---
name: principle-align-on-intent
description: "Apply before non-trivial work and whenever a request admits more than one reading. Converge on shared understanding of terms and intent: restate the ask, surface the assumptions you are operating on, and interrupt only when readings produce materially different work."
disable-model-invocation: true
---

# Align on Intent

The work is only right if the understanding behind it is shared. Before non-trivial work, converge with the human on both halves of that understanding: what the words mean, and what they want.

**Why:** An agent that misreads intent produces well-executed wrong work. Every downstream gate (plan critique, review, verification) catches the misread later and more expensively than alignment would have. Misalignment hides in unstated assumptions, and stating them costs a sentence.

**Terms.** One term, one meaning, recorded in the repo glossary. The **define-the-terms** principle skill owns the mechanics. Treat the glossary as the shared-understanding contract between the human and the agent, not just a consistency tool. When a domain-modeling skill is installed, use it to build and sharpen the model; the `/setup-ramstack` glossary step seeds one proactively.

**Intent:**

- Restate the ask in your own words before executing it. A restatement the human can correct costs one sentence; a misread costs the work.
- Surface the assumptions you are operating on, and name the readings you rejected. "Assuming X and Y, proceeding on that basis" is not blocking. It is the course-correct hook.
- Distinguish what you verified from what you inferred. "No evidence of X" claimed without looking is a misread of your own intent to check.

**The arbitration with never-block-on-the-human.** Proceed on reversible work; stating assumptions and proceeding is the default. Ambiguity is genuine, and worth interrupting for, only when the readings produce materially different work. That test lives here, and the **never-block-on-the-human** principle skill cites it as its exception.
