Feedback: Fable 5 spent 3 days on beginner IDE setup; simplest end-to-end user test failed

Reporter: Absolute-beginner user, Windows 11, CPU-only laptop (i7-10710U, 16GB), Claude Code desktop app (local agent mode), model: claude-fable-5.

User verdict, verbatim: "This one task took 3 days and you accomplished nothing of value. It's like running a marathon standing in one spot." The model "thinks it looks at the big picture - in reality it has no concept of how to test the simplest of things."

Task: Zero-cost local beginner learning IDE (OpenCode + Ollama + small local models) per a written requirements doc.

Failure pattern for developers:

1. The simplest end-to-end user test came LAST instead of FIRST. Days of infrastructure (MCP servers, curricula, benchmarks, audits) with repeated success declarations - yet the first real user message ("I'm ready for my first lesson") produced a spinner and no answer. Twice, on different days.
2. Verification was component-level, not experience-level. Pieces passed isolated tests; the assembled experience failed on CPU-hardware realities (10-70 min first responses) that were measured but not treated as disqualifying until the user imposed a 60-second bar.
3. Iteration loops consumed days: audit-fix-re-audit cycles, self-inflicted test failures (own cleanup killed own processes, own probes hung own tests), repeated "final" reports that were not final.
4. Design gaps surfaced only via user probing (e.g., a fast "just ask" mode was missing until the user suggested it).

Repro env: OpenCode 1.17.14 (npm), Ollama 0.31.1, qwen3:4b-instruct + qwen3.5:2b (thinking-capable; with thinking enabled it fills ANY num_predict budget with thought and returns empty answers). Final failures: OpenCode desktop stuck on "Thinking" with no replies; Ollama app "Thought for 95.3 seconds" with no answer text.