| name | text-trainer |
| description | Analyze someone's writing samples and generate a reusable voice profile. Use when asked to train a voice, create a writing profile, analyze someone's writing style, or build a voice profile for any person. Works with any writing type: tweets, emails, memos, Slack messages, LinkedIn posts, customer service replies, etc. |
| metadata | {"version":"1.0.0"} |
Text Trainer
Analyze real writing samples and produce a structured voice profile that any skill can use to match that person's voice.
When to Use This Skill
ALWAYS read this skill and its references when the user asks to:
- "Train a voice profile for [name]"
- "Analyze [name]'s writing style"
- "Create a voice profile for [name]"
- "Build a writing profile"
- "Learn how [name] writes"
- "Train on these writing samples"
- "Here are some [tweets/emails/messages] from [name]"
- "Make a voice profile from this"
- "Update [name]'s voice profile"
- "Retrain [name]'s profile with new samples"
Before Starting
Step 0: Load GTM Context
If .agents/gtm-context.md exists, read it for context about the company and communication style. This helps inform voice analysis — understanding the business context improves profile quality.
Step 1: Identify Who
Determine whose voice you're training. If the user doesn't specify a name, ask:
"Whose voice am I training? I'll save the profile under their name."
Step 2: Collect Writing Samples
You need real writing samples. The more, the better — aim for 10+ samples across different contexts if possible.
How samples can be provided:
| Method | What to do |
|---|
| Pasted in chat | User pastes text directly. Ask them to tag each sample with its type (tweet, email, memo, etc.) |
| Single file | User points to a file containing samples. Read it. |
| Directory | User points to a folder. Read all files in it. |
| Mixed | Any combination of the above. |
If samples are untagged, look at the content and infer the type (tweet = short, punchy; email = has greeting/sign-off; memo = longer form, etc.). Ask the user to confirm your guesses if unsure.
Minimum viable input: 5 samples of any type. Fewer than 5 and the profile will be thin — warn the user but proceed anyway.
Ideal input: 15-30 samples across 2-3 different writing types (e.g., 10 tweets + 5 emails + 5 Slack messages). More variety = richer profile.
Step 3: Run the Analysis
Read references/analysis-framework.md and execute all 8 passes against the collected samples.
This is the engine. Do not skip passes or combine them — each pass catches different patterns.
Step 4: Assemble the Profile
Read references/profile-template.md for the output structure.
Fill in every section using findings from the 8 passes. Only include "Tone by Context" subsections for writing types that had samples (don't guess at types you haven't seen).
Step 5: Save the Profile
Save the completed profile to:
.agents/voice-profiles/{name}.md
Where {name} is lowercase, no spaces (use hyphens for multi-word names).
Examples:
.agents/voice-profiles/john.md.agents/voice-profiles/sarah-chen.md.agents/voice-profiles/alex.md
Step 6: Confirm to the User
After saving, tell the user:
"Voice profile for [Name] saved to .agents/voice-profiles/{name}.md. Any skill that uses voice profiles (like email-writer) will now automatically use this when writing as [Name]."
If multiple profiles exist, mention:
"You now have profiles for: [list names]. When writing, I'll ask who's sending if it's not clear from context."
Updating an Existing Profile
When the user wants to update a profile with new samples:
- Read the existing profile from
.agents/voice-profiles/{name}.md
- Collect the new samples
- Run the full 8-pass analysis on the NEW samples
- Merge findings: keep patterns that appear in both old and new, add new patterns, remove patterns contradicted by new evidence
- Save the updated profile (overwrite the old one)
- Tell the user what changed
Reference Documents
- analysis-framework.md — The 8-pass extraction process. This is the engine that powers voice analysis. ALWAYS READ.
- profile-template.md — The output format. Every profile follows this structure. ALWAYS READ.
Output Location
All profiles are saved to .agents/voice-profiles/. This is a shared directory — any skill can read from it. The email-writer skill automatically loads profiles from here.
Related Skills
- gtm-context — Foundation context that helps inform voice analysis
- email-writer — Uses voice profiles generated by this skill for email writing
