![\"Pantion\"](\"https://raw.githubusercontent.com/pantion-dialogs/pantion-dialogs/main/.github/images/manakin.png\")
\n\n \n
Agent coding is fast. Your intent is the bottleneck.
\n\nToday's coding agents can outcode humans on speed and execution.\n\nBut great software isn't just code. It's **decisions**: defaults, edge cases, permissions, failure behavior, data boundaries, trade-offs. Agents guess those decisions silently — and when they guess wrong, you end up in correction loops.\n\nThe Pantion Dialog turns those decisions into a **Canon** — a converged dialog with no ambiguity left — before your agent writes a single line. And not just for software. The same dialog works for images, video, and any domain where AI generates from your intent.\n\n### The five-second version\n\n> **You:** \"I want a booking app.\"\n>\n> **Pantion:** \"What happens on double booking?\"\n>\n> **You:** \"First wins. Second gets alternatives.\"\n>\n> **Pantion:** \"Locked. That decision won't be guessed later.\"\n\n## Quick start\n\nPantion works best with **Claude Code**. It also installs on Cursor, Windsurf, Gemini CLI, and Codex — experience varies by model.\n\n**1. Install**\n\n```bash\nnpx @pantion/dialogs install --client claude\n```\n\nOr install for all supported clients at once:\n\n```bash\nnpx @pantion/dialogs install --all\n```\n\nSupported clients: `claude`, `cursor`, `windsurf`, `gemini`, `codex`\n\n**2. Start a dialog**\n\nDescribe what you want to build in natural language. Your agent calls Pantion automatically.\n\n> I want to start a Pantion dialog for a CLI tool that converts markdown files to a static website.\n\nPantion asks targeted questions until your intent has converged, then saves the canon. That's it.\n\n## See it in action: automatic mode\n\nDon't want to answer questions? In **automatic mode**, Pantion generates the entire dialog for you — you just describe your idea and review the result.\n\n> I want to start an automatic Pantion dialog for a web alarm clock with multiple timers\n\nPantion generates a full convergence dialog in seconds:\n\n```\nASSISTANT: What is the core intent?\nPANTION: A browser-based alarm clock with multiple independent timers ⚡\n\nASSISTANT: What are the inputs?\nPANTION: Timer name, duration (HH:MM:SS or MM:SS), optional label ⚡\n\nASSISTANT: What happens when a timer finishes?\nPANTION: Audio alert + visual flash. Audio stops on click/tap. ⚡\n\nASSISTANT: How many timers can run simultaneously?\nPANTION: No hard limit. UI scrolls. ⚡ Performance is the natural cap.\n\nASSISTANT: What does this NOT do?\nPANTION: No accounts, no server, no persistence across browser sessions ⚡\n```\n\nThe `⚡` marks are conservative assumptions — decisions Pantion made for you. Review, adjust what you disagree with, and the canon is ready.\n\nTwo modes:\n\n| Mode | How to start | Best for |\n|------|-------------|----------|\n| **Normal** | \"I want to start a Pantion dialog for...\" | Complex systems, precise control |\n| **Automatic** | \"I want to start an automatic Pantion dialog for...\" | Quick exploration, simple projects |\n\n## How it works\n\nPantion is an MCP server. Your coding agent connects to it, and it guides the conversation — asking one question at a time until there's nothing meaningful left to ask. That's convergence.\n\n```\nIdea → Dialog → Canon → Spec files → Build\n```\n\n1. **Dialog** — Pantion asks targeted questions to remove ambiguity. When new questions stop yielding new behavior, your intent has converged.\n2. **Canon** — The verbatim conversation becomes the source of truth. Decisions, rejected alternatives, rationale — all preserved.\n3. **Build** — Your coding agent reads the spec files or the canon directly via MCP resources. Point your agent at `canon/\n ![\"Pantion](\"https://raw.githubusercontent.com/pantion-dialogs/pantion-dialogs/main/.github/images/manakin-front.png\")
\n Agent coding is here to stay. The Pantion Dialog makes it reliable.\n