Think Better

Think Better targets a very specific pain: your AI writes code fast but makes terrible architectural decisions. The fix is structural rather than cosmetic — it injects MECE, Issue Trees, and Pre-Mortems into the prompt so the model has constraints forcing it to reason like a principal engineer instead of an intern. Two things make that credible. First, the cognitive bias detectors (Sunk Cost, Confirmation, Status Quo) are pushed into the prompt before the AI answers, meaning it actively pushes back when you're wrong instead of agreeing. Second, the BM25 keyword engine over 160 embedded best-practice records across 9 domains picks a framework from your problem type — a query like 'debug latency' routes to an RCA framework, for example — rather than relying on a single canned template. The two interaction modes are well separated: natural language auto-detection for speed, slash commands for control. Depth levels span a 15-second Quick scan at 0.5x up to a 2.5x Executive briefing, and every session writes a step-by-step markdown workspace (problem definition, decomposition, prioritization, analysis plan, findings, synthesis, recommendation, bias warnings, decision log) that you can edit. The single 5MB Go 1.25 binary with zero dependencies and 100% local execution means no runtime to install and nothing sent to an external service during the analysis. The limits are the flip side of the design: it's a CLI, so you need terminal comfort, and it only works with Claude, Copilot, and Antigravity — not ChatGPT or Gemini. Framework discovery depends on the BM25 match to its embedded CSVs, so an unusual problem may not map cleanly. Where it fits: engineers, PMs, and leads making architecture choices, root-cause analysis, and pre-mortems. Where it doesn't: non-technical users, teams that need a GUI, or anyone whose stack runs on unsupported assistants.