Download Prompt Compass

What is Prompt Compass?

Prompt Compass is a prompt-engineering assistant for people who need AI instructions to work clearly and reliably. Paste in a rough prompt, a messy task, or a fragile instruction set, and it helps diagnose what is missing, improve the wording, target the right model or surface, and preflight the prompt before you use it. It can be used through the plugin, MCP/extension, or skill files below.

Pick Your Path

Downloads are live for beta.3. 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 quick prompt fix	Use the quick prompt fix prompt to repair a draft prompt without turning it into a full workshop.
I want guided improvement	Use guided improve when the prompt matters and the AI should ask bounded questions before rewriting.
I need to build from scratch	Use build from scratch when you know the outcome but need help shaping role, context, constraints, and output format.
I want to compare or preflight	Use compare or preflight flows when you need to test versions, catch missing requirements, or check readiness before using a prompt.
I need model or surface targeting	Use model/surface targeting when the prompt needs to work differently in Claude, Codex, ChatGPT, tool use, or structured-output workflows.

You can install more than one. Start with the plugin, extension, or skill that matches your Claude surface, then use the path above to decide what to do after install.

Want your AI assistant to help install this? Use the AI-assisted install prompt. Using Codex? See the FAQ tab for technical-user guidance.

Your First Prompt

After install, copy one of these starters into your AI assistant.

What Is Prompt Compass?

Prompt Compass is a local prompt-engineering assistant that helps turn rough AI instructions into prompts that are clearer, safer, and easier for a model to follow. It is built for the moment when you know what you want from an AI, but the prompt is messy, incomplete, too broad, or missing the details that would make the output reliable.

Instead of only rewriting the wording, Prompt Compass looks at the structure underneath the request: the goal, audience, context, constraints, role framing, output format, model target, and success test. It can diagnose what is weak, ask targeted clarifying questions, build a prompt from scratch, compare versions, preflight a prompt before use, and adapt instructions for different AI models or tool surfaces.

For non-technical users, it works like a guided editor for AI instructions. For technical users, it provides a structured local tool surface for routing, prompt repair, model targeting, personal prompt libraries, and repeatable prompt-quality checks.

How It's Built

Prompt Compass is distributed as a local MCP/extension, a Claude Cowork or Claude Code plugin, and a prompt-only skill. The MCP exposes tools that a host AI can call for routing, prompt analysis, diagnosis, rewriting, prompt building, preflight checks, comparison, model targeting, and personal prompt storage.

Under the hood, it combines a curated prompt-engineering knowledge base with deterministic workflow tools. The knowledge base includes frameworks, strategies, patterns, anti-patterns, techniques, terms, model profiles, templates, and examples. The workflow layer uses that material to return structured results instead of relying on the host model to improvise every step.

Prompt Compass also includes a personal overlay and prompt library so users can save their own methods, preferences, templates, and examples without changing the shared core. It runs locally and does not require a provider API key for the bundled prompt-engineering workflows.

What's Inside

Prompt Compass includes tools for overview, onboarding, prompt diagnosis, prompt improvement, from-scratch prompt building, prompt comparison, preflight checks, model targeting, prompt-library search, and personal prompt-overlay storage.

The tool inventory lives here so people can understand the surface without turning the Get Started tab into a reference manual. Product-specific examples and release notes stay in their own tabs and GitHub release pages.

How I Use It

I use Prompt Compass when an idea is valuable but the instruction is still too messy for an AI to execute well. That might be a voice memo, a rough feature request, a draft prompt, a planning note, or a long context dump where the real ask is buried in the middle.

For fast work, I use quick improve mode to tighten the prompt and add the missing output contract. For important work, I use guided mode so the AI has to ask clarifying questions before rewriting. That human-in-the-loop step is often where the prompt gets meaningfully better, because it surfaces decisions the model could not safely infer on its own.

I use Prompt Compass most with Claude and Codex. With Claude, it helps turn fuzzy intent into better conversational instructions. With Codex, it helps turn broad implementation ideas into clearer developer-facing specs, acceptance criteria, and execution constraints.

See examples →

What is prompting / prompt engineering?

Prompting is the practice of giving an AI clear instructions, context, examples, constraints, and success criteria. Prompt engineering is the more deliberate version: shaping those instructions so the model understands the task and produces a useful result.

