Creator of RepoPrompt
Pioneer in AI-assisted development workflows. Creator of RepoPrompt for context management and multi-model consultation (MCP). His hybrid model approach has become the gold standard for serious AI-powered development.
"Manual context curation beats auto-context. Hybrid setups (Opus + GPT-5.1) outperform single-model approaches. Medium reasoning balances speed and accuracy—avoid high reasoning under context pressure."
The deep thinker. Unmatched for planning and architecture on complex tasks. Use vanilla (non-tool-calling) mode for pure reasoning after building context.
The reliable executor. Impressive for implementation, code review, and translating user rambles into coherent prompts. Efficient "light reasoning" model that gets things done.
The validator. Ideal for second opinions and quick unblocking during iterations. Great bird's-eye view of context via RepoPrompt MCP chat.
The surgeon. Specifically endorsed for context building in large codebases. More surgical than Opus at selecting relevant files without overthinking.
The dark horse. First non-Western model in Repo Bench top 10. Scores on par with Opus 4.5 (non-thinking mode). Worth considering for coding tasks.
Mixed results. Worse at large-context reasoning than Gemini 2.5 Pro. Google fighting on both light and deep reasoning fronts—expect improvements.
Essential stack. Manual context curation avoids "junk" in prompts. MCP enables multi-model consultation—consult GPT-5.1 while Opus executes.
The optimal setup: GPT-5.1 Pro for planning → Opus 4.5 for implementation → GPT-5.1 High for review/validation. Iterate with Opus, consult 5.1 when stuck.
Medium reasoning levels balance speed and accuracy. High reasoning can pollute context with traces. Avoid over-relying on high reasoning or small models under context pressure.
Curate context manually to avoid "junk" in prompts. Auto-context can include irrelevant files. Surgical file selection beats dumping everything.