How builders ship the right thing.
Follow a proposal from request to review, bring evidence into a decision, or prepare the handoff to your coding agent. Pick a workflow or role to find the relevant steps.
Draft your first intent in your repo, no account required
Install the keyless MCP server, describe what you're building, and get a preflighted intent.md in your project root. Assumptions labeled, blockers named, the Pathmode server runs locally. No account required.
Hand the authorized proposal to your coding agent
Carry the product manager's corrections and revision authorization into Claude Code, Cursor, or Codex, then check the implementation against the agreed intent.
Generate a PR description from the linked IntentSpec
The agent that built the work already holds the spec it built against. Have it write the PR description from that spec: objective, outcomes, verification criteria, and edge cases, so reviewers see intent and tests in the same place.
Run a deterministic preflight before your agent starts building
Six calibrated checks, no model call, no account. The preflight reads your intent.md, names the exact gates that block it, and the same spec always gets the same verdict. Repair happens one question at a time, with readiness kept separate from approval to build.
Let your agent pull intent live with the Pathmode MCP server
Stop pasting specs into the agent. With a connected workspace, Claude Code preflights the repo's intent.md, reads the full execution context, stops for open PM requests and for a person to authorize the exact revision, then hands the pull request back for a diff-graded verdict.
Verify the pull request against the spec it was built from
Name the PR and Pathmode reads the diff from GitHub itself, grading every outcome, constraint, and edge case against what the code actually shows, with citations. A summary of your own work is weaker evidence, and the verdict says so.
Let a coding agent correct the spec it's building from
An agent building your spec discovers the spec was wrong. Instead of burying that in a PR comment, it records a finding that rides into the next agent's prompt and flips the falsified check to failing, then proposes the exact correction for a person to accept. The loop converges, and a human closes it.
Review a spec for implementation risk before build
The product side wrote the goal and the evidence. You add what they can't see — the technical edge cases, the constraints, the out-of-scope lines that keep an agent from wandering. Catch the expensive surprises while they're still cheap to fix.
Sequence work so agents have the dependencies they need
Check the prerequisites behind related intents, agree the build order, and hand the agent an authorized proposal based on contracts that are actually available.