# Why the skill reads as useless (2026-09-02, interim)

1. Fresh sonnet-4-6 docs, no skill, setting-sources off: 0.00 viol/100w on both runs. Skill runs: 0.98 and 1.27. The baseline the README beats is not what the user sees today.
2. Committed baseline (56 gens): slop_word = 0. Headline 74.6% = synonym_rotation 33, sentence_over_limit 27, semicolon 16, trailing_condition 16, contraction 11, banned_modal 8. Grammar obedience, not readability.
3. Replies: skill reply keeps em-dashes (12 base / 17 old / 13 plain over 8 files; fresh run 9,7 vs 6,4), bold spans 41/25/27, 14-16 sentences vs cap 5, ~190-250 words. Loophole: "Code blocks and list items do not count" so the model bullets everything.
4. Claude Code's own prompt already says short, direct, no headers for simple questions. Hook adds ~5.9k cached tokens per session for that.
5. The decisive run below: an 8-line prompt against the full skill on 8 reply scenarios.

## Decisive run (8 reply scenarios, sonnet-4-6 low, setting-sources off, 2026-09-02)
| cond | words | sentences | <=5 sent | em-dash | bold | headers | bullets | linter viol/100w |
|---|---:|---:|---:|---:|---:|---:|---:|---:|
| no skill | 249 | 18.9 | 0/8 | 32 | 52 | 19 | 35 | 5.93 |
| skill 2.0.0 (system-prompt.md) | 171 | 13.4 | 0/8 | 17 | 24 | 3 | 22 | 2.65 |
| micro prompt (8 lines) | 150 | 5.8 | 5/8 | 2 | 0 | 0 | 0 | 3.34 |
Skill wins only on its own linter. Micro wins on everything a reader sees.
Root cause: the skill optimizes for evals/ste_lint.py, and 50+ rules dilute the 5 that matter. Loophole: "list items do not count" toward the cap.
Files: why-useless-2026-09-02/ (base__, skill__, micro__ replies; doc__ and reply__ pairs; meta__ answer). Scored with evals/ste_lint.py reader_check, which counts list items as sentences.