How We Work With Claude | Merlino Team Playbook

How it startedStart here

It started with one folder

Mike dropped a folder with 3 files sitting next to each other. They had nothing to do with each other, just three downloads in the same place.

rag-architect.skill (a skill file)
README.md (a memory system from GitHub)
SKILL(1).md (a session-end skill, skipped)

Two of these are unrelated. Don't let them blur together.

Track AThe .skill file

Track A: a new skill (NOT from GitHub)

rag-architect is a helper that designs the right way to let an AI search your data (chat with your docs). We installed it and assigned it to the right agents. One and done.

Installed to your skills
Owned by Merlin, Dan, Memz
Picks the right retrieval design and writes starter code
Came from the .skill file, NOT the github

Track BThe GitHub README

Track B: a memory system on GitHub

The README was an open-source shared brain for AI agents, called Zengram. Interesting, but you already have Master Brain. The question was: is anything in it worth taking?

Open-source, MIT
A shared memory across agents
You already own Master Brain (850K memories)
So: test it before judging

How we decidedTest first

We tested it for real first

Instead of trusting the README, we stood the whole thing up on a server and poked it. The verdict: the newest version broke the one feature that made it special, so we do NOT run it. But 3 of its ideas were genuinely good.

Stood it up isolated on Hetzner
Proved its features with real tests
Newest version broke multi-agent attribution
Verdict: don't run it, take 3 ideas

What we kept3 ideas, no new system

We put 3 ideas into YOUR brain

We rebuilt 3 mechanisms straight into Master Brain. Nothing from their system runs. Just 3 upgrades to the brain you already use.

Secret-scrubbing: keys and passwords get blanked before saving
Decay: old unused facts sink, fresh ones float
Supersede: a new fact retires the old version instead of stacking

rag-architect = a skill, from the .skill file. The 3 brain upgrades = ideas from the GitHub README. Two different tracks, same folder.

The proofNothing broke

Proof we didn't break anything

All 850,000 of your existing memories were left untouched. We ran the same searches before and after, the top results were identical. Every change is additive and reversible.