Deslop Writing

Overview

Strip AI-generated artifacts from existing text while preserving intent and factual claims. Prefer small, local edits over rewrites unless explicitly requested.

Workflow

1) Establish baseline

• Identify the "source of truth" text (before the AI pass) if available.
• Identify the target audience (internal doc, public post, README, PR description, email).
• Preserve facts. If the text makes a claim without evidence, either add the missing specifics or remove the claim.

2) Remove slop patterns

• Delete filler and signposting ("In conclusion", "It's important to note", "Let's dive in", "As mentioned above").
• Replace inflated/marketing language with plain statements.
• Replace vague attributions ("experts say", "many believe") with something checkable or remove it.
• Remove list spam and "rule of three" rhythm when it's not natural.
• Prefer repeating the best word over synonym-cycling.
• Reduce em-dash / bold / emoji habits unless they already match the author's style.

See references/slop-patterns.md for a concrete checklist.

3) Align to the preferred voice (without cosplay)

• Put the punchline first, then the details.
• Use short paragraphs. Cut anything that doesn't earn its spot.
• Make tradeoffs and limitations explicit (what this does not do).
• Sound like a person with skin in the game: specific, candid, non-performative.
• If a repo has a "preferred voice" doc, follow it. If not, default to plain, direct prose.

4) Verify

• Meaning stays the same (unless you explicitly fixed an unsupported claim).
• Names, dates, numbers, and APIs remain correct.
• The result reads like one human wrote it, not a style-transfer demo.

5) Report

• End with 1-3 sentences: what changed, what you didn't change, and any remaining uncertainty.