UX Heuristics Compass · Public beta
Turn screenshots and a focused task flow into a structured review: what looks difficult, what supports the finding, and what to check next.
AI-assisted heuristic review. Scores reflect the supplied evidence and rubric; they do not establish participant outcomes or accessibility compliance.
Interface feedback is hard to act on when the observation, severity, and proposed fix are mixed together. This tool gives that review a consistent structure.
I created UX Heuristics Compass: the audit workflow, evidence and scoring contracts, and shareable report output. Its rubric builds on Nielsen Norman Group’s ten heuristics with four additional profiles.
More work by Kyle HobsonPublic report examples, prompt guidance, source downloads, and the beta install guide. The examples show outputs and review methods; learning or usability improvements have not been measured here.
Project story / May–September 2026
The starting job was simple: help a builder review an interface before showing or shipping it. The difficult part was making an AI-assisted audit explicit about its scope, evidence, and uncertainty.
A May 2 beta exercise requested all 14 heuristics but omitted two optional profiles. The response was an explicit scope lock: show which profiles were scored, which were omitted, and when the review needs to run again.
Recorded implementation: scope controls and confirmation behavior. The public reports expose an “Audit Scope and Omitted Profiles” section.
Human answers initially needed manual translation back into the audit. Checkpointed resumption restored the source, scope, ratings, and review state. Compact status responses kept the conversation usable while full audit data stayed in local files.
Recorded implementation: resumable checkpoints and compact outputs, May 4. Both are present in the published beta source.
The report system gained readiness checks, a finding-linked Evidence Gallery, and clearer separation of AI evaluation from human research. When visual review cannot run, the tool reports that state instead of presenting a generated document as visually checked.
Recorded implementation: report validation, evidence traceability, and visual-review preflight. See the beta.10 release notes and case-study evidence record.
Using the tool on its own ecosystem
A May audit covered five public Compass pages and 23 page/tab states per platform. The review surfaced mobile clipping and crowded navigation that a default desktop screen would miss.
The archived comparison also revealed a reporting problem: “fixed” and “new” could describe differences between platforms, not improvements over time. That distinction informed clearer desktop/mobile language.
These are development and audit records. They do not demonstrate a measured improvement in user performance.
Explore the archived reports
A public beta with downloadable source, a desktop extension, a plugin, standalone guidance, and report examples. The six release files have been checked against their published checksums. The landing page has been reviewed on desktop and mobile; installation compatibility, adoption, and participant outcomes are separate questions.
UX Heuristics Compass is a local-first usability audit toolbox for websites, app screens, prototypes, screenshots, and short task flows. It helps a reviewer find what feels confusing, hidden, inconsistent, inaccessible, cognitively heavy, or hard to recover from, then ranks the fixes that matter most.
The beta is built around a 14-part heuristic system: Nielsen Norman Group's 10 usability heuristics plus accessibility and ease of access, empathetic engagement and inclusion, customer journey and satisfaction, and UX writing.
Not sure what to download?
Start with the host app you use most. You can install more than one format later if you want both the tool connection and the prompt-only fallback. MCP Basics (opens in a new tab).
Plugin Zip. Use this if you already work in Claude Code or Claude Cowork. This bundle includes the local MCP, skill, and host guidance.
Public beta
Download and Open GuideMCPB/Desktop Extension. Use this if you already have Claude Desktop installed. This connects Claude Desktop to the local UXHC tools.
Public beta
Download and Open GuideUse this when you want prompt guidance without the MCP tools, or when your host app cannot use local connectors yet.
Public beta
Download Skill Download Skill ZipDownloads are live for beta.10. Choose the plugin for Claude Cowork or Claude Code, the MCPB/Desktop Extension for Claude Desktop, or the standalone skill when tool access is unavailable.
| I want a fast first read | Use the quick audit prompt to review a screenshot, page capture, or small task flow. |
| I need the full checklist | Use the full audit prompt for scored 0-4 checklist ratings and a UX Health Snapshot. |
| I want the AI to ask questions | Use the advanced audit prompt with human-loop questions for context, calibration, and report readiness. |
| I want multiple perspectives | Use the multi-agent prompt for general UX, accessibility, and UX-writing perspectives. |
| I am testing implementation details | Use the implementation-aware prompt with screenshots, source notes, rendered previews, or browser observations. |
Copy a prompt into Claude, Codex, or another host AI. If native question UI is available, the host should use it for the context questions instead of asking in plain chat.
