Doubt-Driven Development gives an agent an in-flight discipline for testing its own decisions. A decision counts as non-trivial when it adds branching logic, crosses a boundary, asserts a property the compiler cannot verify, depends on hidden context or cannot be undone.
Each cycle writes the claim, extracts the artifact and its contract, sends only those to a reviewer in fresh context with an adversarial prompt, classifies every finding and stops after trivial findings, three cycles or a user override. It also offers a cross-model second opinion with explicit user authorization.
What You Get
- A five-step doubt cycle with a checklist
- An adversarial review prompt to paste into the reviewer
- Rules for optional cross-model review through Gemini or Codex CLI
- Stop conditions, red flags such as doubt theater, and a verification list
Common Use Cases
- Reviewing high-stakes production changes
- Stress-testing a plan before building
- Verifying non-obvious claims in unfamiliar code
- Getting a second model’s view on a design
What Makes It Unique
The skill separates in-flight doubt from the final gate of a review command, withholds the author’s conclusion from the reviewer to avoid agreement bias, and names a checkable signal for doubt theater. Its cross-model step requires a read-only sandbox, stdin prompts and fresh authorization for every external CLI run.
Under The Hood
The skill is a single SKILL.md of instructions, meant for the main session so it can start a fresh-context reviewer. With the user’s approval it can run which, gemini or codex commands, passing a prompt file through stdin in a read-only mode, and it writes that prompt to a temporary file. It ships no scripts.