/align

name: align
description: Resolve consequential choices and shared intent before implementation. Use only when explicitly invoked.
disable-model-invocation: true

Hope Align

Use only when the person invokes `$hope:align`, `/hope:align`, or the host's equivalent skill command or picker. Continue follow-ups within that invoked task without requiring another invocation. Ordinary requests and calls from another skill do not activate this skill; continue authorized work through the ordinary workflow. Reading a shared reference does not invoke its skill.

Discover and test a design with the person to understand its goal and consequential decisions. Honor existing decisions, delegation, and implementation authorization. Align owns its open choices even when an implementation skill runs alongside it. Clear, authorized tasks and routine implementation choices need no new alignment or approval process.

Read `../write/references/writing-standard.md` for user-facing language.

Discover choices through their consequences

Start from the conversation, relevant code, and governing project sources. Keep evidence, user decisions, AI proposals, assumptions, and uncertainty distinct.

Understand the goal by following relevant situations from the starting need to the proposed outcome in its real setting. Investigate transitions that depend on unstated choices or unsupported assumptions. Derive questions from the task and evidence, without a fixed topic checklist or round count.

Use a decision tree to track discovered choices and their dependencies. A choice is material when it could change the goal, observable result, future options, long-lived constraints, or material risk, including architecture and data modeling with those effects. Leave routine, reversible mechanics to implementation. Reversibility alone does not make behavior or success criteria routine.

Work the complete decision frontier

The frontier contains every unresolved material branch whose prerequisites are settled. Ask the whole frontier in each round. Number the questions, explain their consequences, recommend a path, and offer realistic alternatives. Research what the AI can determine instead of asking the person for facts. Continue independent authorized work while answers are pending, but keep dependent work behind its unresolved choices.

After each answer or new evidence, trace its effects in the real situation and its interactions with other commitments to discover choices beyond the current tree. Update the problem understanding and tree, then recompute the frontier. Reopen settled choices when their basis changes. Close each material branch through a decision, deliberate exclusion, or explicit delegation, applying agreement only to the choices it covers. Unstated consequences remain open.

For uncertainty that conversation cannot resolve, identify a useful observation and gather evidence or recommend a small probe. Use actual screens or mockups when judgment depends on seeing the experience. Read `references/design-directions.md` when the person also invokes Design or visual evidence needs preservation. Bring the result back into discovery and agreement; uncertainty alone does not close a branch.

Test the understanding before confirmation

Agreement and sufficient discovery are separate requirements. Once known material choices are resolved, return to the original goal. Walk through concrete situations with the choices applied together to test the outcome and its supporting evidence. Look for missing decisions, conflicts, or unsupported assumptions that the tree never exposed, and simpler ways to meet the goal.

Return material issues to discovery and agreement. Finish when all material branches are resolved and one proportionate walkthrough finds no new material issue. An empty question list or assent to recommendations does not establish this. A clear, narrow request may need no additional question round.

Summarize the goal, problem, success criteria, consequential decisions and effects, and the exclusions, delegation, or assumptions shaping the agreement. Report verification and its limits, distinguishing reasoned walkthroughs from observed behavior. Confirm new shared understanding without reconfirming an unchanged agreement.

After confirmation, use existing implementation authorization without asking again; otherwise leave implementation for the person's decision. An alignment-only request ends with the confirmed understanding.

Preserve understanding when needed

Read `references/artifact.md` when the person supplies an Align record or its HTML page, asks for a durable record, or another session or worker will rely on the agreement. Also use it when a material agreement or later human observation must survive the conversation. Otherwise the conversation is sufficient.

The reference owns record authoring, revision, rendering, migration, and handoff. Run its commands through `node "<skill-dir>/scripts/cli.mjs"`, replacing `<skill-dir>` with the absolute directory containing this file. In Claude Code it is `${CLAUDE_PLUGIN_ROOT}/skills/align`.

Report where the confirmed understanding remains. A receiving implementation AI inspects the current project and uses the agreement for intent and decisions.

The decision-tree and complete-frontier interview is informed by Matt Pocock's [grill-me](https://github.com/mattpocock/skills/blob/main/docs/productivity/grill-me.md).