Five Ideas for Getting AI to Work for You

Chuck Blake // ERA // March 2026

Five Ideas for Getting AI to Work for You

or: How I Spent a Completely Normal Amount of Time Arguing With a Mac mini

This is Gomez.
He lives on that Mac mini.
He has opinions.

Chuck Blake // ERA // March 2026

Everyone's Using AI

Using it isn't the same as leveraging it

You've used it this week.
But is your business actually moving faster?
Easy to confuse activity with progress
I have to remind myself of this constantly.

The goal is not to use more AI. The goal is to get more done.

A Personal AI Chief of Staff

Not a chatbot. A chatbot answers questions. A chief of staff gets things done.

Runs on my machine — my data, my rules
Connected to Telegram, my calendar, my code, my tasks
Always on — memory, tools, real actions

Gomez

A Personal AI Chief of Staff, Built on OpenClaw

An open source local AI agent platform

Connects to Telegram, your tools, your code, your calendar
Memory, skills, evals, principles — all live here
It did not go perfectly — these five ideas are what I'm working on to make it function

I built something, it seemed to work, I moved on. It was not working.

Five Ideas to Improve Your System

Not a framework. Just what's been working.

Memory
Context
Skills
Evals
Principles

1. Memory

Give your AI a memory it can rely on

Four layers: identity, errors, long-term memory, daily log
Semantic search via local vector DB (Mem0 + Qdrant)
Key insight: this is YOUR thinking, not the model's

The agent wakes up fresh every session. The files are the memory.

How Memory Works

Four layers, two systems, one portable brain

EVERY SESSION

Daily Log

Raw notes — what happened, what was decided

AS IT HAPPENS

errors.md

Mistakes logged in real time. Don't repeat.

STABLE

IDENTITY.md

Who I am, how I communicate, hard rules

→

WEEKLY DISTILLATION

MEMORY.md

Curated long-term memory — decisions, patterns, preferences

ONGOING

Mem0 + Qdrant

Semantic vector search — finds relevant context by meaning

↓

EVERY BOOT

Agent loads context

Identity → errors → long-term memory → today's log → semantic search

Wakes up knowing who it is and what matters

2. Context

Organize your personal context into a searchable vault

Curated context (your vault, your errors, your preferences) gets your output
Build a personal knowledge base — notes, highlights, decisions
The model is a commodity. Your context is the moat.

What's In The Vault

An Obsidian knowledge base — personal, structured, searchable

📁 People

Contacts, bios, relationship notes, birthdays

📁 Journal

Daily logs — what happened, what was decided

📁 Decisions

Architectural and product decisions with rationale

📁 Health

Sleep, exercise, mood — logged daily

📁 Meetings

Notes, action items, follow-ups

📁 Knowledge

Ideas, research, reading notes

📁 Projects

GetMusic, SonicSift, IndieCrates — context + status

📁 Readwise

Highlights from books, articles, podcasts

🔍

QMD — the entire vault is indexed and vectorized. Hybrid search: keyword + semantic. Ask a question, get the right note. bash qmd-search.sh query "What is Murat's degree in?"

This is not a notes app. It is the context layer that makes every AI interaction personal.

3. Skills

Build reusable skills, not one-off prompts

A skill is a reusable instruction set for a specific task
Not model-specific — works across Claude, GPT, whatever is next
~75 skills in my setup. Each one took 20 min to build. Each one saves hours.

Agent orchestrates — scripts execute.

Skills in Practice

Same pattern every time: LLM decides, script executes

GOOD-MORNING

🧠 LLM

Ask health questions conversationally

⚙️ Script

Health log

DEPLOY

🧠 LLM

Decides sequence. Watches CI. Reads logs. Verifies deploy succeeded.

⚙️ Script

Git commands. GitHub token auth. Heroku API calls. Version bump.

PROCESS-INBOX

🧠 LLM

Suggest GTD task names

⚙️ Script

MCP calls right to OmniFocus

4. Evals

Replace vibes with a score

An eval is a test for AI output
Without evals: improving by feel ("seems better?")
Autoresearch: runs a skill repeatedly, scores output, mutates the prompt, keeps improvements

You cannot improve what you cannot measure. This is true for software. It is especially true for AI.

What An Eval Actually Looks Like

Six yes/no questions. Run against every output. No subjectivity.

EVAL 1

Does the output contain all 7 required sections?

EVAL 2

Is there any placeholder or error text visible?

EVAL 3

Does the subject line include today's date and a specific summary?

EVAL 4

Is every item actionable — not just informational?

EVAL 5

Does the tone match — direct, no filler openers?

EVAL 6

Does the HTML render cleanly with no broken tags?

Each question is binary. Score = questions passed ÷ total. Baseline was 33%. Four mutations later: 100%.

5. Principles

Scale your taste, not your attention

Most people direct every AI interaction individually
Principles embed your judgment into the system itself
You stop being in every loop — your standards are

The bottleneck used to be execution. Now it is judgment. That is the right problem to have.

What A Principle Looks Like In The Repo

CLAUDE.md — committed to git, read by the agent every session

THE PRINCIPLE (in CLAUDE.md)

# Architecture

API-first: any functionality that exists
in the UI must also exist in the API.
No UI-only features.

All new endpoints require RSpec tests
before merge.

CHUCK ASKS: "Add bulk tag filtering to search"

Agent's plan:

Step 1: Add GET /api/v1/tags/filter
Step 2: Write RSpec tests
Step 3: Build UI against the API

Chuck never mentioned the API.
The principle did.

Write it once. Commit it. Every agent that touches the repo inherits your judgment.

The Model Is Replaceable. The System You Build Around It Is Yours.

The model is a commodity. What you build around it isn't.

Everyone has access to the same AI — GPT, Claude, whatever comes next
Nobody else has your memory, your context, your judgment in the system
That gap compounds. Every skill you build, every correction you log, every principle you embed — it widens.

These ideas aren't OpenClaw-specific. Build your memory in files. Build your context in a vault. Build your skills as text. Any platform that reads them inherits your system.

chuckblake.com/presentations/getting-ai-to-work-for-you/

Find Me

Website: chuckblake.com
X: @chuckblake
LinkedIn: /in/chuckblake
Slides: chuckblake.com/presentations/getting-ai-to-work-for-you/