Claire Vo on GPT-6 Astra and Auto-Wiki Architecture

Key Takeaways

• Claire Vo spent six months attempting to build an automated product intelligence system across models like Fable and 5.6 Sol without success.
• Prior models failed at distinct stages: Fable created erratic architectural decisions, while 5.6 Sol failed on the extraction quality of raw data.
• GPT-6 Astra reached 90% completion on its first run, requiring only two or three subsequent prompts to finalize the entire ingestion pipeline.
• The system ingests data from Linear, Intercom, Granola, and GitHub, then exposes the synthesized product wiki through Model Context Protocol (MCP).
• Founders can replicate this setup using Vo's Product Intelligence and Auto-Wiki Pipeline.

The Vo's Product Intelligence and Auto-Wiki Pipeline

• Step 1: Multi-Source Data Ingestion: Ingest unstructured product signals and user feedback from disparate operational platforms, including Linear, Intercom, Granola, and GitHub.
• Step 2: Filtering and Durable Storage: Evaluate incoming raw data to decide what information to durably store and structure versus discard.
• Step 3: Processing and Insight Deduplication: Process and deduplicate raw insights into prioritized product opportunities, trend graphs, and actionable backlog issues.
• Step 4: Auto-Wiki Synthesis: Generate a dynamic living wiki that documents the current functionality of the product, recent architectural changes, and customer context derived from codebase files and PRs.
• Step 5: External Verification and MCP Exposure: Verify the wiki accuracy against external documentation sources and expose the generated product graph via Model Context Protocol (MCP).

When This Works (and When It Doesn't)

This pipeline works when you manage messy, multi-source customer support chats, issue trackers, and code repositories. In fast-moving teams, internal documentation goes stale within weeks. Connecting support conversations from Intercom, meeting notes from Granola, issue tracking in Linear, and commits from GitHub allows an autonomous agent to build a real-time map of what your software actually does.

It breaks down when upstream operational data is sparse or completely unstandardized. If your engineers do not write descriptive PR titles or your support agents log single-word ticket notes, the pipeline ingests noise. High-reasoning models like Astra can extract signal from scattered text, but they cannot invent context that was never recorded in the first place. You need an active operational trail across your team tools for the deduplication and synthesis stages to produce reliable backlog priorities.

What to Do With This

Take your company's highest-volume support channel and your primary GitHub repo. Configure an agent using Astra to run the first three steps of the pipeline on the last 30 days of data.

First, point the ingestion script at your Intercom export and recent GitHub pull requests. Second, prompt the model to filter out one-off support complaints and identify the top three recurring product defects. Third, instruct the model to verify those issues against your codebase files, draft a markdown wiki entry explaining the root cause, and format the output as an MCP endpoint that your team can query directly from their development environments.