A local-first behaviour guidance layer that helps builders catch missing checks before AI-generated work becomes shipped risk.
AI coding agents make it easy to move fast. Nexpath helps reduce the chance that speed quietly becomes an unverified bug, an unsafe change, an unmaintainable codebase, or hours of manual review later.
Within Nexpath's broader behaviour-guidance vision, Prompt Enhancement is the first feature we're introducing as a practical twist: it saves vibe coders from writing every prompt detail by hand while preserving their original intent. When a task needs more rigour, it can suggest missing development practices, verification, or confirmation steps — encouraging users to give a mature development flow the time it needs to finish.
Demo: Prompt Enhancement in action, on YouTube
- AI can introduce a bug without making the risk obvious.
- Fast generation can leave a codebase harder to maintain after every change.
- Security, rollback, regression, and acceptance checks are easy to miss while momentum is high.
- Reviewing every AI-generated decision manually can cost more time than the generation saved.
Nexpath reduces these risks by noticing what the current workflow is missing and bringing the relevant check into the work at the right moment. It does not replace developer judgement or promise bug-free code; it helps make that judgement faster, more structured, and harder to skip.
- Build with your AI coding agent — Work in the tools and flow you already use.
- Review with Nexpath — Let the Prompt Quality Layer surface missing verification, safety, scope, or maintenance steps while the context is still fresh.
- Ship confidently — Move forward after the important checks are visible, reviewed, and included in the task.
Nexpath is one focused step inside a better shipping workflow — not a new place where development has to happen.
Built during AI Hackfest 2026 by MLH.
- Understand — Nexpath reviews the prompt, the current development stage, and relevant workflow signals while preserving the complete original request.
- Structure — It builds one editable prompt with the sections the task needs, such as scope, constraints, acceptance expectations, verification, or missing-practice guidance.
- Safeguard — Higher-risk work can receive confirmation, safety, rollback, or evidence requirements. Complex work can also receive a sequence-aware breakdown when multiple prompts would be more effective.
- Review — You inspect and edit the result before sending it, or return to the original prompt. Nexpath provides the quality layer; you keep the final decision.
Prompt quality is easy to demonstrate and hard to prove. To check whether Nexpath actually changes what a coding agent produces, we ran it against a public benchmark where the answer is graded by each project's own test suite — not by a person, and not by another model.
Setup: Claude Code (sonnet) attempted 40 tasks from SWE-bench Verified twice — once on its
own, once with the Nexpath-reviewed prompt. The task, starting commit, container, model, time
limit, and tools were identical in both runs. The prompt is the only thing that changed. Grading
used the official SWE-bench harness.
| Claude Code | Claude Code + Nexpath | |
|---|---|---|
| Tasks solved | 27/40 (67.5%) | 29/40 (72.5%) |
| Issue's own tests pass | 27/40 (67.5%) | 29/40 (72.5%) |
| Ran the project's tests before submitting | 38/40 (95.0%) | 39/40 (97.5%) |
Paired outcomes: solved by both — 27 · solved only with Nexpath — 2 · solved only without Nexpath — 0 · solved by neither — 11
Every number above traces to a file below. The full benchmark report (Nexpath_benchmark/README.md) has the method, the two tasks Nexpath won, and the eleven neither run could solve. These are the raw files behind it — all 40 tasks, both runs, and the list we locked before the first run.
| Complete report | Benchmark report |
| Every task, both runs | final_per_task.csv |
| All measurements | final_summary.json |
| Task list and settings, locked before the run | strat40-manifest.json |
| Prompt checks run before starting | preflight-strat40.json |
| Raw data for the confirmation-line test | before · after |
- Verification and test expectations can be added for debugging, maintenance, planning, review, and other tasks that need proof of completion.
- Absence signals can surface missing regression checks, acceptance criteria, reproduction evidence, or project grounding.
- Weak or unrelated signals are not used as filler; guidance must match the current task. The core interaction keeps your request and the added workflow guidance together:
- Sensitive actions can receive explicit confirmation, rollback, backup, or safety requirements before the reviewed prompt is used.
