Video summary

How to Make AI Write in YOUR Voice (Claude Skill Tutorial)

Main summary

Key takeaways

Technology

Summary (technological concepts + tutorial/product features)

  • Claude Co-Work / Claude Skills workflow: The tutorial shows how to build a Claude Skill inside Claude co-work that enables Claude to write in a specific author’s personal voice, rather than producing generic text.

  • Why generic AI output happens: The speaker argues that when Claude “doesn’t know your voice,” it guesses—effectively averaging from prior exposure—resulting in bland “everybody” writing. Fix: create a voice-specific skill.

  • Key idea: voice = writing mechanics (not genre/theme): In this approach, “voice” refers to the craft patterns that shape prose, such as:

    • Sentence rhythm and length
    • Word choice
    • POV handling
    • How dialogue vs. description is structured
    • The use of silence/implication and other non-obvious stylistic habits

  • Voice Skill requires “examples,” not description: You can’t just describe your voice in instructions. Instead, you provide real prose samples—ideally fiction—that reflect your actual published/edited work. Recommended scale and variety:

    • 2,000+ words (ideal 3,000–5,000)
    • Include different scene types (e.g., action, quiet scenes, heavy dialogue)

  • “Extraction step” (main technical method):

    1. Start with an extraction prompt in a fresh Claude chat.
    2. Claude acts like a forensic editor / writing analyst to generate a structured “voice fingerprint” from your sample prose.
    3. The fingerprint captures patterns including:
      • sentence rhythm/length
      • vocabulary and description style
      • dialogue handling
      • POV style
      • naming patterns
      • other compositional habits

  • Building the actual Skill from the fingerprint:

    • After Claude generates the voice fingerprint, you ask it to write a “voice skill file” using the normal Claude skill structure (labels, rules, examples).
    • The rules become voice-specific, aimed at preventing generic writing, using constraints such as:
      • avoid unnecessary adverbs
      • avoid naming emotions directly
      • control subordinate clause usage
      • ensure description does multi-purpose work

  • Skill creation vs. prompt doc:

    • The speaker emphasizes this should be a formal Claude Skill created for Claude co-work/Claude Code using the skill creator, not merely a pasted prompt document.

  • Automated evaluation/testing inside the workflow:

    • After generating the skill, Claude runs a qualitative test suite, including:
      • three realistic prompts (cold open, tense two-character scene, solo action sequence)
      • comparisons between agents with the skill vs. baseline (no skill)
    • It produces evaluation artifacts (e.g., grading files/metadata) and summarizes results.

  • Demonstrated comparison (skill vs. no skill):

    • With the voice skill: more distinctive sensory details and sharper mechanical cadence
    • Without the skill: outputs labeled as “generic,” with more boilerplate phrasing and less distinctive voice

  • Multiple voices / multiple skills concept:

    • You can build multiple voice skills for different pen names or sub-styles (e.g., space opera vs. cozy mystery), each trained on different sample sets.

Key guide/tutorial deliverables

  • A ready-to-use “extraction prompt” to generate a voice fingerprint from your prose (the video notes it will be linked in the description).
  • A procedure to:
    1. Gather 2,000–5,000+ words of fiction prose across scene types
    2. Run the extraction step in Claude
    3. Feed the fingerprint into the skill creator to generate a Claude Skill
    4. Validate via prompt runs and skill vs. baseline comparisons

Main speakers/sources

  • Akello Herod (main presenter/instructor)

Original video