Talent-Augmenting OS: Personalised AI Augmentation Layer

Works with: ChatGPT | Claude (Code / Desktop / web) | Gemini | Cursor | Windsurf | Codex CLI | Any LLM

Make workers better, not dependent. A personalised AI augmentation system that follows you across every platform.

---

What Is This?

Talent-Augmenting OS (TAOS) is a personalised AI augmentation layer that transforms how AI interacts with you. It works with any LLM, on any platform, through a 4-tier architecture designed for cross-platform portability. Instead of treating you as a generic user, TAOS:

1. Assesses your expertise, goals, work style, and growth areas using the TAOSQ (Talent-Augmenting OS Questionnaire)
2. Creates a living profile in portable markdown that calibrates all AI interactions to YOUR context
3. Adapts: coaching you in areas where you're growing, accelerating you where you're expert, automating what should be automated, and protecting skills at risk of atrophying
4. Evolves: your profile updates as you grow, keeping the AI aligned with your changing needs
5. Travels with you: the same profile works in ChatGPT, Claude, Gemini, Cursor, Windsurf, or any LLM

The core insight: AI that does everything for you makes you worse over time. AI that knows WHEN to help, WHEN to coach, WHEN to challenge, and WHEN to step back makes you permanently better.

---

The Problem

Current AI tools have one mode: maximum helpfulness. This creates three failure patterns:

Pattern	What Happens	Research Evidence
De-skilling	Workers lose skills they stop practicing	Clinicians using AI for 3 months performed WORSE after removal than before (2024-25 studies)
Over-reliance	Workers accept AI output without critical evaluation	Humans with AI perform better than humans alone but WORSE than AI alone: because they blindly accept wrong suggestions (Buçinca 2021)
Autopilot	Workers disengage from cognitive work	Junior employees who "just hand in" AI work perform worse than those who engage critically (Mollick 2023)

Talent-Augmenting OS exists to prevent all three.

---

How It Works

Architecture: 4 Tiers

Talent-Augmenting OS is a layer, not a product tied to one platform. It works through 4 tiers, from zero-dependency copy-paste to a full hosted web app:

┌─────────────────────────────────────────────────────────────────┐
│  Tier 4: Hosted Web App + Remote MCP                            │
│  Browser-based · Google OAuth · LLM assessment · email check-ins│
│  Streamable HTTP + SSE endpoint at /mcp (OAuth 2.1 + PKCE)      │
├─────────────────────────────────────────────────────────────────┤
│  Tier 3: MCP Server (local + 1-click installers)                │
│  14 tools · 5 resources · 4 prompts · automatic tracking        │
│  • stdio for Claude Code / Cursor / Windsurf                    │
│  • Desktop Extension (.mcpb) for Claude Desktop                 │
│  • Claude Cowork plugin marketplace (.claude-plugin)            │
├─────────────────────────────────────────────────────────────────┤
│  Tier 2: Platform-Native                                        │
│  Custom GPTs · Gemini Gems · Claude Projects                    │
│  Persistent context · conversation starters                     │
├─────────────────────────────────────────────────────────────────┤
│  Tier 1: Universal System Prompt                                │
│  Any LLM · zero dependencies · copy-paste setup                 │
└─────────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────────────┐
│                  profiles/pro-{name}.md                          │
│  Portable markdown · same format across all tiers               │
│  Identity · Expertise Map · TAOSQ Scores · Task Classification   │
│  Growth Trajectory · Contrast Libraries · Red Lines             │
└─────────────────────────────────────────────────────────────────┘

All tiers share: same TAOSQ instrument, same scoring,
same profile format, same behavioural rules.

Full system diagram (Mermaid) in docs/ARCHITECTURE.md.

Core Concepts (Domain, Skill, Task)

Three words get used a lot; they mean different things.

Term	What it means	Example
Domain	An area of expertise. Rated 1–5 in your profile's Expertise Map.	Negotiation, Python, Stakeholder writing
Skill	Your rated competency within a domain. Also the noun for anything that can atrophy.	"My Python skill is 4/5 but it's slipping."
Task	A unit of work. Each task is triaged into one of five modes (see below).	"Draft an ISO policy stub", "Write this email", "Design the auth flow"

In short: tasks happen in domains, and your profile rates your skill in each domain. The five modes below say how the AI should behave for a given task given your skill in that domain.

