Interview the user relentlessly about the thing they want to work on. Walk the decision tree one branch at a time, resolving dependencies between decisions.
-
Identify the subject. Take whatever the user described — plan, task, design, idea, architecture decision, feature, process — as the root. Ask if unclear.
-
Ground via sub-agents before the first question. Spawn
(typically
or
) in parallel to read the relevant codebase
and anything the user attached or referenced — files, URLs, tickets, docs, screenshots. Delegate the reading; don't burn your own context. Skip only if the subject is genuinely context-free.
-
Map the tree mentally. Using what the sub-agents returned, find the decision that gates the most other decisions. Start there.
-
Ask one question at a time using
. Never batch.
-
Always propose a recommended answer with each question, plus a one-line reason. The user should be able to accept with one tap or push back.
-
Explore instead of asking when you can. If a question is answerable from code, config, git history, attached files, or linked tickets — spawn a sub-agent to find the answer. Only ask the user for judgment, preference, or knowledge you can't derive.
-
Follow dependencies. Each answer unlocks the next most load-bearing question on the current branch. Don't wander to siblings until the current branch is resolved.
-
Keep going until the tree is resolved. Stop only when every decision has an answer, the user ends the session, or what's left is pure implementation detail.
-
Summarize at the end. Produce a concise recap of the decisions made.
Layer these on top of the tree-walk to sharpen each exchange. They don't replace the recommendation-first format — they make sure the answer being recommended (and accepted) is actually load-bearing. Deploy them between tree-walk steps as the situation calls for them, not as a fixed checklist.
Anything the user is trying to figure out — plans, tasks, ideas, features, architecture decisions, process changes. The pattern is the same: find the root, ground via sub-agents, walk the tree, one recommended question at a time.