When to use: You want a prompt-only UX review before the MCP package is installed.
When to use: You have UXHC installed and want a compact audit.
When to use: You want scored findings across the active heuristic checklist.
When to use: You want context, calibration, and report-readiness questions built into the audit.
When to use: You want several UX perspectives without pretending there was a real user study.
When to use: You want help turning findings into practical fixes.
Kyle Hobson created UX Heuristics Compass to make usability reviews more concrete, structured, and useful. His work connects UX research, design, and applied AI. Explore Kyle’s portfolio.
UX Heuristics Compass is a local UX review assistant for visible interfaces. It helps you evaluate a website, app screen, prototype, screenshot, or short task flow before it ships, gets assigned, or gets shown to stakeholders.
The plain-language version: it looks for places where people are likely to get confused, miss the next step, struggle to recover from an error, lose trust, or need more effort than the task should require. It turns those observations into a scored UX Health Snapshot, a short issue ledger, and a prioritized fix list.
It does not pretend to replace real usability testing. It is the faster earlier pass: a structured way to catch likely usability problems, name evidence limits, and decide what should be fixed before you spend time recruiting users or polishing a report.
UXHC is built as a local-first MCP with deterministic heuristic data, checklist-level scoring, source preparation, bounded URL capture, report validation, local report artifacts, human-loop guidance, and agent-audit aggregation.
The host AI owns conversation, visual interpretation, and user-facing explanation. UXHC owns the rubric structure, scoring contracts, report readiness, source boundaries, local persistence, and artifact generation so the audit does not become generic design commentary.
The public beta also includes a standalone skill and plugin packaging. The skill can coach a lighter Markdown-only review when the MCP is unavailable; the MCP adds structured source prep, validation, persistence, report generation, and multi-agent aggregation.
The beta includes 14 heuristic records and 204 possible desktop/mobile checklist items. The core 10 heuristics come from Nielsen Norman Group's usability heuristic framework. UXHC adds four optional general profiles for accessibility, inclusion, customer journey, and UX writing.
This is an early beta. The core audit contract is stable enough to present publicly, but some tool wording and output shapes may still change as local testing tightens the package.
The current MCP surface is listed below so host AIs and technical reviewers can understand what the package exposes.
get_startedux_heuristic_overviewlist_heuristicsprepare_ux_sourcequick_scan_uxaudit_ux_surfacevalidate_ux_report_payloadgenerate_ux_audit_reportcalibrate_ux_sensitivitycalibrate_ux_severitycompare_ux_auditsdraft_project_addon_profilesave_preferencesupdate_preferencesrun_agent_auditget_agent_runrun_usability_teststress_test_heuristic_findingStart with a bounded source: one screenshot, one app screen, a small page capture, or a short named task flow. Tell the AI what the user is trying to do, what platform matters, and whether you want the optional profiles included. UXHC works best when the evidence is specific and the task is clear.
For a fast pass, use the quick audit prompt and ask for the top findings only. For a fuller review, use the checklist audit so the AI has to separate observed evidence from inference and return scored issues. For more sensitive or ambiguous work, use the advanced human-loop prompt so the host asks context and calibration questions before finalizing severity.
When you need a shareable deliverable, use the MCP flow: prepare the source, run the audit, validate the report payload, then generate the report. The Example tab shows the intended shape: a plain-language readout, a scorecard, prioritized fixes, and evidence limits rather than a wall of generic advice.
Choose the install path for the Claude surface you use. These screenshots show the tested UI. If your app does not show this path, stop and use the diagnostic prompt instead of improvising.
