What is grill-for-unknowns?

grill-for-unknowns is a skill you can add to an AI coding agent such as Claude Code, Codex, or Hermes, so the agent pauses and asks questions before it starts building something. Instead of jumping straight into writing code, the agent first looks at the real codebase, documentation, or source files, then interviews the user only on the points that actually matter, before any implementation begins. The idea behind it is to sort what the agent is dealing with into four kinds of knowledge: things it can already confirm from the docs or code, things it needs to ask the user about, terms and language that need to be clarified, and blindspots that nobody has thought to ask about yet. The skill is meant to be triggered when a task is unfamiliar, vague, complex, or has already failed once because of wrong assumptions. The project describes itself as an adaptation of an existing pair of skills by another author, and it says clearly what it keeps from those and what it adds. It keeps the core idea of a relentless back and forth interview, but it adds a defined standard for what counts as a good question, meaning each one must be grounded in something real, matter to the outcome, and be answerable, rather than just being a generic clarifying question. It also works as one self contained package rather than depending on other skills being installed separately, and it extends past the planning stage to produce written artifacts afterward, such as a launch packet for handing work to other agents, notes on new unknowns discovered during implementation, and small design decision records. Installation differs depending on which agent you use. For Claude Code the recommended path is through a plugin marketplace command, while for Codex and Hermes the readme recommends cloning the repository once and creating a symlink into each agent's skills folder, so that a single update later refreshes every agent that uses it. The project is released under the MIT license, and because it builds on someone else's earlier MIT licensed work, both original authors are credited in the license file.