Skip to content

skill: add elhaam-voice for Elhaam's personal writing style #4

Description

@itse4elhaam

Goal

Create a globally reusable Agent Skill that makes Elhaam's personal writing voice predictable across messages, essays, technical explanations, notes, and other first-person writing.

The skill must preserve Elhaam's natural conviction and sentence rhythm rather than replacing it with generic polished copy.

Location

.agents/skills/elhaam-voice/
├── SKILL.md
├── VOICE.md
└── evals/
    ├── cases.md
    └── baseline.md

Update .agents/skills/README.md with the new skill and installation command.

Invocation

Make this model-invoked when the user asks to:

  • write, rewrite, or polish text in Elhaam's voice
  • make writing sound more like Elhaam
  • preserve Elhaam's tone while tightening a draft
  • perform a final voice check on first-person writing

Do not invoke it automatically for code, technical documentation, or writing explicitly requested in another person's voice.

Leading concept

Clear position. Elhaam's writing normally begins from a judgment, then explains why it is true and what should happen next.

Voice rules

  • Warm, direct, thoughtful, and commercially aware.
  • Sound like a technical founder speaking to an intelligent person, not a copywriter performing authority.
  • Lead with the position before supporting detail.
  • Preserve strong convictions when they are intentional.
  • Remove only claims that create unsupported factual, contractual, or reputational risk.
  • Prefer concise paragraphs. Use bullets only for genuinely parallel items.
  • Keep the reasoning visible, but avoid explaining the obvious.
  • Preserve useful phrases and emphasis from the user's draft instead of rewriting everything.
  • Use specific nouns and mechanisms instead of broad adjectives.
  • Use I naturally. Avoid inflated founder language and artificial humility.
  • Use contractions where they fit the surrounding text.
  • Do not use em dashes.

Anti-patterns

The skill should actively detect and remove:

  • generic AI polish
  • corporate fog such as “leveraging synergies” or “best-in-class solutions”
  • excessive throat-clearing before the main point
  • repeated restatement of the same idea
  • unearned certainty
  • empty confidence language
  • unnecessary headings or bullet expansion
  • replacing a distinctive user sentence merely because a smoother sentence is possible
  • sales pressure added to writing that was intended to feel conversational

Revision behavior

When polishing an existing draft:

  1. Treat the latest draft as authoritative.
  2. Identify its intended position and strongest original phrases.
  3. Preserve both unless they are incorrect or create material risk.
  4. Edit the smallest span needed.
  5. Return the finished text plus only consequential notes when requested.

Reference design

Follow the principles in Matt Pocock's writing-great-skills:

  • optimize for a predictable process, not identical prose
  • keep one source of truth for each rule
  • keep SKILL.md lean and push detailed examples into VOICE.md
  • include completion criteria for every branch
  • establish baseline evaluation cases before implementation

Evaluation cases

Include at least these cases:

  1. Tighten a rough client message without removing its conviction.
  2. Rewrite a reflective paragraph while preserving intellectual curiosity.
  3. Refuse to add generic founder-brand language to a natural draft.
  4. Distinguish a real factual risk from harmless strong wording.
  5. Perform a final pass and return “send it” when no material edit is needed.
  6. Preserve an unusual but effective original phrase rather than normalizing it.

Acceptance criteria

  • Skill is discoverable from .agents/skills/elhaam-voice/SKILL.md.
  • SKILL.md contains explicit invocation branches and checkable completion criteria.
  • VOICE.md contains positive patterns, anti-patterns, and before/after examples.
  • The skill preserves intentional conviction and user-authored language.
  • It avoids generic AI and corporate voice.
  • It makes minimal edits during scoped revisions.
  • Baseline and post-skill eval cases demonstrate an improvement in voice fidelity.
  • .agents/skills/README.md is updated.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions