Have the agent relentlessly interrogate a plan, decision, or idea until both parties reach a shared understanding: map all decisions as a "design tree," ask only the questions that can be asked at the current stage in each round while providing recommended answers, and do not proceed until all questions are answered.
skills.sh install count 1016434 (+8175 / 24h)
Security Audit: Audited (no risks found)
Install: npx skills add https://github.com/mattpocock/skills --skill grilling
Rationale
- Specific rules: Decisions with established premises form the "frontier." In each round, number and ask all questions on the frontier, attaching a recommended answer to each; phrasing the question so that answering "yes" implies accepting the recommendation. Questions dependent on answers to other questions in the same round are deferred to the next round.
- Distinguish facts from decisions: When environment information such as files or tools needs to be checked, the agent dispatches a sub-agent with a new context to perform the check, rather than asking the user to do it; while checking, it first asks questions that do not depend on that information. All decisions are left to the user.
- The process ends only when the frontier is cleared (every branch of the design tree has been traversed and no default assumptions remain), and execution begins only after user confirmation.
- The grill-me skill released yesterday merely delegates to this skill; this is the core implementation. The same repository contains 37 other skills (representing only attention level).
- skills.sh installs 8175 in 24 hours, with a cumulative total of 1016434 (representing only attention level).
Usage
- npx skills add https://github.com/mattpocock/skills --skill grilling
- Trigger by saying "grill me" or other phrases containing "grill," or when requesting a stress test of a solution; reply in each round using the format ❓ Q1 / ➡️ Recommended Answer.
- Dispatching sub-agents to check facts requires client support for sub-agents; if unsupported, the agent must perform the checks itself.
Risks
- The audit result is a pass; SKILL.md contains only textual rules, with no scripts, network requests, or keys, consistent with its content.
- Questions will come round after round, requiring traversal of the entire tree even for minor changes, increasing both time and token usage; do not trigger this if you want quick results.
- When sub-agents check the environment, they read local files and tool outputs; the scope is determined by your client's permissions.