What problem does it solve?
It prevents an AI from running ahead, making bundled assumptions, or “doing extra” by enforcing a strict one-change-per-turn workflow that keeps the user firmly in control of architecture and decisions.
Core Features & Use Cases
- One-change-per-turn build loop: Each turn produces the smallest meaningful diff, runs targeted tooling, and asks exactly one next-step question.
- Approval-gated spec and surface mapping: Captures the user’s intent, infers the public interface (endpoints/modules/data shapes/commands), and only then moves to implementation.
- Tooling-verification discipline: Runs LSP/formatters/tests scoped to touched files, distinguishes expected vs unexpected errors (e.g., stubs/forward-refs), and keeps unresolved imports from slipping through unintentionally.
- External API call safety checks: Flags unverified library/API usage when docs access is unavailable or mismatch is detected when it is available.
- Deferred-gap tracking: If the user defers decisions, the skill re-surfaces them at the moment they become load-bearing.
Quick Start
Activate handcraft mode by telling the agent to build your feature function by function, approving each step before it proceeds.