Repository navigation
Stop over-editing informal human text - #318
Conversation
|
Thanks, the human-control results make the case. Two things before I merge. |
ddc830d to
5f9dbeb
Compare
|
Done, all three:
|
|
Rebased on v4.0.0. SKILL.md is now +131 words, with the study cited once in Source (30 LLMs, 'English' in the title). The version bump is dropped, and the changelog entry is under Unreleased. |
Chat messages, rants, and angry reviews were rewritten into polished prose: lowercase and "lmk" formalised, fragments merged, "Worst. Service. Ever." softened, "tried X" turned into "did X". Add a rule to Voice: leave informal text alone unless two unrelated tells appear together (chatbot residue excepted), keep the writer's anger at full strength, and return the text unchanged if no tell is found. Also add a restating "Overall"/"Ultimately" verdict to the closers in pattern 2, note that a plain "let me know" is not chatbot residue, and cite El Attar et al. (arXiv:2606.04177) once, in Source, which found that most linguistic signals of AI text change with the model and the text type.
5f9dbeb to
3e5f816
Compare
Problem
Humanizer rewrites informal human text as if it were AI text. In isolated runs, a Slack message went from lowercase with "lmk" to formal sentences, "Worst. Service. Ever." became "Terrible service.", "Avoid this place" became "I'd skip this place", and "Tried restarting it" became "I restarted it".
Change
SKILL.md(Source) and once in the README sources. It found that lexical richness held across models and domains and that most other signals changed with them. That is the basis for the informal-text caution.SKILL.mdgrows by 131 words over main (5,226 to 5,357 bylen(SKILL.split()), cap 5,500). The citation moved out of the Why section. No version bump: the manifests andSKILL.mdstay at 4.0.0, and the CHANGELOG entry is under "Unreleased". Rebased on main (v4.0.0).Testing
Isolated
claude -pruns (no hooks, tools, or other skills),SKILL.mdas the system prompt, v3.1.0 vs the earlier version of this branch, 10 runs per text. Checks were written before the runs. Eighteen short texts: 8 human informal controls (including 4 held-out ones added after the first results) and 10 AI-written texts.python scripts/validate-package.pypasses on the rebased branch (package v4.0.0).