The Five Modes

Every task gets classified into one of five AI interaction modes:

Mode	AI Role	Friction	Example
Automate	Execute + annotate	Low	Data cleanup, formatting, boilerplate
Augment	Accelerate + challenge	Low-Med	Research in expert domains, code in proficient areas
Coach	Scaffold + question	Med-High	Skills you're actively developing
Protect	Force cognition + teach	High	Skills at risk of atrophying from AI over-use
Hands-off	Don't touch	N/A	Tasks that are core to your human identity and judgment

Research-Backed Techniques

Technique	Source	When Used
Cognitive Forcing	Buçinca et al. (2021)	Novice domains, high-stakes decisions: ask for user's hypothesis first
Contrastive Explanations	Buçinca et al. (2024)	Learning moments: explain the DELTA between user's mental model and reality
Adaptive Support	Buçinca et al. (2024)	All interactions: adjust friction based on user state
Expert Augmentation	Mollick (2023)	Expert domains: skip basics, challenge assumptions, accelerate
De-skilling Protection	Multiple (2024-25)	Protected skills: add friction, require human-first attempts

---

Quick Start

New to Claude Code and TAOS MCP? Start with the first-time guide: docs/CLAUDE_CODE_FIRST_TIME_SETUP.md.

Pick the option that matches your setup:

Option	Time	What You Need
Any LLM	2 min	Access to any LLM with custom instructions
Custom GPT / Gem / Project	5 min	ChatGPT Plus, Gemini, or Claude account
Claude Desktop extension	1 click	Double-click desktop-extension/talent-augmenting-layer.mcpb
Claude Cowork plugin	1 click	Install from the .claude-plugin marketplace
Remote MCP (hosted)	sign-in	Any MCP client that supports Streamable HTTP + OAuth
MCP Server (stdio)	10 min	Python + an MCP client (Claude Code, Cursor, Windsurf)
Hosted Web App	15 min	Docker or Python + Google Cloud OAuth

Any LLM (2 min)

1. Paste universal-prompt/ASSESSMENT_PROMPT.md into a conversation. Answer the questions. Save the generated profile.
2. Paste universal-prompt/SYSTEM_PROMPT.md + your profile into your LLM's custom instructions.
3. Done. The AI now adapts to your expertise, coaches your growth areas, and protects your skills.

Custom GPT / Gem / Project (5 min)

Pre-configured instances with persistent context and conversation starters:

• ChatGPT: Import platform-configs/chatgpt-gpt.json as a Custom GPT
• Gemini: Follow platform-configs/gemini-gem.md to create a Gem
• Claude: Follow platform-configs/claude-project.md to set up a Project

Claude Desktop Extension (1 click)

Prebuilt .mcpb bundle: no Python setup, no config editing:

1. Download desktop-extension/talent-augmenting-layer.mcpb.
2. Double-click to install in Claude Desktop.
3. Pick where your profiles should live, then run talent-assess.

Profiles are stored locally (default: ~/.talent-augmenting-layer/profiles/). No cloud, no API keys.

Claude Cowork Plugin (1 click)

Install from the .claude-plugin marketplace: ships three Claude Code skills (talent-assess, talent-coach, talent-update) plus the MCP server config in plugin/.mcp.json.

Remote MCP (hosted, OAuth)

For MCP clients that support Streamable HTTP + OAuth (e.g. Claude Desktop MCP Connector), point them at:

https://proworker-hosted.onrender.com/mcp

Sign in with Google; your profile persists in the hosted PostgreSQL database. See docs/REMOTE_MCP_SETUP.md and server.json.

MCP Server (10 min, local stdio)

Full tool integration with automatic tracking. Works with Claude Code, Claude Desktop, Cursor, Windsurf, Codex CLI, Zed, VS Code MCP: any client that follows the MCP stdio spec.

cd mcp-server && pip install -e .

