An AI-powered personal growth system for people who think more than they act.
MindOS is a self-contained HTML widget that runs as a Claude.ai artifact or in any modern browser. It turns AI into a thinking partner that helps you capture thoughts, discover patterns, track growth, plan tasks, and get AI supervision — all in one widget.
Works for anyone who thinks more than they act — students preparing for exams, designers building a portfolio, researchers writing papers, professionals tracking career direction, or anyone who wants a quieter way to make sense of their own mind.
01 CAPTURE → 02 UNDERSTAND → 03 GROWTH → 04 REVIEW
记录 理解 成长 回顾
05 SCHEDULE → 06 ACTION → 07 COACH → 08 CONFIG
日程 行动 教练 设置
| Tab | What it does |
|---|---|
| 01 · Capture | Zero-friction thought dump. No categories, no filters. Just write. Energy level tracking with a 5-point mood system. Voice input supported (Web Speech API / Whisper fallback). |
| 02 · Understand | AI reads all your captures and surfaces hidden patterns: recurring themes, emotional trends, things you didn't realize you were thinking about. |
| 03 · Growth | A "growth mirror" — AI generates a narrative of how you've changed, what's rising in your attention, what's fading. Changes you can't see yourself. |
| 04 · Review | AI-generated periodic retrospectives: daily / weekly / monthly / yearly. Monthly uses weekly summaries as input; yearly uses monthly summaries — a hybrid hierarchy that keeps token cost down without losing detail. |
| 05 · Schedule | A lightweight task list. Add tasks manually, set due dates, filter by today/week/done/all, or import next-steps from your latest weekly review with one click. |
| 06 · Action | Not a to-do list. One thing. The single most-worthy action today, connected to who you're becoming — with a WHY. |
| 07 · Coach | AI supervisor. Reads your task completion rate and delivers one of three tones: encourage (≥70% done), push (30–70%), or critique (<30%). Not a cheerleader — a honest mirror. |
| 08 · Config | Pick AI provider, paste API key, set model/base_url overrides, fill in your user context, toggle dark mode. |
Most productivity tools optimize for doing more. MindOS optimizes for understanding yourself.
The eight tabs follow a single direction: chaos → clarity → action → supervision.
- Layer 1 is for the brain that generates more ideas than it can execute
- Layer 2 is for the patterns you can't see when you're inside them
- Layer 3 is for the growth that's invisible without a long enough lens
- Layer 4 turns scattered days into structured retrospectives
- Layer 5 is the bridge between thinking and doing — concrete tasks
- Layer 6 is for finding the one action that matters today
- Layer 7 is the friend who tells you the truth about your follow-through
- Layer 8 is for tuning the system to your provider and context
- claude.ai account (free tier works)
- Settings → Profile → Enable "Create AI-powered artifacts" (Beta feature)
- An API key from one of the supported AI providers (see below). 智谱 GLM-4-Flash 用户无需付费即可使用全部功能 — register at
open.bigmodel.cnand create a free API key. - For voice input: Chrome or Edge browser (uses Web Speech API). Firefox falls back to Whisper API if OpenAI key is configured.
MindOS speaks the OpenAI Chat Completions protocol (/chat/completions) and supports four providers out of the box. Pick one in the 设置 (Config) tab inside the widget.
| Provider | base_url | Default model | Cost | Key required |
|---|---|---|---|---|
| DeepSeek (default) | https://api.deepseek.com/v1 |
deepseek-chat |
~¥0.27 / 1M input tokens | Yes |
| 智谱 GLM-4-Flash | https://open.bigmodel.cn/api/paas/v4 |
glm-4-flash |
Free | Yes (free tier) |
| Claude | https://api.anthropic.com/v1 |
claude-sonnet-4-5-20250929 |
Pay-as-you-go | Yes |
| OpenAI | https://api.openai.com/v1 |
gpt-4o-mini |
Pay-as-you-go | Yes |
DeepSeek's base_url is prefilled by default; switching providers in the dropdown updates base_url and model automatically.
The API key lives only in the browser's localStorage and is never written to window.storage. This is a deliberate security boundary: window.storage (claude.ai's artifact-scoped storage) syncs across sessions and would upload the key to the server, so the widget excludes apiKey from any window.storage write — even when the rest of the config object is synced.
Non-sensitive config (provider, model, base_url, userCtx, theme) is dual-written to both localStorage and window.storage, so your provider preference survives across sessions without exposing the key.
The provider registry lives at the top of the <script> in widget.md:
const PROVIDERS = {
deepseek: { label:'DeepSeek', base_url:'https://api.deepseek.com/v1', model:'deepseek-chat', ... },
// add a new entry here…
};Append an entry to PROVIDERS and an <option> to the dropdown in the config view. As long as the provider exposes an OpenAI-compatible /chat/completions endpoint, it will work without further changes. The architecture mirrors CaseForge's config.example.py — a single registry with base_url + model per entry, resolved at call time.
The microphone button next to the capture textarea uses the browser-native Web Speech API (Chrome/Edge) for free, on-device speech-to-text. No API key needed. Interim results stream live into the textarea; the recognizer auto-restarts on silence to keep recording continuous until you click stop.
If the browser does not support Web Speech API (e.g. Firefox), the widget falls back to OpenAI Whisper API — but only when the user has selected the openai provider and entered their API key in the Config tab. Audio is sent directly from the browser to OpenAI; nothing is routed through any other server.
