| name | wpt-gen-llm |
| description | Best practices for configuring LLM integrations, concurrent networking, context scraping, and managing prompts in WPT-Gen. |
WPT-Gen LLM Skills
This document outlines the best practices for LLM integrations, prompt pipeline safety, and context extraction within the wpt-gen repository.
1. Multi-Provider Support
WPT-Gen supports Google Gemini, OpenAI, and Anthropic models via a unified abstraction.
- Google GenAI (
google-genai): Used primarily for deep context reasoning (e.g., gemini-3.1-pro-preview) and fast generation (gemini-3-flash-preview). API keys are read from GEMINI_API_KEY.
- OpenAI (
openai): Used as an alternative provider. API keys are read from OPENAI_API_KEY.
- Anthropic (
anthropic): Used as an alternative provider (Claude models). API keys are read from ANTHROPIC_API_KEY.
2. Phase-Based Model Mapping
The agentic workflow uses different model categories based on the complexity of the task, as configured in wpt-gen.yml.
- Reasoning Models: Used for complex analysis and extraction.
- Requirements Extraction: Identifies normative requirements from spec text.
- Coverage Audit: Performs deep gap analysis against existing tests.
- Lightweight Models: Used for rapid generation and iterative refinement.
- Test Generation: Produces test code based on audit blueprints.
- Evaluation: Performs self-correction and validation of generated code.
3. Network Architecture (Context Extraction)
Providing context is critical for minimizing hallucinations, but it involves extensive network I/O blockades.
- Concurrent I/O: Always use
asyncio.gather combined with asyncio.to_thread for concurrent network requests (e.g. fetching 5 MDN explainer URLs at once). Blocking sequential for loops across network dependencies creates catastrophic UI lag.
- Missing Failsafes: Reviewers must actively flag blocking network requests that lack exponential retry logic/backoff for
HTTPError 429 (Too Many Requests). Do not depend on LLMs blindly surviving a dropped HTTP request.
- Trafilatura: WPT-Gen uses
trafilatura to extract dense text from W3C Specification URLs linked to web features.
4. Prompt Management & Defensive Context
Prompt structure determines output quality, but unbounded inputs determine runtime OOM exceptions.
- Defensive Paging (OOM/Token Limits): Native memory limits and LLM token limits require defensive programming. If ingesting massive terminal runner logs into the context window, slice the array to only keep the tail end to remain token-efficient. If querying expansive file directory contents, return or exit early when numerical limits are hit rather than crashing.
- Context Bloat: Ensure large, optional prompt dependencies (like MDN explainers) are safely guarded behind Jinja
{% if %} statements to avoid feeding the LLM empty strings when variables map to null.
- Clear Instructions: Ensure system prompts clearly dictate the agent's persona (expert test engineer) and the expected format.
- XML Output: When requesting structured data (like gap analysis or test blueprints), explicitly request XML format and provide the intended schema structure. This allows WPT-Gen to parse and programmatically act on the AI output.