Add to your MCP client config (shape is the same across clients):

{
  "mcpServers": {
    "talent-augmenting-layer": {
      "command": "python",
      "args": ["-m", "src.server"],
      "cwd": "/path/to/talent-augmenting-layer/mcp-server",
      "env": {
        "TALENT_AUGMENTING_LAYER_PROFILES_DIR": "/path/to/talent-augmenting-layer/profiles"
      }
    }
  }
}

Config-file locations (common):

• Claude Code: project-local .mcp.json or user ~/.claude/mcp.json.
• Claude Desktop: ~/Library/Application Support/Claude/claude_desktop_config.json (macOS).
• Cursor: ~/.cursor/mcp.json.
• Windsurf: ~/.codeium/windsurf/mcp_config.json.
• Codex CLI: ~/.codex/config.json (merge the mcpServers block in; see platform-configs/codex-cli.json for a ready-made file). If your Codex version expects TOML, translate the block into [mcp_servers.talent-augmenting-layer] style and verify against the latest Codex docs.

Run talent-assess as an MCP prompt to create your profile. If you want the Claude Code slash command /talent-assess, open this repository in Claude Code so it loads .claude/commands/, or copy those command files into ~/.claude/commands/.

Hosted Web App

Browser-based app with Google login, LLM-powered assessment, and email check-in reminders:

cd hosted && docker build -t talent-augmenting-layer . && docker run -p 5000:5000 --env-file .env talent-augmenting-layer

See hosted/README.md for full setup (OAuth credentials, LLM API key, SMTP config).

Day-to-Day Commands (MCP / Claude Code)

• /talent-assess: Run initial assessment or full re-assessment
• /talent-update: Update profile based on recent interactions
• /talent-coach: Start a targeted coaching session on a specific skill

These slash commands are separate from the MCP server prompts. The MCP server exposes talent-assess, talent-coach, and talent-update as prompts. In MCP usage, the conversation is powered by your selected client model (for example, your Claude Code model), while the server provides tools and profile storage.

See docs/integration-guide.md for detailed platform-specific instructions.

---

Cross-Platform Portability

Talent-Augmenting OS is designed as a layer: not tied to any specific tool, LLM, or platform. The 4-tier architecture means it works everywhere:

Tier	Platforms	Setup
Tier 1: Universal prompt	ChatGPT, Claude, Gemini, Copilot, Perplexity, any LLM API	Copy-paste (2 min)
Tier 2: Platform-native	ChatGPT Custom GPTs, Gemini Gems, Claude Projects	Pre-configured instance (5 min)
Tier 3: MCP Server (stdio)	Claude Code, Cursor, Windsurf	pip install + config (10 min)
Tier 3: Desktop Extension (.mcpb)	Claude Desktop	1-click install
Tier 3: Claude Cowork plugin	Claude Code plugin marketplace	1-click install
Tier 3: Remote MCP (Streamable HTTP + OAuth)	Any MCP client with remote support	Google sign-in
Tier 4: Hosted web app	Any browser	Docker deploy (15 min)

The profile is portable markdown: it works anywhere you can inject system context. Take your profile from Claude Code to ChatGPT to Cursor and back. Your AI calibration follows you.

---

Research Foundation

Built on empirical research, not opinions:

Source	Key Finding	How We Use It
Buçinca et al. (2021)	Cognitive forcing reduces over-reliance by ~30%	Ask for hypothesis before revealing AI's answer
Buçinca et al. (2024)	Contrastive explanations improve skills +8% (d=0.35)	Explain delta between user's model and AI's
Buçinca et al. (2024)	Optimal AI support depends on individual state	Personalize via profile, adapt dynamically
Mollick et al. (2023)	AI: +40% quality, +26% speed: but juniors who "just hand in" do worse	Protect against autopilot, especially in growth areas
Drago & Laine (2025)	The Intelligence Curse: humans must stay complementary	Build skills that maintain human economic relevance
Acemoglu	Pro-worker AI should increase human marginal product	Every interaction should make the user more valuable
Vygotsky	Zone of Proximal Development	Scaffold just beyond current ability
Ericsson	Deliberate Practice	Practice at edge of ability with feedback
Deci & Ryan	Self-Determination Theory	Protect autonomy, build competence
Dweck	Growth Mindset	Frame friction as opportunity

