Portfolio Portal
Operating Model living owner: Azwaan reviewed: 2026-07-05

AI Collaboration Model

How humans, AI assistants, orchestration systems, and future autonomous agents collaborate to operate the ecosystem. This is the operating counterpart to the Context Architecture: context tells each actor what to know; this page describes who does what, together. Implemented behaviour is clearly separated from future vision.

The actors

graph TD
    HUMAN["👤 Human (founder / architect)<br/>capture · review · approve · govern"]
    subgraph Assistants["AI assistants (interactive)"]
        CC["Claude Code ✅"]
        GPT["ChatGPT ✅"]
    end
    SKILLS["🧰 Shared Skills ✅<br/>capability the assistants invoke"]
    subgraph Future["Future autonomy 🔮"]
        OC["OpenClaw ⏳<br/>self-hosted agent"]
        HER["Hermes ⏳<br/>learning loop"]
        ORCH["portfolio-portal-orchestrator ⏳<br/>self-documenting"]
    end

    HUMAN -->|directs| CC & GPT
    CC & GPT -->|invoke| SKILLS
    SKILLS -->|act on| REPOS[(repos · packs · portal)]
    HUMAN -->|governs via ADRs| REPOS
    OC -.consumes services.-> REPOS
    HER -.re-weights packs.-> REPOS
    ORCH -.regenerates.-> REPOS
    HUMAN -->|approves| GATE{{Approval gates}}
    classDef future fill:#eee,stroke-dasharray:4 3;
    class OC,HER,ORCH future;

Who does what

Actor Role Status Evidence
Human (founder/architect) Capture, review, approval gates, architectural governance (ADRs), strategy ✅ Implemented Operating Principles O1–O2
Claude Code Interactive AI assistant: builds/documents repos & this portal, invokes skills ✅ Implemented ai/claude.handoff.md; this portal was built with it
ChatGPT Interactive AI assistant with drop-in-to-templates working rules ✅ Implemented (working rules defined) ai/chatgpt.handoff.md
Shared Skills The capability layer the assistants invoke (169 skills; deterministic scripts + method) ✅ Implemented cap-reusable-ai-skills
AI-in-product (assessment) LLM explains findings/recommendations from structured facts only ✅ Implemented (engine) cap-website-assessment
OpenClaw Self-hosted personal agent; planned governed consumer of packs/services ⏳ Planned (FIP Phase 5) Intelligence Platform
Hermes Planned learning loop re-weighting intelligence from outcomes ⏳ Planned (not built) Intelligence Platform
portfolio-portal-orchestrator Planned skill that auto-maintains this portal ⏳ Spec only SPEC

Collaboration patterns

Implemented today

  1. Human-directed, skill-powered assistance. The founder directs Claude Code / ChatGPT, which invoke Shared Skills to build repos, author packs, and maintain the portal. Handoff files (AI-HANDOFF) give any assistant the working rules; the portal + llms.txt supply context.
  2. AI-as-narrator inside products. In the assessment engine, deterministic rules decide and the LLM only explains — a bounded, auditable use of AI (Operating Principle O3).
  3. Human governance. Structural decisions are ADRs; nothing external is auto-sent.

Future vision (not built)

  1. Governed autonomous agents. OpenClaw becomes a service consumer — it pulls intelligence via the (future) retrieval service and assessment service, under one claims/tone guardrail, rather than a bespoke integration.
  2. Closed learning loop. Hermes observes outcomes and publishes re-weighted pack versions (one-way, O9), making recommendations outcome-driven instead of inference-derived.
  3. Self-documenting portal. The orchestrator regenerates digests, diagrams, registries, and llms.txt from source — keeping the operating manual true without manual upkeep.

The division of labour (target)

graph LR
    H[Humans] -->|judgement, approval, governance| WORK((Work))
    AI[AI assistants + in-product AI] -->|drafting, extraction, explanation| WORK
    AUTO[Deterministic automation] -->|scoring, retrieval, versioning| WORK
    AGENTS[Future agents] -.->|routine consumption & upkeep| WORK

Principle: as autonomy increases, the human approval boundary is preserved — automation expands within stages, but capture/review/approval and architectural governance remain human (Operating Principles).

  • Keep AI explaining, not deciding in customer-facing outputs; stricter review for regulated verticals (health).
  • Route future agents through service contracts, not venture repos (ADR 0005, proposed).
  • Converge AI prompt surfaces onto a shared claims/tone guardrail (Architecture Risks R9).