Packaging startup advisory into an agent-agnostic 'second brain' skill pack

Problem

If you advise a portfolio of early-stage companies, you've probably hit this: the best of what you do is high-touch, lives in your head, and resets to near-zero with every new engagement. Your output is bottlenecked on your own time, and the same hard-won playbook gets re-derived from scratch each time. At Delphi Labs — a venture studio that researches, writes small checks, and provides heavy zero-to-one advisory — Kevin Simback faced exactly this: the tangible work was repeatable in principle, but trapped in one person's practice.

Approach

Kevin built his advisory on top of a personal knowledge system, with everything routed back to one substrate:

• An Obsidian "second brain" at the center. A local markdown knowledge base whose bidirectional links form a personal knowledge graph; when you point an LLM at it, it's good at pulling all the connected information across files.
• Two self-reinforcing functions. Research/content and advisory feed each other — better research means better advice, and more advisory interactions inject more taste back into the research.
• An agent-agnostic layer. Workflows run through interchangeable agents (Hermes, Claude Code, Codex), so the same task runs from a phone on the go or a PC when cranking on something. Crucially, every agent both pulls from and feeds back into the second brain, keeping it in sync.
• An advisor skill pack covering seven disciplines — positioning, investor slide decks, investor intros, competitive landscape, product strategy / PMF, GTM launch, and roadmap (plus ongoing coaching) — packaging learnings and best practices into repeatable workflows.
• An advisor portal built on the Claude Agent SDK: per-project files (research, founder inputs, roadmap, sessions), deliverable tracking, top-line metrics, and a chat window pre-loaded with the relevant skill.

What he was candid about: he took a deliberately experimental approach since agents emerged, and one tool didn't stick — he no longer really uses OpenClaw, though some instances still run. He also noted the skills are time-consuming by design — invoking the deck-creation skill pulls in a large system prompt and runs a full process, "way more time than this session would allow," so he couldn't run it live.

Results

The studio's advisory is packaged into repeatable, context-aware deliverables — positioning, investor decks, PMF, GTM, roadmap — tracked per engagement, so the work is no longer bottlenecked on one person re-deriving it each time. Beyond invoking skills, the portal doubles as a working tool: chatting with any project for meeting prep or to cut to where a project is struggling right now. Because impact was shown through a portal seeded with fictitious data (for confidentiality) and the most time-consuming skill was described rather than run end-to-end, the demonstrated outcome is the working system itself rather than stated metrics.