- Open a new Claude conversation
- Paste the contents of
widget.mdinto the prompt - Ask Claude to render it as an artifact
- Open the 设置 tab, pick a provider, paste your API key, and start capturing
# Clone the repo, then serve widget.md as an HTML test page
git clone <repo-url> mindos-skill
cd mindos-skill
# Extract the HTML block from widget.md and wrap with a test harness,
# then serve with any static server, e.g.:
python -m http.server 8766A Node.js build script (build_mindos_test.js) is included in development setups to extract the HTML block from widget.md and wrap it with a mock window.storage harness for local browser testing.
Open the 设置 (Config) tab inside the widget and fill in the 用户背景 (User Context) field with a one-line description of your situation and goals, e.g.:
- A student preparing for exams:
正在备考的学生，主攻方向是 XX - A designer building a portfolio:
设计师，正在准备作品集 - A researcher writing a thesis:
正在写毕业论文，研究方向是 XX - A general user: leave blank — the AI will infer from your captures
When filled, this line is appended to each AI system prompt (Understand / Growth / Action / Review / Coach) at call time. When blank, no context is added — the system works fine without it.
Toggle the LIGHT/DARK button in the top bar. Your preference is saved to window.storage under mindos:cfg:theme and persists across sessions. The dark theme uses the same --ms-* CSS variables as the light theme, just with different values — no separate stylesheet.
The widget defaults to Chinese (zh-CN). To switch to English:
- Change all
toLocaleDateString('zh-CN', ...)calls - Update placeholder text in the textarea
- Replace
MNmood label array with English equivalents - Update all Chinese UI strings
The design uses:
- Crimson Pro (all AI-generated content, in italic — insights, reviews, coach messages, action text)
- Instrument Sans (body text, inputs, UI labels)
- DM Mono (metadata, timestamps, labels, eyebrows)
Mood colors are hardcoded hex values in the CSS (.m1–.m5). Change these to match your aesthetic.
MindOS uses two storage layers with a strict security boundary:
window.storage (artifact-scoped, syncs across sessions):
mindos:cfg → non-sensitive config object {provider, model, baseUrl, userCtx} (apiKey always empty)
mindos:cfg:theme → 'light' | 'dark'
mindos:captures → thought entries [{id, text, mood, ts}]
mindos:cache → AI analysis cache (avoids redundant calls)
mindos:actions → daily action history [{text, why, date, done}]
mindos:reviews → periodic retrospectives keyed by period + revKey
mindos:tasks → schedule tasks [{id, text, dueDate, createdAt, done, doneAt, source}]
mindos:coach → coach history [{tone, message, ts, completionRate}]
localStorage (browser-only, NEVER sent to any server except the chosen AI provider):
mindos:cfg:apiKey → API key (plaintext, lives only in the user's browser)
mindos:cfg:provider / model / baseUrl / userCtx → also mirrored here as fallback
mindos:cfg:theme → theme preference (mirrored)
The API key is intentionally excluded from window.storage writes. When saveSetting writes the config object to window.storage, it replaces apiKey with an empty string first. When loadSettings reads from window.storage, it overrides apiKey with the value from localStorage. This guarantees the key never leaves the browser even if other config syncs across sessions.
Cached AI results stay until you manually re-trigger analysis. To fully reset, clear both window.storage keys (prefix mindos:) and the mindos:cfg:* keys in localStorage.
- Multi-provider AI (DeepSeek / 智谱 / Claude / OpenAI)
- Dark mode toggle
- Weekly / monthly / yearly AI digest (Review tab)
- Schedule / task management with import from weekly review
- AI coach with tone-based supervision (encourage / push / critique)
- Voice input (Web Speech API + Whisper fallback)
- JSON export / import
- Tag system for captures
- Timeline view
- Multiple life areas / projects
- English UI mode
- Obsidian / Notion export
mindos-skill/
├── README.md — This file
├── widget.md — Complete widget implementation (HTML/CSS/JS, single self-contained file)
├── LICENSE — CC-BY-NC 4.0
├── .gitignore
└── docs/
├── devlog.md — Development journal (design decisions, pitfalls, demo screenshots)
└── devlog-assets/ — 8 demo screenshots referenced by devlog.md
The widget is a single self-contained HTML file inside a markdown code block. No build step. No npm. No framework. Just paste into Claude and render as an artifact.
For the full development story — architecture decisions, pitfalls, and screenshots of all 8 tabs in action — see docs/devlog.md.
Why magazine typography? AI-generated insights use Crimson Pro in italic — creating the sensation of reading a magazine about yourself. This is intentional: it signals that the AI's output is a different kind of object than your raw input.
Why one action? Productivity systems fail by optimizing for quantity. One action, with a reason connected to your long-term self, is harder to ignore than a list of three.
Why no forced categories in Capture? Because categorization is a form of pre-judgment. The system learns your categories from the patterns — not the other way around.
Why a coach instead of a habit tracker? Habit trackers let you lie to yourself. The coach reads your actual completion rate and responds with the tone you've earned — encouragement when you're following through, a push when you're slipping, critique when you're avoiding the work. No badges, no streaks, just honesty.
Why hybrid review hierarchy? Monthly reviews read weekly summaries as input; yearly reviews read monthly summaries. This avoids feeding the AI hundreds of raw captures at year-end (expensive and lossy) while preserving narrative continuity across periods.
This project is built as a Claude skill and is designed to be iterated on. Areas where contributions would be most welcome:
- English translation / i18n support
- Export/import functionality (JSON, Markdown, Obsidian)
- Additional AI analysis dimensions
- Additional AI providers
- Mobile-optimized layout
- PWA / offline support
Built with Claude.