agent-nextup: Skill for Claude Code

align-scope — settle the scope before building

Goal: when the work starts, there is nothing left that "we will figure out later" — and every assumption behind it is either verified or written down as an assumption.

Use it when the ask is broad ("make the reporting better"), when two readings of the request would lead to materially different work, or when the cost of building the wrong thing is high. Skip it when the next step is obvious — an alignment ritual on a one-line fix is just friction.

1. Separate what you can find out from what only the human knows

Before asking anything, split the unknowns into two piles:

• Findable: the answer exists in the code, the files, the history, or the environment. Look it up yourself and spend the human's attention only on what needs their judgement. When a lookup will take a while, hand it to a sub-agent if your host has them and carry on: only the questions that depend on its answer wait for it.
• Human-only: priorities, trade-offs, what "good" means here, what happens to whoever uses this, what the deadline really is. These are the questions worth asking.

Say what you found before you ask. "The task files already carry a priority field, so ordering is possible today — the question is whether…" is a better question than "how should tasks be ordered?"

2. Ask in rounds: the whole frontier at once

The open decisions form a tree — some can only be asked once others are settled. The frontier is every question whose prerequisites are already answered: the ones you can ask now without guessing at answers you have not heard yet.

Ask the whole frontier in one round. Number each question, put your recommended answer under it with the reason in a line, and word the question so that "yes" accepts your recommendation:

Q1 — <short title>: <the question; the options, if it is a choice>
   Recommended: <your answer> — <why, in one line>

Q2 — <short title>: …

Lead the round with the question whose answer changes the most about what gets built. A question that depends on another question in the same round belongs to the next round. Then wait for the answers: each round settles part of the tree and moves the frontier outward, so recompute it and ask the next round.