Personal AI Workflow — Corey Hobson

How it works

My work runs through a setup that combines plain text files, an AI assistant in the terminal, and a tagged structure that connects everything. The core idea is routing: not every task needs Claude Opus. Most don’t.

The system classifies every request and sends it to the right tier:

Local tier: free, zero latency. Qwen3-30B-A3B running 4-bit quantized on LM Studio. Handles classification, summarization, simple text transforms, and basic scripting. This is where 80%+ of all AI interactions land. Zero API cost. Zero network round trips.

API tier: pay-per-use. OpenRouter with DeepSeek and Claude Sonnet as fallbacks. The default for most real work: code generation, content writing, analysis that needs more than 4-bit reasoning. Roughly $5/month in practice.

Premium tier: flat subscription. Claude Opus for hard reasoning, multi-file refactors, and planning sessions where quality matters more than speed or cost. Reserved for the work that actually needs deep thought.

The pieces

Everything I write lives in .md files. A .md file is plain text with light formatting: a hash mark for a heading, a dash for a bullet. No special software needed. My whole workspace is a folder of these files: meetings, decisions, project briefs, research, journal entries.

I use them because they’re durable. Small, fast, searchable, and not trapped inside an app the way a Word document or a Notion page is.

The second piece is an AI assistant that runs in the terminal. I type a request in plain English. The AI reads the relevant files in my workspace, answers, and edits files when I ask.

Plain text plus an AI that can read and edit it is the combination that makes the rest of this work.

What it runs

Daily brief. Every morning the system scans active workstreams, meeting notes, and the kanban board. Generates a compact briefing in the journal. Takes 30 seconds instead of 15 minutes of tab-switching.

Session logging. After a work session, the assistant writes a third-person prose log: what got decided, by whom, what’s still open. These accumulate into a searchable history that spans months. No more “what did we agree about that?”

Kanban dispatch. Large tasks get decomposed into specialist worker agents with dependency chains. The dispatcher runs them in parallel where possible, serializes where there are gates. I review the outputs; the AI handles the routing.

Code review delegation. PRs go to Claude Code with quality gates and project-specific rules. The review comes back with inline comments and a summary. I read the summary and spot-check the diffs.

Project management. Builds, deployments, and configuration for personal projects (UXRStudy, coreyhobson.com) run through the assistant. I describe what I want; the system figures out how to ship it.

Evening recap and weekly review. At the end of each day, the assistant writes a recap of decisions, progress, and open questions. At the end of the week, it reviews all sessions, identifying patterns, suggesting skills to improve, and sometimes recommending a book or a course. I dump everything in there: what went well, what went badly, directions that didn’t work. It finds the signal.

What I didn’t expect

It built its own knowledge graph. Without me asking, the AI started organizing everything into datasets, interconnecting references, creating categories, building flows between related topics. I didn’t design this. It emerged from the system figuring out that cross-referencing made recall faster.

The habit is the hard part. You have to consciously remember to use it. Ad-hoc conversations in meetings don’t get captured. I can send voice notes via text for personal work, but I can’t do that at work, for security and privacy reasons. I need better habits around writing notes while waiting on things, managing agents between meetings. The tool can’t fix the discipline problem.

If someone asked me where to start: small. Build a couple workflows for daily tasks. Automate research. Use it as a sounding board. Don’t try to architect the whole system upfront. Let it grow.

Is this the permanent way of work? I don’t know. It’s still new. But the emergent behaviors, the dataset organization and the weekly reviews spotting patterns I missed, make me think it might be.

Why this works

Context persistence. The AI doesn’t reset every session. It knows what I worked on yesterday and what’s due this week. Long-running work doesn’t have to be re-explained every time.

Two-Mac continuity. Same assistant, same memory, synced via git every 5 minutes. Work Mac and personal Mac stay in lockstep without manual synchronization.

Zero vendor lock-in. The knowledge graph is plain markdown. If the assistant disappears tomorrow, every decision, note, and dataset is still readable in Obsidian. No vector database. No black-box embedding store. Just files and git.

Cost efficiency. 80%+ of AI interactions run on the local model at zero marginal cost. Cloud API spend is reserved for tasks that genuinely need the reasoning depth.

Stack

Orchestration: Hermes Agent with multi-profile routing
Local inference: LM Studio running Qwen3-30B-A3B (4-bit quantized)
Knowledge base: Obsidian vault (CoreyWorkspace) with git-backed sync
Task orchestration: Hermes Kanban with multi-agent dispatch
Coding delegation: Claude Code CLI for complex implementation work
Hardware: Apple M1 Max, 64GB unified memory