By name: content-writer description: Generate content in each team member's voice with adaptive voice learning via WhatsApp. version: 1.2.0 author: GroundUp Toolkit actions: - generate - test
Content Writer
WhatsApp-triggered content generation that writes in each team member's authentic voice. Supports LinkedIn posts, Substack notes, LinkedIn messages, and newsletters. Uses per-member voice profiles, audience data, and brand context combined with web research and a humanizer pass.
Voice Learning: The system gets smarter over time. Submit writing samples and accept generated content to build an adaptive voice profile that improves with every interaction.
Content Types
LinkedIn Post
Short-form content (150-300 words). Direct, insight-driven, observation-first. Trigger: "write a post about...", "linkedin post about..."
Substack Note
Ultra-short content (1-10 sentences). Punchy, quotable, single-insight format. Trigger: "write a note about...", "substack note about..."
LinkedIn Message
Direct message (2-5 sentences). Warm but direct outreach, no generic templates. Trigger: "write a message to...", "linkedin message to...", "reach out to..."
Newsletter / Article
Long-form thought leadership (800-1500 words). Subject line options, sectioned headers, skim-optimized. Trigger: "write a newsletter about...", "write an article about..." Delivered via WhatsApp preview + full version by email.
Hebrew messages produce Hebrew output with English tech terms.
Voice Learning
The system learns your writing style over time through two mechanisms:
Submit Writing Samples
Send examples of your existing writing to teach the system your voice:
voice sample: [paste your post/article/message here]
Send at least 2-3 samples for the system to start learning patterns.
Accept Generated Content
After receiving generated content, reply:
keep
This saves the content as an accepted voice sample.
Check Your Voice Profile
my voice
Shows how many samples you have and when they were last analyzed.
How It Works
- Samples stored in
profiles/<name>/samples.json(auto-created) - When 2+ samples exist, Claude Haiku extracts a style fingerprint
- The fingerprint + up to 2 raw examples are included in every generation prompt
- Max 20 samples per member (oldest pruned, accepted content prioritized)
Profile Structure
Each team member needs a directory under profiles/<first_name>/ with:
voice.json— Voice DNA (tone, style, phrases, boundaries)audience.json— Target audience (pain points, language, aspirations)brand.json— Business context (positioning, differentiators, offerings)samples.json— Auto-created when voice samples are submitted
See profiles/example/ for template files.
Actions
generate
content-writer generate "<message>" "<sender-phone>"
test
content-writer test
Configuration
Required environment variables:
ANTHROPIC_API_KEY— Claude API for generation + humanizer + voice analysisBRAVE_SEARCH_API_KEY— Brave Search for topic research (optional but recommended)