Interview Yourself For Context

I've noticed the same failure point over and over when starting something with an agent: I sit down with a rough idea, type a few sentences, and get a plausible-looking answer to a problem I hadn't actually finished defining. The agent isn't guessing badly, it's working with a half-formed brief. The fix isn't a better prompt or a smarter model. It's making myself answer a handful of questions before the agent gets involved at all, which is what I built context-compass to do.

/context-compass A fast, always-on warm-up that interviews you about your own idea before you hand it to an agent, turning a rough thought into a ready-to-run brief in a few minutes.

That's the actual mechanism: context-compass interviews me, not the agent. Most of the time I think I know what I want until I'm asked a direct question about it, and then I notice I hadn't actually decided. Scope, audience, whether this is a one-off or something I'll come back to, what's off-limits. None of that is unknowable, it's just unasked, because nobody asks it, including me, when I'm moving fast and just want to get started. The interview runs five short steps: describe the idea, say what's already in place, resolve a couple of things I hadn't decided, name what's off-limits, say where the work goes next. Every question but the first offers concrete options rather than a blank box, because recognizing the right answer among a few real ones is faster than generating one from nothing, and that's exactly when I'm least sure what I want.

The step where I actually learn something

The core of it is the step that surfaces what I haven't decided yet, pulling two or three relevant angles from a small library by task type: scope and lifespan for something I'm building, audience and length for something I'm writing. It only offers what fits what I've already said, not the whole list.

Insight: the point of this step isn't to collect information I already have. It's to ask me the two or three questions I would have skipped past on my own, and force a real answer instead of an assumption.

This is what separates a self-interview from a form. A form records what you already know. This step exists because the most useful thing an interview does is point at the gap between what you said and what you meant, and I'm consistently the person least likely to notice that gap on my own. More than once I've picked an option I hadn't consciously considered and realized, mid-answer, that it was the actual crux of the task. The interview also knows its own limits: if the idea is bigger than a few minutes deserves, it says so and points toward a longer session instead of forcing a quick pass through something that needed more.

Why it holds

The brief that comes out the other end holds up because every line in it is something I decided, not something an agent inferred from a few rushed sentences. Interviewing myself first means the agent is executing against decisions I already made and can stand behind, not reconstructing my intent from thin context. That's sturdier than a good prompt, because a good prompt still asks the agent to guess at what I left out. This skips the guessing by making sure I've done the deciding first.

It's the same instinct behind everything else I build for how I work with agents day to day. If you want to see how the rest of it fits together, my skills live at github.com/Kimotep/skills.