Give Cursor or Claude Code an IntentSpec in intent.md beside the code it governs. Pathmode's MCP server reads that repo-bound file first, checks whether its objective, outcomes, constraints, edge cases, and verification are ready, and carries the decision into the coding session. Connected teams can add private evidence and review context without moving the decision out of the repo.
A team pointed Claude Code at a refactor of their onboarding flow. It cleaned up the code, passed every existing test, and merged. It also dropped a retry-on-failure step that three support tickets had identified as load-bearing. The agent saw a function to simplify. The repo never told it which outcome that function protected.
Why a capable agent still ships the wrong change
A coding agent optimizes for the task and repository in front of it. A one-off prompt is a thin slice of someone's judgment: useful for the immediate instruction, but easy to lose on the next session or branch. More implementation capability cannot recover a product decision that was never made explicit.
The durable handoff is a small, reviewable artifact beside the code:
The repo gets the decision. Pathmode keeps the private evidence and human judgment behind it.
The handoff, end to end
Evidence → agent draft → human authorization → intent.md → implementation → verification
- Evidence informs the draft. Tickets, interviews, observed friction, metrics, and codebase facts are inputs. They can support or challenge a choice; they do not choose for you.
- The agent drafts an IntentSpec. It turns the current understanding into an objective, observable outcomes, constraints, edge cases, scope, and verification checks.
- A human reviews and authorizes it. Assumptions stay visible. Waivers and judgment calls remain explicit rather than being smuggled in as generated certainty.
- The decision is saved as
intent.md. The file is versioned with the code, travels with the branch, and is readable by any agent that works in the repository. - Cursor or Claude Code implements against it. The agent can make local technical choices while preserving the authorized outcomes and boundaries.
- Verification closes the loop. Checks and shipped signals show whether the implementation satisfied the spec and whether anything important regressed.
What the MCP server gives the agent
- The repo's active decision — local mode reads
intent.mdfirst, so “current intent” is bound to the code being changed rather than guessed from a workspace. - A deterministic readiness verdict — six gates identify thin objectives, unobservable outcomes, missing constraints, uncovered edge cases, and vague verification before implementation starts.
- Useful exports and workflow prompts — the same IntentSpec can produce agent-native context without creating a second source of truth.
- Optional team context — connected mode can supply private evidence, named guest feedback, signed-in authorization, open findings, and cross-repo state that should not be committed to the repository.
Prompt-only vs repo-bound intent
| Prompt-only | Repo-bound IntentSpec | |
|---|---|---|
| What governs the work | The current instruction | The authorized decision in intent.md |
| What survives a new session | Whatever is re-explained | Objective, outcomes, boundaries, and checks |
| Who can review the decision | Whoever sees the prompt | Anyone reviewing the branch |
| Where private evidence lives | Often scattered across tools | Referenced from the file; retained privately in team mode |
| How “done” is judged | The task appears complete | The implementation is checked against explicit outcomes |
Cursor and Claude Code are not downstream recipients of a hidden workspace decision. They read the decision where they already work: the repository. Pathmode helps make that decision ready, portable, and traceable.
Frequently asked questions
- How does Pathmode connect to Cursor or Claude Code?
- Add the Pathmode MCP server to your coding agent. With no API key it works locally and reads intent.md from the repository. Connected mode can enrich that same repo-bound intent with private evidence, review state, and implementation findings from a Pathmode workspace.
- Is this a replacement for Cursor or Claude Code?
- No. Cursor or Claude Code remains the code brain: it inspects the repository, plans the implementation, writes code, and runs checks. Pathmode helps compile and preflight the product intent that governs that work. The repo gets the decision; Pathmode can keep the private evidence and human judgment behind it.
- Who decides what goes in intent.md?
- An agent can draft the IntentSpec from your description, codebase, and available evidence, but a human reviews and authorizes the product decision. Evidence informs the judgment; it does not make the decision on its own.
- What does verification actually check?
- Verification names concrete checks for the outcomes and regressions in the IntentSpec: the fastest feedback loop, higher-confidence tests, manual fallbacks, shipped signals, and regression guards. The implementation agent runs the applicable checks and reports what the results support.