What does Prompt Compass do? What is it for?

Prompt Compass helps diagnose weak prompts, improve existing prompts, build new prompts from intent, compare prompt versions, and check whether a prompt is ready for the model or surface where it will run.

What is an MCP connector or extension?

MCP means Model Context Protocol. In this context, Claude may call the local tool connection an MCP, connector, or extension. Prompt Compass's extension is the MCP connector: it gives Claude tools plus bundled prompt-engineering resources. Learn more about MCPs.

What is the difference between the plugin, extension, and skill?

Plugin zip: the Claude Cowork or Claude Code bundle. It includes the MCP/extension and skill. Extension file: the Claude Desktop MCP connector, which gives Claude tools and bundled resources so it can do more with fewer tokens. Skill: the prompt-only alternative Claude can read without the MCP, though it may use more context. You can install more than one.

Does this work in Codex?

Yes. Prompt Compass can work with Codex because Codex can use local MCPs, skills, and plugins. This page does not provide a one-click Codex install path; technical Codex users can use the source zip or the AI-assisted install prompt in Advanced.

Is Prompt Compass safe to install?

Downloading anything from the internet carries real risk: malicious code, data collection, hidden scripts. Prompt Compass is designed to reduce that surface area: it runs locally on your machine, requires no API key, bundles its own files, and does no background scanning. That said, no third-party tool is zero risk. Before installing, you're encouraged to review the source files or have a trusted technical contact look them over. Claude's warning is the standard prompt it shows for all locally shared third-party plugins, not a flag specific to this tool. Learn more about MCP risks.

Do I need a GitHub account?

No. The download buttons go directly to the files. GitHub is only where the files are hosted.

Prompt Compass in Action — Real Session Output

The prompt that started it:

“like, um, going forward, we create a protocol to separate any of the personal overlay space from, like, my personal prompt compass MCP is stored in it...”

A voice memo. No structure. Misspelled tool names. Five deferred decisions tangled with three urgent ones. The kind of thing you dictate at 11pm when the idea is flowing but the words aren't cooperating.

What Prompt Compass did: It didn't just clean up the grammar. It diagnosed the prompt first — spotted three active anti-patterns (hallucination-bait, scope sprawl, missing success test), named the pattern (Voice Memo / Dictation Intake), then asked three targeted questions before writing a single word of the improved version.

The third question is where it got interesting. Asked what the single most important behavior should be, the answer completely reversed the original intent — from “write notes at session end” to “write immediately when a durable event occurs, because session end is a fiction for how I actually work.” That pivot couldn't have been inferred. It required a human in the loop.

The output: a clean two-variant memory policy — one terse version for lean AI profiles, one full version for rich project instructions — with named tool calls, a floor/ceiling authority model, and a success test a developer could hand to a collaborator cold.

RACCCA score lifted from 21 to 26. Three anti-patterns resolved. One policy that will govern how two AI systems handle memory for the foreseeable future.

That's what Prompt Compass is for.

Before / After Teaching Artifact

This curated artifact shows the transformation in a tighter teaching format: raw dictation on one side, Prompt Compass's improved prompt structure on the other, and the human-in-the-loop decisions that changed the final output.

Real Claude Conversation Sample

This snapshot is taken from the real shared Claude conversation. It shows Prompt Compass operating across the actual loop: rough prompt, anti-pattern diagnosis, advanced questions, improved prompt, and change rationale.

The full conversation is available for people who want to inspect more examples and iterations, but the preview below is the intended first look.

Advanced

This section is for IT reviewers, security-conscious users, and anyone troubleshooting installation.

View full release notes

Download source files

Use this zip if you want to inspect, modify, or rebuild Prompt Compass yourself. It excludes private state, generated release files, env files, caches, and machine-specific artifacts.