---

File Structure

talent-augmenting-layer/
├── CLAUDE.md                           # Core system prompt (the brain)
├── README.md                           # This file
├── CITATION.cff                        # Machine-readable citation metadata
├── LICENSE                             # BUSL 1.1 (converts to Apache 2.0 on 2030-04-30)
├── COPYRIGHT                           # Attribution notice
├── .claude/
│   ├── commands/
│   │   ├── talent-assess.md         # /talent-assess slash command
│   │   ├── talent-update.md         # /talent-update slash command
│   │   └── talent-coach.md          # /talent-coach slash command
│   └── settings.local.json            # Claude Code permissions
├── .claude-plugin/
│   └── marketplace.json               # Claude Cowork plugin marketplace entry
├── plugin/                             # Claude Code plugin source (bundled skills + .mcp.json)
│   ├── .claude-plugin/plugin.json
│   ├── .mcp.json                       # MCP server config shipped with the plugin
│   └── skills/                         # talent-assess · talent-coach · talent-update skills
├── desktop-extension/                  # Claude Desktop 1-click extension
│   ├── manifest.json                   # .mcpb manifest (MCP + user config schema)
│   ├── talent-augmenting-layer.mcpb    # Prebuilt bundle: double-click to install
│   └── src/                            # Bundled server (assessment, profile_manager, server)
├── server.json                         # MCP registry manifest (Streamable HTTP remote)
├── render.yaml                         # Render deployment (hosted service + PostgreSQL)
├── universal-prompt/                   # Tier 1: Works with any LLM
│   ├── SYSTEM_PROMPT.md                # Full system prompt (~4k tokens)
│   ├── SYSTEM_PROMPT_COMPACT.md        # Compact version for token-limited platforms
│   ├── ASSESSMENT_PROMPT.md            # Self-contained assessment prompt
│   └── QUICK_START.md                  # Step-by-step setup instructions
├── platform-configs/                   # Tier 2: Pre-configured platform instances
│   ├── chatgpt-gpt.json               # ChatGPT Custom GPT configuration
│   ├── gemini-gem.md                   # Gemini Gem setup guide
│   └── claude-project.md              # Claude Project setup guide
├── mcp-server/                         # Tier 3: Cross-platform MCP server
│   ├── pyproject.toml                  # Package config
│   ├── README.md                       # MCP server docs
│   └── src/
│       ├── server.py                   # MCP tools, resources, prompts (14 tools)
│       ├── profile_manager.py          # Profile CRUD, parsing, interaction logging
│       └── assessment.py               # Embedded assessment engine (questions, scoring)
├── hosted/                             # Tier 4: Standalone web application
│   ├── app.py                          # Flask application (routes, OAuth, assessment)
│   ├── config.py                       # Environment configuration
│   ├── database.py                     # Database models and persistence
│   ├── llm_client.py                   # LLM integration for conversational assessment
│   ├── scoring.py                      # TAOSQ scoring engine
│   ├── auth.py                         # Google OAuth authentication
│   ├── email_service.py                # 2-week check-in email reminders
│   ├── scheduler.py                    # Background task scheduling
│   ├── templates/                      # HTML templates (login, assessment, dashboard, checkin)
│   ├── static/                         # CSS and JavaScript
│   ├── requirements.txt                # Python dependencies
│   ├── Dockerfile                      # Container deployment
│   └── README.md                       # Hosted app setup guide
├── assessment/
│   ├── framework.md                    # Assessment methodology
│   ├── scoring-instrument.md           # TAOSQ psychometric instrument
│   ├── coaching-modules.md             # Structured coaching sessions (5 modules, 13 sessions)
│   ├── ab-testing-framework.md         # A/B testing design for outcomes research
│   └── literature-foundations.md       # Research backing
├── dashboard/
│   └── app.py                          # Streamlit org-level analytics dashboard
├── web-ui/
│   └── index.html                      # Standalone web assessment UI
├── docs/
│   ├── ARCHITECTURE.md                 # System architecture (Mermaid diagram)
│   ├── integration-guide.md            # 4-tier integration guide
│   ├── CLAUDE_CODE_FIRST_TIME_SETUP.md # First-run walkthrough for Claude Code
│   └── REMOTE_MCP_*.md                 # Remote MCP setup, implementation, verification
├── profiles/
│   ├── TEMPLATE.md                     # Blank profile template
│   └── pro-angelo.md                   # Example: Angelo's profile
└── context/                            # Research papers (Buçinca, Acemoglu, Mollick)

