Intent Interviewer
Clarify the real problem through one-question-at-a-time interviews before planning or coding.
The skill is limited to interactive questioning and invokes no external tools. It explicitly forbids use in non-interactive contexts and requires explicit user confirmation before saving an intent, supporting least privilege, confirmation, and external-effect control. Deductions apply because sensitive-data handling, data-flow disclosure, save-location boundaries, and rollback or undo are unspecified; publisher identity is also unverified.
The triggers, process, stopping condition, and failure branch are generally self-consistent, including an instruction to stop and explain when convergence cannot be reached. Points are deducted because the roughly 95% confidence threshold and “predict the next three questions” test are subjective, and no executable tests cover the skill's key paths; static review caps this dimension at 10.
The skill defines target scenarios, explicit trigger phrases, non-fit requests, and non-interactive environment limits, and distinguishes downstream skills. Points are deducted because finer input/output boundaries, false-trigger handling, Chinese-language guidance, and mainland-China environment evidence are absent.
The documentation is well layered, with overview, triggers, process, example, skill interactions, rationalization rebuttals, red flags, and a verification checklist. Shared repository context supplies MIT licensing, installation guidance, CI, and maintainer information. Points are deducted because the skill lacks its own version, changelog, maintenance responsibility, and troubleshooting guidance; some governance evidence is only indirect repository-level context.
The goal is clear: produce a user-confirmed intent statement with Outcome, User, Why now, Success, Constraint, and Out of scope fields. The example demonstrates the value of reframing “dashboard” into an “experiment tracker.” Points are deducted because there is no real execution evidence or comparative proof of output quality; static review caps this dimension at 7.
The skill provides an auditable checklist requiring a hypothesis, confidence, one-at-a-time questions, restatement, explicit confirmation, and a stopping test. Points are deducted because the supplied CI validates repository structure and generic evaluations but does not show dedicated coverage of interview-me's key behaviors or independent third-party reproduction; static review caps this dimension at 5.
- The skill depends on live, sustained user responses and cannot reliably complete its deliverable in asynchronous, batch, or unattended environments.
- The “about 95% confidence” and “predict the next three questions” criteria lack objective operational tests, risking premature termination or excessive questioning.
- The skill does not instruct agents to avoid collecting passwords, tokens, or other sensitive information, nor does it define access control or deletion procedures for saved intent documents.
- No Chinese prompts or localized examples are provided, so semantic triggering and interaction quality for Chinese users require additional validation.
What does this skill do, and when should you use it?
Interview Me is a requirements-clarification workflow for separating what users actually want from conventional wording. It asks one focused question at a time about the user, motivation, success criteria, and binding constraints, while exposing the agent’s current guess. It continues until the agent can predict the user’s intent with high confidence, then asks for explicit confirmation. It is designed to precede definition, specification, and planning work.
States a one-sentence hypothesis about the user’s intent and gives it a confidence estimate; asks one focused question at a time with a reasoned guess attached; probes convention-driven or sophistication-signaling answers such as “scalable” or “clean architecture”; restates the converged intent as outcome, user, why now, success, constraint, and out of scope; waits for an explicit confirmation rather than treating “sounds good” or “whatever you think” as approval; and stops to flag missing foundations if several rounds still do not make the user’s likely reactions predictable.
- A product or engineering lead asks to “build a dashboard” without identifying its users, purpose, or success criteria.
- A developer asks to “make it faster” without specifying a target, context, or binding constraint.
- A user explicitly says “interview me,” “grill me,” or asks to stress-test the idea before starting.
- An agent is tempted to create a plan, specification, or code while key requirements remain unstated.
What are this skill's strengths and limitations?
- Finds requirement misalignment before plans, specifications, or code create switching costs.
- Makes the agent’s assumptions visible by attaching a guess to every question.
- Requires an explicit out-of-scope boundary and explicit user confirmation before stopping.
- Shows no dependency on shell, network, filesystem, or MCP tools.
- Requires an interactive user and cannot be used directly in CI, scheduled jobs, or autonomous loops.
- Prioritizes verification over immediate execution, so it conflicts with an explicit request for speed over verification.
- The source provides no automated test suite, persistence mechanism, or dedicated structured-output tool.
- Its deliverable is a confirmed statement of intent, not a specification, plan, or task breakdown.
How do you install this skill?
Install only this skill from the repository:
npx skills add addyosmani/agent-skills --skill interview-me
Or install the complete collection:
npx skills add addyosmani/agent-skills
The repository documents integrations for Claude Code, Cursor, Codex, Copilot, Cline, and other agents; follow the relevant client-specific setup guide for native or local installation.
How do you use this skill?
Use it in an interactive session after stating an underspecified request, or invoke it with “interview me,” “grill me,” “are we sure?” or “stress-test my thinking.” The skill begins with a hypothesis and confidence number, then asks one question at a time. It requires a live, responsive user and is not intended for CI, scheduled runs, or autonomous loops.
How does this skill compare with similar options?
It sits at the earliest point of the Define phase: clarify intent before handing the result to idea-refine, spec-driven-development, or downstream planning skills. Unlike doubt-driven-development, it extracts intent before a decision or plan exists rather than adversarially reviewing an existing artifact.