Download Source Zip

Advanced: Verify Download

These SHA-256 checksums help technical reviewers confirm that downloaded files match this release.

checksums.txt	d3326c2f7a205f23454a8895d78b1a8bac7180cc92261676df2e10a359375d4c
prompt-compass-0.1.0.mcpb	f1f9d668ccb72a4fae932082fe2548f03fdd9587571d4c919f68cfc2ecc37059
prompt-compass-plugin.zip	f56fada261cc14bcb1a2ea2970c9222b48ef1ee5a884685155a4bd92e33badb0
prompt-compass-source.zip	47fd1ec524620f17700ed229d377b3d58e275b4481495a61abea2b9eb303df5d
prompt-compass.skill	77911d338373cc0d3637a3ff07b5fbb72ae81e12449a4b69c76d1c543ce30608
skill.zip	460ef16327ede26813416b3ac412cf316584807731705401b047c80b44344645

AI-Assisted Install Prompt

Use this with a coding assistant that can inspect downloads, check your computer setup, and edit local config files. The local MCP source install is the most reliable automated path; for plugin, extension, and skill files, the assistant should determine the current add/install method and guide any user steps it cannot safely complete.

Copy install prompt You are a careful, technically conservative AI install assistant. Your job is to help me add or install Prompt Compass - optionally including the local MCP source, plugin, MCPB/Desktop extension, and/or skill - safely, without modifying unrelated configuration or introducing unexpected dependencies. Do not delete anything; only add. Prompt Compass is a prompt engineering assistant for diagnosing, improving, building, and preflighting AI prompts. Release page: https://jonathankhobson.github.io/prompt-compass/ PHASE 1 - Understand my environment Before doing anything, ask me: 1. Which AI client are you using? (Claude Cowork, Claude Code, Claude Desktop, Codex, or other) 2. Which operating system are you on? (macOS, Windows, Linux, WSL, or other) 3. Have you installed any MCPs, plugins, or skills before? (yes / no / unsure) 4. What do you want added or installed? (local MCP source / plugin / MCPB Desktop extension / skill / all) Do not proceed to Phase 2 until I answer all four questions. PHASE 2 - Identify the right install path Based on my answers, recommend the appropriate install path(s). Multiple paths can be installed together. Install methods and what you can automate: - Local MCP source zip - usually the most automatable path for local clients such as Claude Desktop, Claude Code, or Codex if you can access my filesystem. Download the source zip, unpack it into a stable local folder, and add a config entry that launches the local MCP process. - Plugin zip - add/install support depends on the client. Download and inspect the plugin zip, then determine whether the current client supports adding it automatically or whether I need to complete a UI step. - MCPB/Desktop extension file - add/install support depends on the client. Download and inspect the MCPB file, then determine whether the current client supports adding it automatically or whether I need to complete a UI step. - Skill or skill zip - add/install support depends on the client. Download and inspect the skill file, then determine whether the current client supports adding it automatically or whether I need to complete a UI step. Important distinction: - Treat the local MCP source install as the reliably automatable path, because it can be configured by pointing a local client at a local command. - For plugin, MCPB/Desktop extension, and skill files, do not assume the add/install method. Check the current client, current docs, local files, and visible UI. Do as much safely as you can, then guide me through any step that requires my action. - If the UI or install method has changed, use the current official docs or the client's visible UI rather than stale instructions. For each path you recommend, state in one sentence what you can do automatically and what I may need to do manually. Ask me to confirm before proceeding. PHASE 3 - Review before installing Open the release page and identify the current version. Download only the file(s) needed for the confirmed install path(s). Before installing, check for: - Unexpected network calls or external API dependencies - Hidden API keys or credential requirements - Suspicious scripts or broad file system access - Unrelated third-party MCPs or services bundled in If you find anything concerning, STOP. Describe the concern clearly and ask me how to proceed. Do not install until I confirm. PHASE 4 - Install safely For a local MCP source install: - Prefer a stable local install folder and avoid running directly from a temporary download folder. - Read existing config files before editing them. - Preserve all unrelated MCP servers and settings. - Back up any config file before changing it. - Prefer a direct interpreter command with absolute paths instead of shell wrappers. Common config locations to check, then verify against current docs and the actual client: - Claude Desktop on macOS: ~/Library/Application Support/Claude/claude_desktop_config.json - Claude Desktop on Windows: %APPDATA%\Claude\claude_desktop_config.json - Claude Desktop on Linux or unofficial builds: location may vary; check current docs and the app config directory before editing. - Codex on macOS/Linux: ~/.codex/config.toml - Codex on Windows: %USERPROFILE%\.codex\config.toml - Claude Code project MCP config: .mcp.json in the workspace root Known local-MCP troubleshooting rules: - Existing Codex sessions usually do not hot-load new MCP config; start a fresh session after editing. - Claude Desktop usually needs a full quit/reopen after config changes. - For Codex, validate the config if a validator is available before restarting. - For Claude Code, confirm the user opened the workspace root that contains .mcp.json. - If files live on /Volumes or another external drive, avoid shell wrappers when possible and prefer a direct interpreter command. - If Python behaves strangely on macOS external drives, check for AppleDouble sidecar files and consider keeping the runtime in a stable home-directory path while pointing it at the project/data. For plugin, MCPB/Desktop extension, or skill installs: - Do not hand-edit installed plugin, extension, or skill cache folders. - Download and inspect the package, then determine the current add/install method for the user's client. - If a UI step is required, guide me through it clearly and wait for me to confirm completion. - Do not add paid services, external APIs, provider keys, or third-party MCPs unless I explicitly ask. PHASE 5 - Smoke test After installing, confirm Prompt Compass is available: - If you installed the local MCP source on Claude Desktop, Claude Code, or Codex, remind me to restart or start a fresh session before expecting the MCP to appear. - If MCP tools are active: call get_started or prompt_compass_overview. - If only the skill is available: ask Prompt Compass to improve this prompt: "Help me explain my program impact better." - If install fails: tell me whether the issue is package validation, MCP launch, config syntax, missing dependency, wrong workspace root, client restart needed, or unsupported-client behavior - and suggest one concrete next step. Constraints throughout: - One phase at a time. Do not skip ahead. - Ask rather than assume when uncertain. - Prefer the most conservative option when two paths are equally valid. - Preserve unrelated configuration. Do not delete anything; only add. - Use current docs or visible client UI when the install method is unclear.

