All skills
oaustegard avatar

/declauding

@6777f98

Load at the start of any task whose deliverable is prose another person will read — a PR description, commit message, README, doc, postmortem, blog post, issue or review comment, report, release note or essay — before drafting it, whether the result is pushed, posted, published or saved to a file, without being asked. Draft, then run this pass on the draft before handing it over. Also use when someone says "de-claude", "de-slop", "humanize this", "this reads like AI", "make it sound human", or asks for a voice, tone or register edit. Rewrites the constructions that mark prose as model-written (staged reveals, verdict headers, aphoristic closers, "it's not X, it's Y", em-dash drama, forced triads, flat-certainty adverbs) into plain technical prose and checks the rewrite kept every claim. Not for fiction, poetry, code, or quoted text; for a full adversarial review of a deliverable use challenging.

  • 23 files
  • 249 KB
  • Updated last week
  • GitHub

Use this Skill: https://skilld.dev/gh/oaustegard/claude-skills/declauding

This session only. Nothing lands on disk.

referencesmodelssonnet.md

≈320 tokens on demand. Your agent reads this file only when SKILL.md points to it.

Sonnet 4.6 and Sonnet 5

Editing a draft Sonnet wrote

  1. Work through the linter's em-dash and negation-first flags first; on Sonnet 5 they carry most of the pass. Rewrite the reversal as a plain statement (entry 2): The cache wasn't broken — it was doing exactly what we asked it to do. Apply the entry-16 dash guard unless the author's sample uses dashes at that rate.
  2. On Sonnet 4.6, merge fragment runs into sentences (entry 15): The p50 improved. The p99 got demolished. Cut subjectless asides (entry 35): Straightforward stuff.
  3. Read the closers; both versions end some paragraphs on a verdict (entry 12).

Running the pass as Sonnet, or delegating it to Sonnet

Sonnet cuts the evidence along with the framing: "5,697-line gather cut at line 120" became "thousands of lines" in the PR #791 audit.

  • Run declaude_diff.py on every draft, however short, and restore each LOST number, date or command value.
  • In a delegate's brief, state that dates, measured numbers and copy-pasteable values stay even when the sentence around them is cut.
  • Review every edit that is mostly deletion before accepting it.

Evidence: oaustegard/experiments model-register-drift/ (three samples, one prompt) and claude-skills PR #791.

Source: SKILL.md on GitHub

No third-party reports yet.

Signed by skilld at 6777f98. This ties the file your Agent reads to that commit on GitHub. It does not review the instructions.

Last checked against GitHub yesterday.

Activeupdated last week
metadata
{
  "version": "0.9.3"
}

README badge

README badge for oaustegard/claude-skills/declauding