The screenshots below reflect the Claude Code, Claude Cowork, and Claude Desktop surfaces available during testing. Claude's install UI can differ by platform, account, and version. Local MCP/plugin and Desktop extension installs are not claimed for Claude.ai in a web browser.
Do not hand-edit Claude config, or let an assistant rewrite it, unless it first backs up the file, reads the existing server names, merges only ux-heuristic-compass, writes UTF-8 without a BOM, and verifies that no existing MCPs disappeared.
ux-heuristic-compass-plugin.zip.get_started or use the first prompt on the Get Started tab.If you only see Skills and Connectors, and Connectors only accepts a remote MCP URL, stop. That is not the plugin upload path.
ux-heuristic-compass-0.1.0.mcpb.If your Claude Desktop build does not show Settings → Extensions, do not hand-edit config as a beginner workaround. Use the diagnostic prompt.
ux-heuristic-compass.skill or skill.zip.Use the skill when the MCP/plugin path is unavailable, or keep it as a prompt-only fallback alongside the tool connection.
get_started or use the Skill Audit prompt.Do not improvise in Connectors unless that surface clearly accepts the exact file type you downloaded. Use the skill/reference path or the diagnostic prompt instead.
Restore the latest config backup or remove only the ux-heuristic-compass entry from mcpServers. Never reset the whole config to fix one install.
Technical users can download ux-heuristic-compass-source.zip, inspect the source, run the self-test, and connect the MCP manually. This is a fallback for review, development, or local troubleshooting, not the beginner install path.
Use the diagnostic prompt in the Advanced tab before editing files manually. Most failed installs come from an unavailable package, a host app that has not reloaded the MCP, or a path that needs local-machine adjustment.
Explore previously published reports and their supporting captures. These are archived review outputs, not current assessments of the sites. AI scores and researcher personas are evaluation aids, not participant-study results.
This public-safe example shows UX Heuristics Compass turning bounded page evidence into a scored heuristic audit, a plain-language readout, prioritized findings, evidence limits, and recommended next validation steps.
Example output B- / 68.5%
This example shows how UX Heuristics Compass turns a provided source into a ready-to-read audit report. It does more than list UI issues. It prioritizes a reported progress-persistence issue and separates the report’s observations from what still needs validation.
| Top finding | The archived report flags progress persistence as its top concern; this is not a current verification of all accounts or flows. |
| Primary themes | Progress persistence, lesson loading feedback, visible exit path, and screen-reader loading state. |
| Evidence limit | The audit is bounded to selected pages and is not a full WCAG compliance test. |
These snapshots show the report-style output: a scorecard with the overall grade, then a compact plain-language read that turns the audit into a usable decision point.
These are Claude Code conversation screenshots from the audit flow. Claude chat and Claude for Work can produce similar results, with some variation based on the model, host surface, available tools, and evidence supplied.
This bonus example shows a follow-up synthesis where Claude combines UXHC results with a human evaluation panel and another AI evaluation. This is a demonstration of what Claude can do with the audit output afterward; it is not an inherent MCP report function at this stage.
This example shows UX Heuristics Compass evaluating the public Compass Suite surface across multiple audit sources. It includes a merged Claude + Codex read, separate desktop and mobile reports, and a desktop/mobile comparison so the reader can see where the experience holds up and where mobile stress changes the priority order.
Example output 3-source suite audit
This example shows how UXHC can move from a single product-page audit into a cross-surface public-site review. It is useful for seeing how separate evaluations can be compared, merged, and turned into a remediation roadmap.
These snapshots are captured directly from the HTML reports. They are meant to show the shape of the finished artifacts before you open the full examples.
Use the buttons below to preview each report inside this tab. The new-tab links above remain the most reliable way to read or share a full report.
This example shows UX Heuristics Compass running a logged-in learner audit across a set of critical Oppia user journeys. The AI was given the journey set, then moved through each flow on its own to gather evidence, score the experience, and surface cross-journey patterns.