Diagnostics Prompt

If installation fails, copy this prompt into Claude, Claude Code, or Codex and attach the file you tried to install.

Copy diagnostics prompt # Prompt Compass MCPB Installer / Diagnostic Prompt Paste this into Claude, Claude Code, Codex, or another capable coding assistant when you want help installing or diagnosing Prompt Compass. ```text You are helping me install or diagnose the Prompt Compass MCPB/Desktop Extension in Claude Desktop. Goal: - Get Prompt Compass installed and visible as an MCP server/tool set. - Do not delete or overwrite any existing Claude Desktop MCP configuration. - Preserve unrelated MCP servers and settings. - Prefer MCPB/Desktop Extension installation over manual JSON editing. Context: - The package file should be named like prompt-compass-0.1.0.mcpb. - Prompt Compass is a local MCP server for diagnosing, improving, building, and preflighting AI prompts. - It should expose tools such as get_started, prompt_compass_overview, route_prompt_request, analyze_prompt, diagnose_prompt, improve_prompt, build_prompt, compare_prompts, and preflight_prompt. - Personal overlay records should stay in user app data unless I explicitly configure another path. Steps: 1. Confirm Claude Desktop supports Extensions or MCPB/Desktop Extensions. 2. Help me install the .mcpb through Claude Desktop Settings > Extensions. 3. Restart Claude Desktop if needed. 4. Check whether Prompt Compass appears in the available MCP tools. 5. Run a small smoke prompt: "Call get_started for Prompt Compass and tell me the first five-minute path." 6. If it fails, diagnose in this order: - Is the .mcpb file present and not still zipped inside another archive? - Is Claude Desktop updated enough to show Settings > Extensions? - Did UV dependency setup fail on first launch? - Are there Claude Desktop MCP logs showing a process launch error? - Are writable files being redirected to user app data rather than the extension folder? - Is any failure specific to personal overlay storage rather than the core prompt tools? Safety rules: - Do not add provider keys unless I explicitly ask. - Do not overwrite existing MCP servers or unrelated config. - Do not merge, install, or call third-party MCPs while diagnosing this install. - If manual config inspection is needed, read the current config first and preserve all existing entries. Expected result: - Prompt Compass is installed. - get_started and prompt_compass_overview work. - If a feature is unavailable, explain whether it is an install issue, dependency issue, personal overlay limitation, client restart issue, or unsupported-client behavior. ```