Voice

"Be conversational" isn't a voice guide. This track teaches you to document voice precisely enough that AI can replicate it—and validate when it drifts.

5 levels 6-10 weeks artifact: Complete voice system with validation

Demonstrated, not asserted

  • Signal Dispatch voice guide read The shipped voice spec this track teaches you to write — voice documented precisely enough that an agent replicates and validates it. Read it and copy the pattern. verified 2026-07-15
  • claude-recall-cli clone The corpus + recall tooling that makes the voice measurable and driftable — it mines transcripts into a voice stack. Now public; clone it. verified 2026-07-15

Who this is for

This path is for you if

  • ·You produce content regularly and want consistent voice
  • ·You want AI to sound like you, not like generic AI
  • ·You're a blogger, writer, or thought leader with an audience
  • ·You write for different contexts (blog vs docs vs email)

Probably not if

  • ·You're just starting to explore AI (see the Explorer track)
  • ·You don't have existing writing to analyze
  • ·You need architecture docs, not content (see the Architect track)
  • ·You want generic AI output, not your voice specifically

Why Voice Guides Fail

Most voice guides are too vague to be useful. Vague and useless: "Be conversational and approachable." Specific and actionable: "Start with a question. Second-person you. No sentences over 20 words. Contractions always." A good voice guide is specific enough that you could check compliance automatically.

The levels

0 Analyze Your Own Writing 1-2 days

Find the patterns you already use—before trying to codify them.

Analysis method

  1. Gather 10-20 pieces of your writing (posts, emails, docs)
  2. Read them looking for patterns, not quality
  3. Note: sentence length, opening patterns, word choices, tone shifts
  4. Look for what you do consistently (intentional or not)

Patterns to look for

  • Sentence Rhythm — Short punchy? Long flowing? Mixed?
  • Opening Hooks — Questions? Bold claims? Stories? Data?
  • Tone Markers — Formal/casual? Confident/tentative? Direct/exploratory?
  • Structural Habits — Lists? Paragraphs? Headers? Mixed?
  • Signature Phrases — Words or phrases you use repeatedly

Checkpoints

  • ·List of 5+ patterns you noticed in your writing
  • ·Examples of each pattern from your actual work
  • ·Notes on which patterns feel intentional vs accidental

Done when — You can identify 5 distinct patterns in your existing writing

1 Write Your First Voice Guide 1-2 weeks

Document your voice so others (or AI) can replicate it.

Voice guide sections

  • Core Identity — Who is speaking? What do they believe? What makes them distinctive?
  • Tone Spectrum — Where on the formal/casual, confident/humble, direct/exploratory axes?
  • Structural Patterns — How are ideas organized? What formats are preferred?
  • Language Choices — Words to use, words to avoid, signature phrases
  • Examples — Before/after showing voice applied

Requirements

  • ·Be specific—"conversational" is too vague
  • ·Include anti-patterns (what NOT to do)
  • ·Show don't just tell (examples > descriptions)

The worked example

  • ·The published Signal Dispatch voice guide is exactly this artifact — a voice spec precise enough that an agent can replicate and validate it. It is the voice that runs this blog. Read it and copy the pattern.

Checkpoints

  • ·3,000+ word voice guide with all 5 sections
  • ·At least 3 before/after examples
  • ·Someone else can read it and understand your voice

Done when — You have a 3,000+ word voice specification

2 Create Multiple Voice Profiles 2-3 weeks

Define distinct voices for different contexts or audiences.

Voice contexts

  • Audience — Technical readers vs executives vs general public
  • Purpose — Teaching vs persuading vs documenting
  • Medium — Blog posts vs documentation vs social media
  • Domain — Thought leadership vs architecture (inverted voice)

Technique

  1. Start with your base voice
  2. Identify what changes for each context
  3. Document the delta, not a complete rewrite
  4. Create switching rules (when to use which voice)

Checkpoints

  • ·2 distinct voice profiles documented
  • ·Clear switching rules between them
  • ·Examples of same content in both voices
  • ·Someone else can identify which voice a piece uses

Done when — You have 2 documented voice profiles that are clearly different

3 Build Automated Validation 2-3 weeks

Create checks that catch voice drift without manual review.

Validation approaches

  • Word Lists — Flag banned words, suggest alternatives
  • Pattern Matching — Check for required/forbidden sentence structures
  • LLM Scoring — Prompt an LLM to score adherence to voice guide
  • Comparative Analysis — Compare new content to a corpus of approved content

Implementation

  1. Start simple—word list checks work for most cases
  2. Add LLM scoring for nuanced violations
  3. Build a feedback loop (violations → guide updates)
  4. Track metrics (violations over time, common issues)

The real system

  • ·Comparative analysis against a corpus is exactly what claude-recall-cli does: it mines your Claude Code transcripts into a searchable corpus and a voice stack you can score new writing against. It is the machinery behind this blog's voice, and it is public — clone it.

Checkpoints

  • ·Working validation tool (any approach)
  • ·Can catch at least 3 types of violations
  • ·Feedback mechanism to improve the guide
  • ·Documented false positive rate

Done when — You have automated voice validation that catches violations

4 Orchestrate Multi-Voice Systems 3+ weeks

Manage multiple voices across teams or content types at scale.

System components

  • Voice Registry — Catalog of all voices with metadata (owner, use cases, last updated)
  • Selection Logic — Rules for which voice applies to which content
  • Validation Pipeline — Automated checks run on all content before publish
  • Drift Detection — Monitor for voice inconsistency over time
  • Governance — Who can create/modify voices, approval workflows

Metrics to track

  • ·Voice consistency score across content
  • ·Time to validate new content
  • ·False positive/negative rates
  • ·Voice drift over time

The actual system behind this blog

  • ·This is not hypothetical. The Signal Dispatch voice guide plus claude-recall-cli are the running version: a documented voice spec, and the recall machinery that keeps it measurable and flags drift. The pair is exactly what this track walks you toward building for your own voice.

Checkpoints

  • ·Voice registry with 3+ documented voices
  • ·Automated selection and validation
  • ·Governance process documented
  • ·System works without you for 2+ weeks

Done when — You have a voice system that works without your direct involvement

Start at Level 0 Read the voice guide

Tool links and level references live in a dated data file checked by the link sensor on every build — this page can't silently point at a moved doc or a dead demo.