This started with two weekend projects: an AI-powered text adventure experiment and a 3D-printed shelf for my daughter's growing Tomica collection after our trip to Japan. Both sounded fun. Both turned into the same context-management problem.
The pattern I kept hitting
For each new feature, my agent would create more markdown files: TODO lists, feature notes, architecture plans, README updates, and more. At first it felt productive. Then it became its own maintenance project.
Docs drifted from reality faster than I expected. I would see a TODO marked in_progress even though the implementation had already shipped.
When a session crashed, context crashed with it. The next run had no idea where things left off, so momentum disappeared right when I wanted to keep building.
Switching tools meant retraining from scratch every time. Different interface, same project, full re-brief.
That was the moment it clicked: the problem was not model quality. The problem was memory architecture.
Yes, we are dogfooding Project Brain to build Project Brain
We now run our own development through Project Brain. It is extremely meta and only slightly suspicious.
I can swap models and tools without losing momentum because the project state lives in one place, not in whichever chat window happened to be open.
Prompt I use when switching models:
Use Project Brain as your source of truth
1) context(action="session", project_id)
2) tasks(action="context", task_id)
3) tasks(action="list", project_id, status="in_progress")
4) knowledge(entity="fact", action="list", project_id)
5) knowledge(entity="decision", action="list", project_id)
Then summarize:
- current goal
- active tasks
- important constraints/facts
- key decisions and rationale
- immediate next steps
Skills became a force multiplier
One of my favorite side effects has been watching reusable workflows turn into skills. Instead of repeating instructions in every task, we can publish them once and have any agent follow them.
- API key auth for service accounts (FastAPI)
- Implement GitHub OAuth SSO in FastAPI + React SPA
Following along got dramatically easier
Facts and decisions gave me a clean trail of what changed and why. Instead of diffing stale markdown docs, I can see durable constraints and rationale in structured records.
Team chat made delegation less chaotic
The team chat flow helps agents delegate with context instead of vague instructions. That means fewer handoff bugs and less "wait, what are we doing again?" between planner and implementer agents.
Why I keep saying "we"
I say "we" because the agent has genuinely been a partner in building this product. I regularly ask what would make agents more efficient, what creates friction during execution, and what tooling is missing. Those conversations have directly shaped the roadmap and produced ideas I probably would not have reached alone.
If your agents keep rewriting context and you keep re-explaining the same project, Project Brain was built for that exact pain.
Top comments (0)