Stanislav Kremeň on using Claude's Plan Mode to fix messy AI project specs
Stanislav Kremeň describes how shifting from writing a "perfect" upfront spec to using Claude Code's Plan Mode on an empty folder — and explicitly documenting non-goals — fixed a recurring pattern of mid-project rewrites.
Score breakdown
The post identifies a concrete workflow — using Plan Mode on an empty project combined with explicit non-goals stored in `CLAUDE.md` — that addresses the common problem of AI agents silently making structural decisions the developer never intended.
- 01Author Stanislav Kremeň identified the root cause of his project failures as a missing upfront plan, not bad AI-generated code.
- 02He recommends a minimal starting spec: user type, main job, 3–5 screens, data objects, failure cases, and non-goals.
- 03Explicit non-goals (e.g., 'no payments in v1') act as guardrails that prevent the agent from inventing out-of-scope features.
Stanislav Kremeň describes a recurring failure mode in his AI-assisted development: starting projects with no real preparation, then spending the rest of the project chasing the structure Claude had invented on its own. The root cause, he concludes, was not the AI's code quality but the absence of a guiding plan. His solution is a lightweight starting skeleton covering user type, the main job, three to five screens, data objects, failure cases, and — critically — what not to build yet.
The insight he found most surprising was the value of explicit non-goals.
The insight he found most surprising was the value of explicit non-goals. Lines like "no payments in v1" or "no social login yet" act as guardrails; without them, the agent rediscovers out-of-scope ideas and pursues them. He recommends keeping non-goals in the spec until the project ships and storing permanent rules in `CLAUDE.md` so they persist across every prompt.
The workflow he landed on uses Claude Code's Plan Mode (`Shift+Tab`) on an empty folder — a use case he had previously assumed was only for changes inside existing projects. In Plan Mode, Claude explores and proposes a structure without touching any files, letting the developer review and push back before a single line of code exists. He also cautions that free-form AI brainstorming produces "a huge document full of praise," and recommends giving the agent explicit rules to always ask for pros, cons, and alternatives rather than agreeing with everything.
Key facts
- 01Author Stanislav Kremeň identified the root cause of his project failures as a missing upfront plan, not bad AI-generated code.
- 02He recommends a minimal starting spec: user type, main job, 3–5 screens, data objects, failure cases, and non-goals.
- 03Explicit non-goals (e.g., 'no payments in v1') act as guardrails that prevent the agent from inventing out-of-scope features.
- 04Non-goals should be kept in the spec until the project ships and moved into CLAUDE.md to persist across prompts.
- 05Claude Code's Plan Mode (Shift+Tab) works on an empty folder, surfacing structural decisions before any code is written.
- 06Unconstrained AI brainstorming produces praise rather than useful critique; the fix is giving the agent rules to always ask for pros, cons, and alternatives.
Topics
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Jun 9, 2026 · 17:05 UTC. How this works →