Personal Operating Cockpit

Goal-aligned bounded autonomy for solo builders.

Comment

Inspiration

What it does

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for Personal Operating Cockpit

Inspiration

As I built more AI-assisted projects, my main problem changed.

The question was no longer whether AI could complete tasks. The harder question was whether it was working on the right things, supporting my actual life goals, respecting decisions that should remain human, and keeping project state trustworthy.

I therefore built a Personal Operating Cockpit for solo builders.

What it does

At the top of the cockpit are confirmed human life goals. Every project must explain how it contributes to one of those goals before it consumes time or autonomous AI work.

The cockpit provides a shared portfolio view across multiple AI-assisted projects. For OpenAI Build Week, I implemented one complete synthetic Evidence-to-Action reference workflow:

  1. A project signal or change is detected.
  2. GPT-5.6 performs a bounded, structured analysis after an explicit user action.
  3. The change is connected to a confirmed human goal.
  4. A Human Decision Gate presents two predefined options.
  5. After the human decision, only registered local tools may run.
  6. Their results are collected into an Evidence Bundle.
  7. Proposed state changes are reviewed field by field.
  8. Only explicitly accepted or edited fields become confirmed project state.

The AI cannot invent commands, select arbitrary paths, register new tools, or silently turn its own output into operational truth.

How I built it

The demonstration runs as a local Node.js application with a deliberately restricted demo boundary.

It includes:

  • A synthetic six-project portfolio
  • A persistent Decision Run lifecycle
  • A structured GPT-5.6 analysis contract
  • A Human Decision Gate
  • A fixed tool registry
  • Read-only local tool adapters
  • Evidence Bundle generation
  • Field-level state review
  • A separate synthetic runtime state
  • Automated functional, syntax and visual tests

The submitted demo contains no real customer, applicant, financial or private project data.

Codex supported the architecture, implementation, debugging, testing, privacy review and fresh-clone verification. I made the core product decisions around goal alignment, reserved human authority, tool boundaries, evidence requirements and confirmed state.

GPT-5.6 performs the bounded structured analysis inside the application.

What was built during Build Week

The broader personal cockpit existed before Build Week.

During Build Week, I meaningfully extended it with the complete workflow demonstrated in the video:

  • Goal-linked structured analysis
  • Persistent Decision Runs
  • Human Decision Gates
  • Registered tool execution
  • Evidence Bundles
  • Field-level state review
  • Synthetic demo and privacy boundary
  • Automated tests and reproducible setup documentation

The repository documents the distinction between the pre-existing cockpit and the Build Week implementation.

Challenges I ran into

The main challenge was not giving the AI more autonomy. It was defining useful autonomy without transferring human authority to the model.

The system had to distinguish between:

  • Analysis and decision-making
  • Tool output and trusted evidence
  • Proposed state and confirmed state
  • A useful demonstration and exposure of private project information

Another challenge was making this architecture understandable within a three-minute video.

Accomplishments that I am proud of

I am especially proud that the control boundaries are implemented technically rather than described only as principles.

The submitted project includes:

  • A complete end-to-end synthetic workflow
  • A hard boundary between demo and private workspace state
  • Fixed registered tools instead of model-generated commands
  • Evidence requirements before state review
  • Explicit accept, edit or ignore decisions for every proposed field
  • 65 passing automated tests
  • Fresh-clone verification records and setup documentation

What I learned

Reliable AI autonomy requires more than model intelligence.

It requires explicit goals, reserved human decisions, constrained tools, verifiable evidence and a clear distinction between AI proposals and confirmed operational truth.

The most useful system is not necessarily the one that acts most independently. It is the one that knows where its authority ends.

What is next

The next step is to extend the same controlled workflow to additional project types and registered adapters.

Future work includes:

  • Better cross-project goal prioritization
  • Detection of conflicts between projects and life goals
  • Additional read-only evidence sources
  • Improved portfolio-level recommendations
  • Optional self-hosted deployment
  • Controlled integrations for explicitly approved external actions

The long-term goal is a personal AI operating layer that helps people use autonomous systems without losing control over what matters, what is decided and what becomes trusted state.

Built With

Share this project:

Updates