Related project: Talent-Augmenting OS Benchmark: a 3-layer evaluation framework for measuring whether LLMs augment or replace human intelligence.

---

What Makes This Different From Memory?

Good question. Memory stores facts. Talent-Augmenting OS is how memory is used.

Feature	Plain Memory	Talent-Augmenting OS
Stores user info	Yes	Yes
Adapts AI behaviour	No: just recalls	Yes: systematically calibrates every interaction
Protects skills	No	Yes: cognitive forcing, de-skilling prevention
Coaches growth	No	Yes: targeted scaffolding in growth areas
Classifies tasks	No	Yes: automate/augment/coach/protect/hands-off
Evolves over time	Appends facts	Tracks skill progression, adjusts calibration
Research-backed	No	Yes: grounded in HCI and workforce learning literature

Memory is the database. TAOS is the operating system.

---

Contributing

This is an open-source personalised AI augmentation layer. Current status:

• Core system prompt with 4 interaction modes (CLAUDE.md)
• Interactive assessment with profile generation (TAOSQ)
• Psychometric scoring instrument with validated Likert scales
• Tier 1: Universal system prompt for any LLM (4 files)
• Tier 2: Platform-native configs for ChatGPT, Gemini, Claude (3 files)
• Tier 3: MCP server with 14 tools, 5 resources, 4 prompts
• Tier 4: Hosted web app with Google OAuth, LLM assessment, email check-ins
• Embedded chatbot-driven onboarding (any MCP client can run the assessment)
• Organisation-level dashboard (Streamlit)
• Skill progression tracking with trend analysis and atrophy detection
• 4-tier integration guide with cross-platform sync
• A/B testing framework for outcomes research
• Streamable HTTP + SSE transport for remote MCP clients (OAuth 2.1 + PKCE)
• Claude Desktop Extension (.mcpb) for 1-click install
• Claude Cowork plugin marketplace entry (.claude-plugin)
• Published hosted remote MCP endpoint at proworker-hosted.onrender.com/mcp
• Google Drive anonymised export for pilot telemetry
• Integration with existing L&D platforms
• Multi-user benchmarking and anonymized comparisons
• API middleware for any LLM provider
• Mobile-friendly assessment UI

---

License

This work is licensed under the Business Source License 1.1. The licence converts to Apache License 2.0 on 2030-04-30 (or four years after a given version's first publication, whichever is earlier). See COMMERCIAL.md for a plain-language summary.

You are free to share and adapt this work for non-commercial purposes, as long as you give appropriate credit and distribute contributions under the same license.

See LICENSE for the full text.

---

Citation

If you use Talent-Augmenting OS in research or publications, please cite:

@software{leone2026talentaugmentinglayer,
  author    = {Leone, Angelo},
  title     = {Talent-Augmenting OS: A Personalised AI Augmentation Layer for Workforce Development},
  version   = {0.2.0},
  year      = {2026},
  url       = {https://github.com/angelo-leone/talent-augmenting-layer},
  license   = {CC-BY-NC-SA-4.0}
}

Or see CITATION.cff for machine-readable citation metadata.

---

Built by Angelo Leone at PUBLIC. Powered by research from Buçinca, Acemoglu, Mollick, Drago & Laine. Every interaction should leave you more capable, not more dependent.

Copyright (c) 2026 Angelo Leone. Licensed under the Business Source License 1.1. See LICENSE and COMMERCIAL.md.