Preflight · a free local workflow by Pathmode
Run a preflight before
your agent builds.
Coding agents build whatever they are handed. Preflight is a deterministic readiness check for the intent you are about to hand over: six calibrated gates, a named verdict, and one targeted repair question per failure. No API key, no signup.
title: "Checkout"
objective: "Make the checkout experience seamless and intuitive"
outcomes:
- Checkout feels faster
- Users are happier
constraints:
- fast
- secure
edge cases:
(none)
verification:
- test it
✗ Preflight failed. 6/6 checks blocking.
✗ Objective is too vague — name the actor and the concrete failure they hit today.
✗ Title is missing or generic — name the intent in a short, specific noun phrase.
✗ Outcomes are not all measurable — at least two-thirds need a testable or observable threshold.
✗ No hard constraint — name at least one thing that must never happen.
✗ No edge case with a defined expected behavior.
✗ No concrete verification — describe at least one check specific enough to run.
What short, specific name distinguishes this work from a generic feature or fix?
Answer it in your own spec on the Your spec tab; the verdict reruns as you type.
This verdict is computed in your browser, right now, by the same deterministic gate that scores every Pathmode spec. Nothing above is mocked.
Buildability only: this local check does not evaluate supporting evidence or decide whether this is the right work.
The gate is local and has no memory. It reads your file and scores it. Connect your repository to add private evidence the file cannot hold, decisions attributed to the person who made them, and review by a teammate who never opens the terminal.
Choose your coding agent
Preflight is free and uses the coding agent you already have. No Pathmode account or API key is required.
/plugin marketplace add pathmodeio/claude-plugin
/plugin install pathmode@pathmodeInstalls the Pathmode plugin and MCP server for Claude Code.
Then run /preflight in the repository you're about to change.
Your agent calls the check_intent_readiness tool and returns this exact verdict on your own spec, in your terminal. The Claude Code plugin also bundles the compile, pressure-test, and handoff skills. Your spec lives in intent.md in your repo, plain markdown you own.
Using another MCP client, or prefer automatic setup? This detects the clients installed on your machine and writes their configuration:
npx @pathmode/mcp-server@latest setupAlready added the MCP server with npx? Remove the manual entry first so you do not end up with a duplicate pathmode server.
No intent.md yet?
Start with the problem. Your agent will help write the spec.
Paste a support ticket, describe a user complaint, or explain what is broken. The bundled compile-intent skill asks one pointed question at a time, reads the repo for implementation context, and saves the result as intent.md. Then Preflight runs the same six gates before anything gets built.
> Help me write an intent spec for [the problem you want to solve]It is part of the local workflow above—no separate product, account, or configuration.
Six gates. Pass or fail, with the reason named.
No model grades your spec. Every check is plain code with a hard floor, so the same spec gets the same verdict every time, and a failure tells you exactly what to fix.
Title
Names the intent in a short, specific noun phrase. "Bulk edit" passes; "New feature" is a label, not a name.
Objective
Names the actor and the concrete failure they hit today. Buzzword objectives with no concrete number are rejected.
Outcomes
At least two, and at least two-thirds carry a testable or observable threshold. "Users are happier" does not count, even with a number attached.
Constraints
At least one thing that must never happen. "Fast" and "secure" are adjectives; a constraint is violable.
Edge cases
At least one real scenario with a defined expected behavior, not a restatement of the outcome failing.
Verification
At least one check specific enough to actually run. "Test it" is a gesture, not a check.
The gates are calibrated against a labeled corpus of 98 real spec fragments, good and vague, and regression-guarded in CI. Short, specific writing passes. Long, dressed-up vagueness does not.
Your agent will not push back. Preflight will.
Ask a model whether your spec is ready and it will say yes, then build its best guess of what you meant. A preflight is the opposite kind of answer: reproducible, explainable, and indifferent to how confident the prompt sounds. When a gate fails, the repair is a conversation. When all six pass, the handoff is earned.
Preflight never blocks you. It tells you the risk you are taking, and you decide.
Working with a team? The same gates run in Pathmode with shared evidence, decision history, and verification against what actually shipped.