Skip to main content

FOR EVERYONE

Skill gaps — asked once

Sometimes an agent reaches for a capability it doesn't have — a way to talk to GitHub, a memory store, a connection to a tool you use. When that happens, you don't want a stream of interruptions. You want one clear question, asked at the right moment, and never asked again.

That's what smart skill prompts do. They keep the conversation calm.

What you get​

  • Asked once. If an agent needs something it's missing, it raises it a single time — up front, before it gets deep into the work — not halfway through, and not over and over.
  • Only when it's not obvious. If the right answer is clear from context, the agent just proceeds. The question only appears when your input genuinely matters.
  • Remembered. Whatever you decide is recorded for the project. The next time the same need comes up, the agent already knows — so it doesn't ask again.

The whole point: no nagging. One thoughtful question, then quiet.

What it feels like​

You ask for some work. Before diving in, the agent notices it would benefit from a capability you haven't wired up — say, access to GitHub. It pauses to ask:

This task would go smoother with GitHub access (the gh CLI). Want me to use it, or should I work without it?

You answer once. From then on, in this project, it remembers your choice and carries on without asking.

If the need is obvious — or if you've already answered — you'll hear nothing at all. The agent simply gets on with it.

Where your answers live​

Your decision is tied to the project, not to a single chat. Close Claude Code, come back tomorrow, start a fresh session — the agent still remembers what you told it. That's part of the kit's Memoria value: context that survives, so you're never re-explaining yourself.

How it works under the hood
  • The rule. /kit-init scaffolds a smart skill-gap rule into .claude/rules/. It's always-on guidance that tells agents: when you detect a capability gap (a "skill" or tool you lack) and the resolution is non-obvious, surface a single up-front prompt to the user; otherwise proceed silently. The rule explicitly forbids repeated or mid-task nagging.
  • Non-obvious test. The rule asks the agent to prompt only when your answer would actually change the plan — e.g. whether to enable an opt-in tool like the GitHub CLI (gh), MemPalace memory, or Google Workspace (gws). When the right call is clear from context, no prompt is shown.
  • The remembered record. Your answer is written down for the project so it isn't re-asked. Where MemPalace memory is enabled at /kit-init, the decision is stored in the project's wing and recalled in later sessions; otherwise it's noted in the project's .claude/ config. Either way, the lookup happens before any new prompt, so a resolved gap stays resolved.
  • Relation to /kit-customize. This is the lightweight, in-flow path. When you want to deliberately add, wire, or search for skills and tools on your agents, that's /kit-customize. Smart skill prompts handle the gaps that surface while you're simply trying to get work done.