- Double confirmation, cross-confirmation, and final confirmation are used only when the task's complexity or risk justifies them.
- The quality layer does not silently expand authority, auto-send work, or treat an agent response as proof of completion.
- Scope, constraints, acceptance expectations, affected surfaces, and behaviour-preservation needs stay visible instead of being lost in a fast request.
- Complex work can be decomposed into a clear current task plus a compact, ordered sequence plan.
- Your complete original request stays visible inside one editable, quality-reviewed prompt.
- You can edit the result, use it, or return to the original request before the reviewed version is sent.
Nexpath CLI is built for prompt capture across AI coding agents.
| Agent | Status in v0.1.5 |
|---|---|
| Claude Code | Fully supported — end-to-end tested |
| Cursor | Fully supported — end-to-end tested |
| Windsurf | Fully supported — end-to-end tested |
| Replit | Fully supported — end-to-end tested |
| Lovable | Fully supported — end-to-end tested |
| Bolt.new | Fully supported — end-to-end tested |
# Clone and build from source
git clone https://github.com/hi0001234d/nexpath.git
cd nexpath
npm install
npm run build
npm link
# Register with your coding agent and verify
nexpath install
# Verify
nexpath --versionSetup notes:
- Install asks how Nexpath should run — one credential, two ways to give it: your own OpenAI API key (get one), or a Nexpath token from a free account at parseos.tech/nexpath if you would rather not create an OpenAI account. If both are configured, your OpenAI key is the one used.
- During install you pick your project role (what kind of work you do) so Nexpath tailors its guidance to how you build.
- Nexpath's core is Prompt Enhancement (PE), with Multi-Prompt Sequence (MPS) and Prompt Enhancement Feedback (PEF) — these work automatically as you code.
# Remove the Nexpath CLI
nexpath uninstall
npm uninstall -g nexpath
npm unlink -g nexpath
# Verify it's gone
npm list -g nexpath
which nexpath
# Clear local data and caches
rm -rf ~/.nexpath
rm -rf ~/.config/nexpath
rm -rf ~/.local/share/nexpath
rm -rf ~/.cache/nexpath
# Clear the npm cache
npm cache clean --forcenexpath uninstall disconnects Nexpath from all detected agents and offers to clear the stored
API key. The remaining steps remove the global package, any leftover binary, and all local
data and caches.
All data is stored locally only at ~/.nexpath/. Only targeted LLM calls used to classify or
prepare relevant guidance leave your machine.
- Automatic secret redaction — API keys (
sk-*,ghp_*,ghu_*), bearer tokens, and PEM blocks are automatically stripped from prompts before storage. - Telemetry is off unless you turn it on — install does not ask and does not enable it;
it starts off and stays off until you run
nexpath config set telemetry.enabled true. Local prompt capture and remote telemetry are independent — disable capture anytime vianexpath store disable(if you do this, nothing will work).
# Delete prompts for a specific project you no longer want to keep
nexpath store delete --project <path>
# Delete all stored prompts permanently
nexpath store delete -y| Platform | Default location | Inspect with |
|---|---|---|
| macOS | Keychain | Keychain Access.app → search "nexpath" |
| Linux | Secret Service (libsecret) | secret-tool lookup service nexpath account openai_api_key |
| Windows | Credential Manager | Control Panel → Credential Manager → Web Credentials |
| Fallback (any OS) | ~/.nexpath/config.json (mode 0600) |
cat ~/.nexpath/config.json |
Use nexpath config show-key-source to confirm which layer is currently active.
- Major League Hacking (MLH) — For organizing AI Hackfest 2026
- Anthropic — For Claude Code, our primary development environment
- OpenAI — For models used in targeted classification and prompt-quality tasks
- Google — For Gemini AI, planned as an alternative LLM provider alongside OpenAI
Built with insights from the vibe coding community and developers building real projects with AI coding agents, coding AI tools, and AI developer tools.