[scarletkc]

Motivation

Tweets written with talk-like-scarletkc plus x-content had become flat and formulaic. The causes:

• Duplicated and conflicting rules. Many rules were stated three to six times across files: the banned-opener and both-sides lists, the defensive tail clause, analogies, and edit scope. Several rules also contradicted each other. For example, the core voice allowed contrast while anti-patterns discouraged it. Writing one tweet loaded about 60 KB across 7 files, mostly prohibitions.
• Faithful-only drafting. Both skills told the model to transcribe the supplied material faithfully and add nothing. Drafts therefore restated the material instead of adding an angle or stance.
• Examples copied as templates. Sample tweets were reused as patterns, so posts converged on the same openings, phrases, and structure.
• No engagement guidance in x-content. It only covered where readers get confused, not why they would stop scrolling.

Changes

• talk-like-scarletkc
  • Merge voice-profile.md and anti-patterns.md into SKILL.md so each rule appears once.
  • Split behavior by task. Voice rewrites (polish, rewrite, translation, local edits) stay faithful. Drafts may add angle, stance, reactions, humor, analogies, and rhetorical questions.
  • Both modes still forbid inventing verifiable facts. When the assistant chooses a stance, it says so outside the draft instead of stopping to ask.
  • Write tweets from rules only. Delete examples.md and fold its chat habits into surface-profiles.md as rules.
  • Keep dialogue-samples.md, which is read only when ghostwriting one-on-one chat.
  • Add tweet rules distilled from the author's higher-engagement posts. The posts themselves are not stored. The rules: lead with a verdict or reaction, state emotion fully, voice speculation as the author's own, and allow an ending that adds a new judgment.
  • Add a no-repetition rule: vary openings, structure, endings, and catchphrases across posts.
• x-content
  • Add "why readers stop" and first-line guidance.
  • Allow drafts to add an angle and stance.
  • Add quote-post structure and the no-template rule.
  • Remove the voice rules it duplicated from talk-like-scarletkc.
• lint_style.py: drop the negative-framing and vague-experience-word rules. They contradicted the guidance to keep reasons and concrete experience words.
• telegram_corpus.py: the candidate checklist now points to distilling rules into surface-profiles.md instead of examples.md.
• Evals
  • Update talk-like-scarletkc 6 and 14, and x-content 10, for the new drafting behavior. Cases that previously required asking for a missing stance now expect a draft that says the stance was chosen by the assistant.
  • Add talk-like-scarletkc 43: a real contrast and punchline survive a voice edit.
  • Add x-content 27: author attitude leads instead of a changelog.
  • Add x-content 28: a draft adds an angle to flat material without inventing facts.

Compatibility: this removes the reference files references/voice-profile.md, references/anti-patterns.md, and references/examples.md. Local copies or external links that point to them will break. There are no runtime code, public API, or eval JSON schema changes.

Validation

• python -m unittest discover -s tests (Python 3.13.14): 99 tests passed.
• python scripts/update_readme_skills.py --check: passed.
• python scripts/package_skill.py talk-like-scarletkc and x-content: both packaged.
• gh skill publish --dry-run: passed with advisory repository-security warnings.
• git diff --check origin/main...HEAD: passed.
• Informal behavioral comparison: subagents wrote the same three fictional posts plus one voice-only edit once per skill version.
  • Original: drafts restated the material.
  • Pruning only: nearly identical output.
  • Allowing additions plus example tweets: stronger hooks, but every post followed the same verdict, facts, punchline shape.
  • Final rules-only version: varied structure, and the voice-only edit stayed faithful. The tweet path now reads 3 files, about 19 KB.

This is a small, single-run sample. It shows a direction but does not establish better engagement. The model-driven eval suite was not run.

---