Draft your first intent in your repo, no account required

The first spec most teams hand an agent is the prompt itself, composed in the chat box and gone when the session ends. Three weeks later nobody can say what the feature was supposed to do, only what it does. The fix doesn't need a platform: your first intent is a markdown file in your repo, drafted in conversation, preflighted before you trust it, and committed next to the code it governs. No account exists anywhere in the loop.

The input

A repo and a rough idea of what you're about to build. No intent.md yet, no workspace, no API key. The conversation you were going to have with the agent anyway is the raw material.

The flow

1. Install once, keyless. In Claude Code: /plugin marketplace add pathmodeio/claude-plugin. In any other MCP client: npx @pathmode/mcp-server with no key runs in local mode. Local mode makes no network requests at all; your spec stays on your machine.

2. Describe what you're building. The compile-intent prompt turns the conversation Socratic: the agent drafts a spec from what you said and what the repo shows, and every field it inferred rather than heard is labeled an assumption. You correct a draft instead of composing from zero.

3. Preflight the draft. check_intent_readiness scores it against six calibrated gates: title, objective, outcomes, constraints, edge cases, verification. Deterministic, no model call. The same draft always gets the same verdict, and a failing one names the exact blockers.

4. Repair one question at a time. For each blocker the agent asks one targeted question and proposes its best-guess answer from the repo, so you correct rather than compose. A failing verdict never blocks work: the gate reports, you decide.

5. Save it where the code lives. intent_save writes intent.md to your project root with the current verdict stamped in its frontmatter. It diffs in review and travels with the branch. Need agent-native formats too? intent_export emits CLAUDE.md, AGENTS.md, or .cursorrules from the same spec.

The output

• An intent.md in your repo root: versioned with the code, reviewed like code, readable by any agent
• Assumptions labeled as assumptions, so inferred intent never masquerades as confirmed intent
• A readiness verdict in the file's frontmatter, so the spec carries its own state
• No account, anywhere in the loop. If the team later wants a shared workspace, npx @pathmode/mcp-server setup pm_live_xxx syncs the same file up; until then the repo is the source of truth.

Why this beats writing the markdown yourself

You could write intent.md by hand, and plenty of engineers do. What the hand-written version lacks is the gate: nothing tells you the outcomes aren't observable or the verification has no fastest check until an agent has already built against them. Drafting in the compile conversation gets you the structure for free; the preflight tells you, deterministically, whether it's load-bearing.

Try it yourself

1. /plugin marketplace add pathmodeio/claude-plugin in Claude Code, or npx @pathmode/mcp-server in any MCP client
2. Say what you're about to build and ask for a first intent
3. Answer the repair questions until the verdict flips
4. Commit intent.md with your next PR

Related

• Use case: Run a deterministic preflight before your agent starts building
• Use case: Hand off a spec to Claude Code without losing context
• Playbook: Your First Spec in 10 Minutes