Skip to content
Back to skills

Writing

ASecurity

Use when drafting or revising human-facing prose, titles, articles, posts, marketing copy, READMEs, or docs; also when asked to humanize text, remove AI writing patterns, or match a writer's voice. Keep code cleanup in its own lane.

  • 22 stars
  • 0 votes
  • 0 copies
  • 2 views
  • Added September 30, 2026
researchgoreact

Security analysis

A100/100

Pro scans all 6 files and shows the line behind each finding

Scanned September 30, 2026

npx -y skills add lev-os/agents --skill writing --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Writing?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Writing
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/lev-os-writing/badge)](https://www.skillsdirectory.com/skills/lev-os-writing)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: writing
description: Use when drafting or revising human-facing prose, titles, articles, posts, marketing copy, READMEs, or docs; also when asked to humanize text, remove AI writing patterns, or match a writer's voice. Keep code cleanup in its own lane.
---

# Writing

Make something worth understanding clear to its intended reader. Preserve what
the evidence says and the writer actually means.

Read this file and all four references completely before editorial work:
[patterns](references/patterns.md), [approved language](references/approved-language.md),
[review](references/review-contract.md), and [coverage](references/coverage.md).
Also read the complete house voice and supplied examples when the caller names
them. Finish truncated reads. A delegated worker loads the material in its own
context; a parent's reading is not inherited proof. Reload after a context reset.

The references supply editing detail. This file owns the procedure. Source and
claims constraints always apply; explicit user direction and a supplied writing
sample or house voice govern style, then the requested format. Preserve exact
meaning over punctuation preferences. Read every reference, then apply the
branches that fit the job; a title request does not require a full article.

## 1. Establish the job

Name one reader in a real situation, the question or decision, the requested
format, source text, permitted claims and voice owner. Read an entire named
file before editing it. When a sample is supplied, note its sentence length,
diction, paragraph openings, punctuation and transitions; preserve deliberate
quirks. For a technical explanation or incident record, faithful explanation can
be the whole contribution; personality is optional.

For explicitly fictional work, invented story details belong to the requested
genre; do not present them as real evidence or personal experience.

For authored editorial work, distinguish fact, interpretation and thesis.
Identify the useful calculation, comparison, synthesis, operating distinction,
verified experience or supported judgment. An owned opinion comes from a human
or approved institution, not a model pretending to have experience. State the
strongest limitation and whether the value is timely, durable or reusable.
Return a missing premise to the author instead of polishing an empty argument.

## 2. Earn the title and opening

Before prose, state the reader payoff and the evidence that earns it. For a
headline assignment, develop at least five materially different angles when
the source supports them: finding, consequence, concrete tension or number,
mechanism, and an ordinary-language question or judgment. Do not fill a quota
with synonyms or unsupported promises; identify the evidence ceiling instead.

A cold reader should recognize the subject and a reason to read. Put secondary
scope, dates and formal labels in the deck when the title remains true without
them. Keep proposal status and any qualification necessary to avoid a false
claim in the headline. Avoid unexplained acronyms, persona prefixes, source-file
labels, vague clues and recurring contrast formulas. Compare adjacent work for
repeated mechanics. Length is a readability check, not a source of truth.

Write the title, deck and opening together. The opening two paragraphs must pay
the promise; the conclusion must answer it. If a new title changes the premise,
revise the argument before declaring the retitle complete. Short work has the
same truth and usefulness duties as long work; never pad a complete argument.

## 3. Draft, then edit manually

Draft from the approved premise and sources. Put evidence near material claims;
keep publication, event, effective and measured-outcome dates distinct. Use
concrete nouns and mechanisms, simple verbs, and rhythm suited to the writer.

For humanizing, preserve the sample's deliberate habits, uncertainty and real
human details. For cleanup, use the full pattern reference and read line by
line: restore specifics, fix clusters, and remove unsupported inflation. Never
invent a story, reaction, factual detail or citation to make prose feel human.

When a sentence remains awkward, rewrite the paragraph around its actual point
rather than patching one flagged phrase at a time.

For mannered prose, replace each flourish with the literal statement it avoided.
Keep meaning and structure, leave direct sentences untouched, and do not shorten
for its own sake. Keep an internal table of flourish → plain wording → implied
connotation, and name the repeated habit. Show that table when requested; keep
it out of publishable copy.

## 4. Compare and verify

Ask: “What still sounds generated?” and “Did the edit add or remove any fact,
name, number, date, quote, citation, ranking or other claim?” Repair unsupported
additions and lost claims. Then apply every relevant lane in the
[review contract](references/review-contract.md).

The calling publication owns delegation and model selection. If it requests an
external variant pass, use `skill://coder` for the exact authorized provider and
model, an independent input packet and a bounded result. Never use Luna unless
the user explicitly requests it for that run; never silently substitute a model.
Generic writing does not require a worker launch. Do not let a variant author
approve its own work. Compare anonymous candidates against sources and examples,
not model reputation, and allow reject-all. Model preference is not owner taste.

## 5. Return the requested result

For pasted humanizing work, the default is draft → short remaining-pattern
critique → final. A request for final text only overrides that display. For a
named file, write only final prose and preserve code, YAML/data and link targets
unless explicitly asked to change them. Embedded work returns the final text.
For recommendations, show proposals without editing. Automations show readable
output and compact editorial reasoning before artifact links.

## README opening and example
For a README, put audience, purpose, and the first useful action in the opening. Include one command or example verified against current package scripts or source, and state only limitations that affect use.

Files in this skill

  • LICENSE.humanizer1 KB
  • SKILL.md5.9 KB
  • references/approved-language.md1.6 KB
  • references/coverage.md1.8 KB
  • references/patterns.md28.6 KB
  • references/review-contract.md1.9 KB

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…