de-llmlisted
Install: claude install-skill tae898/de-llm
# De-LLM: Remove the Tells, Keep the Meaning
Generated prose is rarely wrong. It is recognizable. This skill removes what makes it recognizable.
Every claim here is measured or cited, and `references/sources.md` says which,
including what each source does **not** establish and which claims have no source
at all.
## Read the register first, then pick a tier
The tells are not universal, and running the wrong tier first wastes the pass.
**Structure and vocabulary rise in every register with a baseline**, two to four
times. Tier 2 and Tier 3 are the ones to trust everywhere.
**Formatting is register-specific and usually is not a tell.** Decide by how the
document was written, not by whether it contains markup:
| Document | Start at | Because |
|---|---|---|
| Written end to end by an agent | **Tier 1** | em dashes and bold lists run about ten times a maintained repository |
| A README a human maintains | **Tier 2** | em dashes have *fallen* since 2022 and title case is flat; both are house style. Check the inline-header bold list first, the one formatting tell that rose |
| An academic paper | **Tier 2** | section headings are journal template, not authorship. "2. Materials and Methods" is what the form requires |
| Plain prose: abstracts, email, commit bodies | **Tier 2** | only the em dash carries over from Tier 1. The rest needs markup to exist |
Figures, corpora and the before-and-after comparisons behind this table are in
`references/sources.md`.
Related but separat