The Ten Questions

Adapted from The Socratic Intent Engineer by Martin Montero. Each question carries up to four lenses — First-Principles, Intent, Agent Skills, and the Human Edge — applied where they earn their place. Run these before you point an agent at anything; the Build Studio turns your answers into a project constitution, spec, and agent prompt.

I. Understanding the goal

1 · What exactly do I want the AI to do? Dissect the outcome — creative piece, summary, code, a multi-step workflow, an autonomous task? First-principles: what’s the irreducible problem? (Not “a report” but “three people understand three risks in five minutes.”) Intent: am I defining the task, or the goal behind it? Human edge: what actually needs to exist — the thing nobody asked for but everyone needs?

2 · Why do I want this output? The why illuminates the what. First-principles: challenge the why itself — do they need a weekly report, or to stay informed of changes? Intent: this is where a measurable proxy and the actual purpose diverge. Connect the work to the goal before writing a prompt.

3 · What are the constraints and success metrics? Audience, prior knowledge, emotional impact, measurable criteria. First-principles: for each constraint — hard requirement, or “the way it’s always been done”? Intent: set metrics at the right altitude (lifetime value, not resolution time). Agent skills: build evaluation into the skill — self-checks and escalation rules.

II. Defining the input

4 · What essential information does the AI need? Don’t assume it shares your knowledge. Intent: for agents, context is organizational knowledge — values, trade-off hierarchies, decision boundaries, not just data. Human edge: your lived experience — the launches that flopped — is context no training set contains.

5 · How should I present this information? Structured data as JSON/CSV; long text with clear hierarchy. Intent: at agent scale this is architecture — goal structures, delegation frameworks with decision boundaries, feedback loops that close.

III. Shaping the output

6 · What is the ideal format? First-principles: not “what format?” but “what does the recipient need to do, and what removes the most friction from that action?” Agent skills: a skill might emit a spreadsheet with live formulas, or a dashboard, not a static table.

7 · What is the precise style and tone? Specific references, not vague descriptors. Human edge: style is taste made visible — name the exact voice, and recognize it when it arrives.

8 · What are the non-negotiable success criteria? State them before you run anything. Intent: are you measuring the right thing? Resolution time looked right until it wasn’t. Define success at the level of purpose, not task completion.

IV. Iterative refinement

9 · Did the output truly meet my criteria? First-principles: when it fails, resist tweaking the surface — is the failure in the prompt, or in your assumptions? Human edge: not “does it meet the criteria” but “does this have the aliveness that makes it worth putting into the world?”

10 · How do I revise to guide the AI toward the outcome? Refine constraints, add examples, switch techniques. Intent: close the loop — was the agent’s decision aligned with intent, and how do you measure drift over time?

Operationalized in the Build Studio

You don’t have to hold all ten in your head. The Build Studio asks the ones that shape the build — goal, the why behind it, success criteria — and folds your answers into the generated constitution, spec, and agent prompt, so the agent is aimed at your real intent from the first commit.

The complete workflow

A practical sequence whenever you design a prompt, build a skill, or architect an agent workflow:

1. Identify the fundamental problem — strip away format and convention; write the irreducible need in one sentence.
2. Connect to organizational intent — what goal does this serve? Encode the goal, not just the task (trade-offs, boundaries, escalation).
3. Question every inherited assumption — fundamental truth, or habit?
4. Decompose into irreducible components — smallest units, one purpose each.
5. Reconstruct from first principles — assemble the pieces that solve your problem; don’t copy a template.
6. Apply your taste — does this reflect your actual intent and your read on what the audience needs?
7. Test against reality — real scenarios, edge cases, adversarial inputs; evaluate against criteria you set beforehand.
8. Iterate with discipline and courage — one gap per pass; sometimes the right move is to throw it away because the direction was wrong. Then close the loop.

Then hand it to the agent — through PIE and the Build Studio — and keep every change green with the enforcement engine.