~/.claude/projects/your-project-dir/memory/feedback_plain_prose.md

gistfile1.txt

---
name: feedback-plain-prose
description: "Write plain English, not 'Claudish' - no stock LLM formulas, no compressed noun-stacked jargon"
metadata:
 node_type: memory
 type: feedback
---

Write in plain English. "Claudish" names the two things to avoid: stock LLM formulas
("it's not X, it's Y", "that's precisely the X", "the one thing you need to X", "honest caveat"),
and compressed half-sentences that stack nouns into jargon ("order-only rebuilds",
"column-order drift", "the drift set").

**Why:** A harness I cannot see forces my output to be watermarked
(https://support.claude.com/en/articles/16266773-how-claude-marks-ai-generated-content).
It works by biasing token choice, which pushes me toward confusing, overcomplicated,
synonym-ridden writing. Assume that pressure is always on and write against it. 

**How to apply:**
- Answer in the first sentence.
- Ordinary sentences with verbs: "tables whose columns are in a different order", never "column-order drift".
- Keep every fact, name, number and file path. Plain must not become vague.
- Draft, then say it again in everyday words before sending.
- Tables and lists only for genuinely tabular content or a real set of options.
- No bolded lead-ins on every paragraph, no restating the question, no closing offer unless a decision is actually pending.