Layer 01
Identity
SOUL.md came first. The instruction file for Romulus’s tone, modes and when to push back.
Designed on day one
Romulus
My personal AI setup on OpenClaw.
scroll
00 · The Brief
A full-time job, evenings and weekends, and an AI setup to help.
Romulus is my personal AI setup on OpenClaw. I use it to research ideas, help build them and keep track of the work. I wrote Romulus’s instructions, organized its memory, picked models and tools for each job and set what needs my approval. This covers the first six months, from February 2026. Counts, models and statuses are a snapshot; the work continues.
0
Months covered
From February 2026
0+
Ideas reviewed
Researched in that period
0
Project examples
The main projects below
0
Memory stages
Flat files to a wiki layer
The setup, the workflow, what broke and what came of it.
01 · The Constraint
I had about ten hours a week and too many ideas.
I have a full-time product job. Evenings and weekends aren’t enough to research, build and test more than one idea at a time.
I wanted to see how much I could hand off to an AI assistant with steady instructions, notes and limits.
I also had a model budget, and some jobs, like morning briefs, had to run without me watching.
Most of the work turned out to be writing clear tasks and checking what came back.
02 · The Setup
Six parts I configured on OpenClaw. Some came later, after something broke.
Layer 02
Memory
Flat files, an Obsidian vault, QMD search, then a wiki layer.
Rebuilt three times.
Layer 03
Research
Brave Search, X scans, Firecrawl, scripts and model critiques before I decide to build.
No permission needed to research.
Layer 04
Routing
Each model got a job, a budget, a context limit and a known failure mode.
Added after Failure III
Layer 05
Delegation
I approve a spec, Claude Code builds it, then checks and staging.
Revisited after Failure IV
Layer 06
Approval rules
The Fortress: no messages, payments, account changes, public posts, credentials or production deploys without my OK.
Set in instructions and config.
03 · Who Does What
In these projects, I wrote specs and reviewed. Agents wrote the code.
I chose what to build, approved specs, reviewed results and made release calls.
Romulus researched, drafted specs, kept notes and coordinated builds. Claude Code wrote the code.
Agents can be confidently wrong. When one looped, the cause was usually my task description, so I fixed the spec first.
I
Research Lead
Set the question and decide when there’s enough evidence.
II
Product Strategist
Decide what to build, park, drop or reframe.
III
Spec Writer
Turn direction into specs with acceptance criteria.
IV
Delegation Manager
Pick the model or tool for each task.
V
Design Lead
Set the interaction model and quality bar.
VI
Reviewer
Read diffs, test flows and reject output that only looks right.
VII
Release Manager
Keep staging, previews and my sign-off before production.
VIII
Postmortem Owner
Write up failures and change the setup to match.
04 · The Loop
The workflow, in eight steps. Question, research, critique, spec, routing, build, review, notes.
01
Frame the product question
I decide what we’re trying to learn, prove, build or drop. A vague question makes every later step worse.
Written first: audience, constraint, success signal.
02
Research the terrain
Market, competitors, pricing, pain, timing and a path to revenue.
Pain and timing first, so a trend doesn’t pass for an opportunity.
03
Pressure-test the thesis
Grok, ChatGPT, Claude and Romulus critique the idea separately. I’m looking for contradictions before code.
Example: Forbidden got reframed here.
04
Write the spec
Product and technical requirements, acceptance criteria, edge cases and known risks.
Later, the build gets reviewed against it.
05
Route the work
Early on: Qwen for routine jobs, Claude Code for code, Grok for long context.
Weighed: cost, context and failure modes.
06
Delegate implementation
Claude Code builds against the approved spec.
Tracked: branch, build, commit and review state.
07
Verify and stage
Build, lint, test, stage, review the preview, then decide. The agent isn’t allowed to run vercel --prod.
Production waits for my sign-off.
08
Write back to memory
Notes, decisions and postmortems go back into the vault.
For next time. The next project starts from them.
MEMORY.md flat file · Obsidian linked vault · QMD semantic index · Wiki Layer session retrieval
05 · Memory
Memory took four tries.
Romulus started with one MEMORY.md. Within days it was a notebook with no chapters.
Then an Obsidian vault of daily notes, project notes, decisions and wikilinks.
QMD added search by meaning. The wiki layer holds notes Romulus checks before work.
Sessions end. The vault stays, so the next session can pick up the work.
Flat Files
MEMORY.md, project files, daily notes. Too flat.Obsidian Vault
83 files moved over at that stage, with wikilinks.
QMD Search
340 files indexed at that stage, searchable by meaning.
Wiki Layer
27 sources at that stage, checked before sessions.
Without notes, each session started from scratch.