The era of AI-assisted coding is hitting a friction point with silent assumptions. A new pair of npx tools, kokoro-baton and kokoro-aegis, attempts to solve the problem of AI agents making critical product decisions without user input. The workflow forces developers to explicitly decide on edge cases, data persistence, and UI behaviors before Claude Code writes logic, then verifies those decisions post-build without using another AI model.

The Problem With Silent Guesses

When Claude Code builds an application, it routinely makes choices that were never specified. The source article lists common questions left unanswered: does the data survive closing the app, what shows when there is nothing in the list, does deleting ask first, and how many items does it need to handle? While these guesses are often reasonable, the trouble is that developers find out which ones they disagree with only after the build is complete, when changing them is expensive.

Decide First With kokoro-baton

The first tool, kokoro-baton, runs in an empty folder and takes a one-sentence description of the app. Instead of a blank prompt, it presents multiple-choice questions about undecided aspects, explaining the consequences of each option. The output is a kokoro-spec.md file that documents decisions, flags open questions, and defines verification steps. Crucially, the author warns against asking the AI to run Baton for you. If the AI answers the questions, it defeats the purpose of forcing human intent into the spec.

Verify Without AI Blind Spots

After handing the spec to Claude Code, the second tool, kokoro-aegis, performs a check. Unlike Baton, Aegis uses no AI and sends no code to a server. It generates a report with two parts: static analysis holes found by reading the code, and a manual checklist derived from the spec. This separation is vital because a checker built on the same model as the builder shares its blind spots. Aegis ensures that verification is independent of the generation logic.

Infrastructure and Cost Model

Both tools are free, require no sign-up, and run via npx with Node.js 22+. Baton uses an Anthropic API key for its questioning engine, costing under 1ยข per round, while Aegis is entirely local. The author highlights a privacy nuance: Baton sends prompts and answers to usekokoro.com, which holds the prompts and calls Anthropic with the user's key, but does not store the text or key permanently. A short hash is kept for rate limiting and deleted after 2 days.

Key Takeaways

  • Explicit Decision Phase: kokoro-baton forces users to answer multiple-choice questions about app behavior before building, preventing silent AI assumptions.
  • Independent Verification: kokoro-aegis checks code without using AI, avoiding the shared blind spots of same-model validation.
  • Spec-Driven Workflow: The output is a markdown spec that guides the AI builder and defines manual checklists for human review.
  • Low Friction: Both tools are npx-based, free, and require no installation, with minimal API costs for the decision phase.

The Bottom Line

This workflow is a necessary correction to the over-reliance on AI intuition in early-stage development. By separating human intent from machine execution, these tools turn vague prompts into actionable specs, ensuring that the final product reflects what the developer actually wanted, not just what the model guessed.