Directors, producers, and planners ask AI to do their work every day. But they end up not getting the output they actually want — spending time on settings and corrections, and doing the work by hand anyway.
Where the bottleneck is:
- Lack of ability to articulate and structure their own workflow for AI
- AI behavior is a black box — no idea how to fix it when it goes wrong
- Every correction means starting the instructions over from scratch — time just disappears
- No habit of writing down "how I work" for AI — and often can't even articulate it themselves
Aun is a product where Claude Opus 4.7 runs four modes:
| Mode | What happens |
|---|---|
| Interview | Just talk about your work by chat or voice. Aun structures your intent, deliverable, and steps into an executable brief. |
| Visualize | The brief becomes a node-flow diagram. Review the full pipeline and edit any step. |
| Build (first time only) | Writes your style, judgment criteria, and templates as a Skill (.md). This becomes the foundation of your personal AI. |
| Production | A Managed Agent runs autonomously overnight — injecting your Skill automatically — and generates PPT, Word, Excel, or Markdown. |
| Reflection | When you refine and save, Opus 4.7 reads the why behind your edits and rebuilds the Skill from scratch. The more you use it, the more it becomes yours. |
Browser
→ POST /api/interview ── Opus 4.7 structures intent → brief
→ POST /api/run ── Managed Agent session created → session_id returned instantly (< 5s)
→ GET /api/status ── polling every 10s until idle → artifact download
→ POST /api/skill ── rebuilds Skill from edit history
→ POST /api/history ── records Run to Blob → appears in Library screen
One Opus 4.7. Four modes via system-prompt switching. Zero multi-agent complexity.
| Framework | Next.js 15 (App Router, Node.js runtime required) |
| AI Model | Claude Opus 4.7 (all 4 modes) / Claude Haiku 4.5 (voice formatting, lightweight tasks) |
| Long-running execution | Claude Managed Agents (beta: managed-agents-2026-04-01) |
| Output generation | Code Execution Tool (python-pptx / python-docx / openpyxl) |
| File attachments | Anthropic Files API (beta: files-api-2025-04-14) |
| Skill format | Claude Agent Skills (portable .md persona) |
| Persistence | Vercel Blob (Skill / History) |
| Frontend | TypeScript strict + Zod + Tailwind CSS v4 |
git clone https://github.com/peco-glhf/aun
cd aun
pnpm install
cp env.example .env.local
pnpm devOpen http://localhost:3000 in your browser.
Environment variables (see env.example for details):
| Variable | Required | Description |
|---|---|---|
ANTHROPIC_API_KEY |
✅ | Anthropic API key |
AGENT_ID |
optional | Pre-created Managed Agent ID (auto-created on first run if blank) |
ENV_ID |
optional | Pre-created Managed Environment ID |
BLOB_READ_WRITE_TOKEN |
optional, not recommended | Vercel Blob token. Leave unset for the public demo. Blob persistence only works when ENABLE_PUBLIC_BLOB_STORAGE=true is also set. |
ENABLE_PUBLIC_BLOB_STORAGE |
optional, not recommended | Explicit opt-in for public Blob persistence. Do not enable for shared demos. |
- Demo Video: https://youtu.be/9BT94WgQP_Y
Security note for the live demo
The live demo is protected by Vercel Deployment Protection. Aun is a single-tenant hackathon MVP and does not include built-in authentication or rate limiting.
For deployment, set only
ANTHROPIC_API_KEY. Do not setBLOB_READ_WRITE_TOKENorENABLE_PUBLIC_BLOB_STORAGEfor the public demo. Persistence is in-memory by default. Vercel Blob persistence is opt-in and not recommended for shared demos.
Opus 4.7 handles all modes through system-prompt switching alone — no fine-tuning, no extra tooling. With adaptive thinking, it reads between the lines in Interview mode and structurally integrates the why behind edits in Reflection mode.
Managed Agents are the engine of Production mode.
/api/run creates a session and returns session_id in under 5 seconds.
The agent runs autonomously on Anthropic's servers, using code_execution to save output files to /mnt/session/outputs/.
The browser polls /api/status and downloads artifacts via the Files API once complete.
The session continues even after the browser is closed. This is what makes "submit at night, receive in the morning" possible.
MIT © 2026 Ryu Inoue — see LICENSE