Your First Intent in Pathmode

If you're reading this, you probably understand the philosophy of Intent Engineering. But how do you actually do it in Pathmode?

Here's a step-by-step walkthrough of creating your first Agent-Ready IntentSpec. The short version: you have a conversation, and a spec crystallizes.

Pathmode organizes work as Workspace → Products → (Evidence, Intents, Build Queue). A Product is a slice of your software with its own evidence, vocabulary, and intents — Checkout, Onboarding, Admin Console.

From your workspace home, open an existing Product or create a new one. Everything below happens inside that Product.

Intent Engineering starts with a problem, not a solution. Before you write a single line of spec, you need proof.

1. Go to the Evidence tab inside your Product.
2. Click Add Evidence.
3. Choose a type — friction, quote, observation, metric, or request.
4. Paste in real user signal. e.g., "I tried to pay but the screen just hung there for 10 seconds. I had no idea if my card was charged."
5. Save.

Add a few more while you're here. Evidence with one data point is an anecdote; evidence with a cluster is a pattern.

1. Open the Intents tab in the same Product.
2. Click New Intent.
3. In the prompt box, describe what you want to improve in plain language. e.g., "Checkout payment step is hanging for 10+ seconds and users don't know if it worked."
4. Pathmode surfaces related evidence from your Product. Check the items that apply.
5. Click Create.

You're now in the Intent Editor.

The editor isn't a form. The left side is a conversation with a Socratic thinking partner; the right side is the Spec Panel — your spec as it crystallizes.

The AI will challenge you. Expect questions like "What does 'fast' mean here in numbers?" or "What should happen on a network timeout — retry, fail, or queue?" Each question comes with a recommended answer grounded in the evidence you attached, so you can accept, refine, or push back.

Answer in plain language. The AI extracts structure as you talk.

When the AI extracts something concrete — an objective, an outcome, an edge case — it merges into the spec immediately. An InlineSpecCard appears in the conversation showing what changed, with a 10-second Undo in case it misread you.

You'll see two indicators above the conversation:

• Topic Coverage — passive fill levels for Objective, Goal, Outcomes, Edge Cases, Verification.
• Thinking Quality — scores for specificity, testability, coherence, and tension (0–3 each).

Aim for 2 or 3 across the board before exporting.

As the spec takes shape, Pathmode automatically links your evidence to the parts it justifies — an outcome, an edge case, the objective.

Open the Spec Panel and look for the N ev. badges next to each section. Click one to see the linked items. If something looks miscategorized, drag it to a different section or detach it. Manual anchoring is now a refinement step, not the starting move.

What could the AI agent do wrong when fixing this? Tell it.

In the Spec Panel, add Constraints — hard boundaries the agent cannot cross.

• "No double-charge on retry"
• "Must support iOS Safari 15+"

Or just say it in the conversation: "It can never double-charge on retry." The AI will add it for you.

How will the agent prove it worked? Verification isn't just tests — it's the full feedback loop.

• Fastest check — the unit or integration test that catches a regression in seconds.
• Manual fallback — what a human runs if automation can't reach.
• Observable signal in production — the metric, log, or event that confirms it shipped correctly.
• What shouldn't regress — adjacent flows the change must not break.

Fill these in the Verification section of the Spec Panel, or describe them in the conversation.

When Topic Coverage is full and Thinking Quality looks healthy, you're done.

Click Export to copy the spec, hand it to Claude Code or Cursor directly, or pipe it through the Pathmode MCP Server so your agent reads it from the workspace. The agent now knows what you want, why you want it, what it cannot do, and how to prove it worked.

That's the loop. Evidence in. Conversation through. Spec out. Agent ships.