Fabrizio Ferri-Benedetti's Context Curator Approach

Table of content

Fabrizio Ferri-Benedetti is a Principal Technical Writer at Elastic and co-maintainer of OpenTelemetry docs. Based in Barcelona, he writes about the intersection of technical writing and AI at passo.uno. His approach: tech writers should build their own LLM tools rather than depend on generic solutions.

Fabrizio created Impersonaid for testing documentation against LLM-powered user personas, Valegen for generating Vale linter rules with RAG, and various MCP servers for documentation workflows.

Background

Career path: Tech journalist and content strategist since 2008, then API designer and technical writer
Current role: Principal Technical Writer at Elastic
Previous: Splunk, New Relic
Training: Cognitive psychology
Languages: English, Spanish, Italian, Catalan
Blog: passo.uno since 2023
GitHub: @theletterf

He describes himself as knowing “enough coding to be dangerous” and creates custom tools when existing solutions fall short.

The Context Curator Philosophy

Fabrizio argues that technical writers should become “context curators” for AI systems. The core insight: developers increasingly rely on AI tools that consume documentation as context. Well-structured docs directly improve AI performance.

From his blog post:

“All AI requires to do the right thing is great context and a gentle nudge.”

This shifts the writer’s role from presentation-focused content to designing context-rich information structures. Specific recommendations:

Technique	Purpose
CLAUDE.md files	Project guidelines and commands for AI context
Context folders	Organized directories for AI consumption
Semantic tagging	Information architecture for machine parsing
llms.txt files	LLM-optimized content delivery
DITA markup	Standardized semantic formats

Impersonaid: Documentation User Testing

Impersonaid simulates user interactions with documentation using LLM-powered personas. Instead of traditional user research, it provides quick feedback loops during development.

Install and run:

git clone https://github.com/theletterf/impersonaid.git
cd impersonaid
npm install && npm link
impersonaid create-sample --name beginner_developer
impersonaid simulate --persona beginner_developer \
  --doc "https://your-docs.com/quickstart" \
  --request "How do I authenticate?"

Personas include attributes like:

Expertise level (beginner to expert)
Patience and attention to detail
Communication style preferences
Learning patterns

Supports OpenAI, Claude, Gemini, OpenRouter, and Ollama. Results save as Markdown with the persona profile, request, and simulated response.

Valegen: RAG-Powered Linter Rules

Valegen generates Vale linter rules from natural language descriptions. Rather than writing YAML configurations manually, describe what you want in plain English.

The tool uses retrieval-augmented generation:

Search a vector database of Vale documentation and style guides
Retrieve relevant examples from Vale core, Google, and Microsoft styles
Ground the LLM response in actual syntax patterns
Generate three rule approaches with confidence ratings

git clone https://github.com/theletterf/valegen.git
cd valegen
python -m venv venv && source venv/bin/activate
pip install -r requirements.txt
python setup.py  # Builds vector database
flask run --port 5001

Supports Gemini, GPT, and Claude. The RAG approach prevents hallucinated YAML that looks plausible but fails.

Aikidocs: Own Your Prompts

For custom documentation workflows, Fabrizio built Aikidocs, a NodeJS tool that processes local files as context for LLM requests.

His philosophy: “Tech writers must own the prompt.”

The workflow:

Initialize a git repository (critical for rollback when LLMs stray)
Define folder structure: context, prompts, credentials, output
Configure LLM provider and API key
Run the script to process context

Key insight: use iterative prompting rather than comprehensive requests. Build up capabilities through sequential refinements.

Daily AI Uses

Fabrizio uses Claude and Copilot for professional work. Specific applications:

Task	Approach
Markdown fixing	Tables and links faster than regex
Code boilerplate	React components, UI widgets
Code explanations	Understanding unfamiliar functions
Writer’s block	Request closing paragraphs for blog posts
Style matching	Rewrite content to match voice
User testing	Computer Use for automated screenshots
Test generation	Selenium code for documentation workflows

From his AI usage post:

“LLMs augment existing skills by removing tedious work at margins, not replacing core work requiring judgment and creativity.”

Other Projects

His GitHub includes:

Project	Description	Stars
english-lang	“The English Programming Language”	505
otel-sonifier	Calm monitoring for OpenTelemetry Collector	13
Vale-MCP	MCP server for Vale linting	16

He also contributed to making Splunk’s Vale style guide open source.

Key Takeaways

Principle	Implementation
Build your own tools	Custom scripts beat generic solutions
Context over prompts	Structure docs for AI consumption
Test with personas	Simulate users before publishing
Use RAG for rules	Ground generation in real examples
Git everything	Version control enables rollback

Links

Next: Jesse Vincent’s Superpowers Framework