Example output 9 critical journeys B / 73.39%
This example shows how UXHC can move from a single-screen review into a multi-journey audit. It is especially useful for seeing how an AI assistant can inspect authenticated flows, compare journey-level failures, and then support a follow-up synthesis that combines AI findings with a human heuristic review.
| Main audit | Logged-in learner UXHC audit across 9 critical journeys plus a new dashboard preview. |
| Human + AI output | A follow-up synthesis that merges the AI journey audit with a human evaluator's heuristic review. |
| Primary lesson | Multi-journey audits reveal system-level blockers that a single-page review can miss. |
These snapshots show the two finished report surfaces: first the AI-led logged-in learner audit, then the Human + AI synthesis output that incorporates a human evaluator's heuristic read.
These stitched Claude Code screenshots show the multi-journey workflow in motion: the prompt, journey setup, evidence gathering, user-journey traversal, report generation, and Human + AI synthesis follow-up.
Quick answers about UXHC, heuristic evaluation, MCPs, and install expectations.
It helps an AI assistant review a visible interface with a structured UX checklist instead of giving generic design commentary. It returns scoped findings, severity, evidence limits, and fixes to prioritize.
UX means user experience: what it feels like for someone to use a product, service, website, or app. Usability is the part of UX focused on whether people can understand, navigate, complete tasks, recover from mistakes, and do what they came to do without unnecessary friction.
A heuristic evaluation is a structured expert review. Instead of waiting for a full research cycle, a reviewer checks an interface against known usability principles and flags likely problems. Nielsen Norman Group's 10 usability heuristics are one of the most widely used sets of principles for this kind of review. UXHC uses those 10 as the core, then adds four optional profiles: accessibility and ease of access, empathetic engagement and inclusion, customer journey and satisfaction, and UX writing.
It can use implementation evidence when a host AI provides it, such as rendered previews, source notes, browser observations, or external audit results. UXHC does not replace code review or accessibility testing, and implementation evidence must still be mapped back to visible usability findings.
MCP means Model Context Protocol. In this context, Claude may call the local tool connection an MCP, connector, or extension. UXHC's extension is the MCP connector: it gives Claude structured local tools for preparing sources, running audits, validating reports, and generating artifacts. Learn more about MCPs.
Plugin zip: the Claude Cowork or Claude Code bundle. Extension file: the Claude Desktop MCP connector. Skill: the prompt-only alternative a host can read without the MCP. You can install more than one when the release assets are available.
Yes. Codex can work with local MCPs, source installs, and prompt guidance. This beta.10 landing page does not provide a one-click Codex install path yet; technical Codex users should wait for the source package and AI-assisted install prompt.
Downloading anything from the internet carries real risk: malicious code, data collection, hidden scripts. UXHC is designed to reduce that surface area: it runs locally, uses bounded files, and does not need an external service for core workflows. No third-party tool is zero risk, so review the source or ask a trusted technical person before installing. Claude's warning is the standard prompt it shows for local third-party extensions, not a flag specific to this tool. Learn more about MCP risks.
No. Once the release assets are published, the buttons will go directly to the files. GitHub is only where the files are hosted.
This section is for IT reviewers, security-conscious users, and anyone troubleshooting installation.
Use this zip if you want to inspect, modify, or rebuild UX Heuristics Compass yourself. It excludes private state, generated release files, env files, caches, and machine-specific artifacts.
These SHA-256 checksums help technical reviewers confirm that downloaded files match this release.
ux-heuristic-compass-plugin.zip | 927649bd3e152b7551e3dff935cc4a62d8a54d5ff28c7946a025324490ab3567 |
ux-heuristic-compass-0.1.0.mcpb | 6892f583afaeb5e39a2c7640327264eeca9b908f89fd17155cef1b574119b04d |
ux-heuristic-compass.skill | 18507c693379eefb37ac0fdb89224d11c93bd4a70573e39026076137466872d6 |
skill.zip | 83a885597a48ee44665bb726648720c856cc49f1adb831d234a785a3ef7bcf93 |
ux-heuristic-compass-source.zip | 33a97cb4d74a19287f04eb321bf37e49a21ee68407c927589021e5dc8d1c183a |
checksums.txt | SHA-256 manifest |
Use this prompt when you want a careful AI assistant to help inspect and install the package.