or Python terms (responsibility, contract,
Protocol, extension point) rather than coined metaphors like "seam".
When proposing an abstraction, include thin pseudo-Python of current versus target
so the delta is concrete. Keep snippets reaction-sized, not implementation-ready.
For diagrams, load the mermaid-diagrams skill rather than restating its rules
here; it covers type selection and density, and Linear renders mermaid natively.