Descriptions:
Dylan Davis, who runs an AI consultancy, addresses one of the most common and least visible failure modes when working with AI tools: context window degradation. When users upload large files to ChatGPT or Claude in the browser, the model may return a confident, well-formatted answer while having silently processed only part of the document. Davis presents a structured three-phase workflow using desktop coding agents — specifically OpenAI Codex and Anthropic’s Claude Cowork — to systematically work around this limitation.
The core mechanism is splitting the task across three separate conversations with fresh context windows. In the Map phase, the AI catalogs the structure of all files and records its findings in a notes file. In the Search phase, a new AI instance reads the map and extracts only the evidence relevant to the stated goal into a focused evidence file. In the Read phase, a final fresh AI instance reads only the pre-filtered evidence and produces the answer. Each handoff avoids the intelligence drop that occurs once context fills past roughly 50–60% capacity.
Davis also clarifies that context consumption varies significantly by file type — plain text is lightest, PDFs moderate, spreadsheets and PowerPoints heavier, and video files heaviest — and that the same model behaves differently depending on the platform. Claude Opus 4.7 supports 200,000 tokens inside Claude Cowork on the desktop but up to 1 million tokens when accessed through Claude Code. Specific copy-paste prompts are provided for all three phases, including instructions for how the AI should record its map, flag relevant sections, and carry forward notes between sessions.
📺 Source: Dylan Davis · Published May 16, 2026
🏷️ Format: Workflow Case Study
