Claude + Obsidian: Closing the Gap Between Ideas and Execution

#ai #agents #programming

Claude Code closed a loop for me today where everything - meetings, decisions, specs, code, documentation - are all connected using AI.

Not just stored.
Connected.

Most of the time, information is scattered everywhere: Teams, Slack, Notebooks (physical and digital), GitHub, Confluence, and in your head.

A lot of time can be lost on large systems — not writing code but figuring out what’s going on.

So, I started experimenting with a different approach.

Step 1 — Capture: Meetings Become Part of a Knowledge Base

The first problem to solve was capturing information automatically. I started using Plaud.ai Notepin last year to improve my notetaking, but I was originally exporting markdown files to save to Obsidian.

Meetings are recorded, transcribed, summarized, and automatically stored as markdown files in an Obsidian vault. That means meetings are no longer just conversations — they become searchable, permanent knowledge.

But now, I am automating it and giving agents access to it alongside the code.

Here’s the meeting capture flow:

Step 2 — Spec: Knowledge Becomes Requirements or Revisions

Once meetings are stored as markdown, the next step is where things start to change.

Instead of writing specs from scratch, what if the agent reads meeting notes as part of its planning phase.

Here’s the spec-driven workflow:

The agent is not working from a prompt.
The agent is working from a living spec.

The spec becomes the source of truth, and it is continuously updated based on meetings, new decisions, and implementation feedback. But we don't have to stop at just meetings - we could have other documentation, architecture notes, and business rules.

Step 3 — Memory: The System Can See Everything

Now comes the part that really changes things.

Because meetings, specs, code, and tests are all stored in accessible locations, the agent can reason across all of them.

This enables completely different kinds of questions:

What decisions from meetings are not in the spec?
What parts of the spec are not implemented?
What changed between the original request and the final implementation?
What risks were mentioned but never addressed?
Write a status update based on meetings, specs, and commits.
Generate documentation for stakeholders based on everything above.

This is more than automation.

This is intelligent traceability.

We now have a knowledge base for each feature, for each system, for each department, or a business in its entirety.

Version Control for Decisions

We version control code. But we don’t always version control decisions. When AI can see meetings, specs, code, and documentation together, it stops being just a coding assistant.

What I'm envisioning is a system where: