5 hours, one client, one workspace. And now I can't unsee it.

Last post I told you I externalized the persona. Vector store for the knowledge. Redis for the session memory. The system prompt shrank to the essentials.

That was step one.

Step two was realizing the persona isn't the only thing that doesn't belong in the prompt.

## The session ##

Paul, a client-friend, came in with a real demand. An ecotourism business in the mountains of Espírito Santo. Real briefings, real region knowledge, real brand voice, real stages — from a demand landing in WhatsApp to the Instagram post going up after the trip.

We sat down together. Five hours. By the end, the system was mapped, built, and tested.

What we built isn't an "AI app." It's a workspace.

ecoa-experiencias/

├── CLAUDE.md ← operating manual for the agent

├── CONTEXT.md ← task router

├── brand-vault/ ← identity + voice rules

├── knowledge-base/ ← regions, partners, biodiversity (grows over time)

├── shared/ ← seasonal calendar, languages, qualification criteria

└── stages/

├── 01-triagem/

├── 02-planejamento/

├── 03-proposta/

├── 04-execucao/

├── 05-avaliacao/

└── 06-publicacao/

Each stage knows what to load and what not to load. Triage doesn't load brand voice. Publication doesn't load qualification criteria. The agent only sees what matters for the task it's on.

Public repo: github.com/marcosaccioly/ecoa-experiencias

## What I can't unsee ##

This week I started a different project — a chapter for a 20-year commemorative book at the federal education center I work with. Different domain. Different audience. Different stakes.

And I caught myself building the same shape before writing a single sentence. CLAUDE.md at the root. Sections as folders. A skill for voice. Slash commands for the moves I knew I'd repeat. A contexto/ folder of source material the agent isn't allowed to invent past.

It wasn't deliberate. It was muscle memory.

## The shift ##

A system prompt is a static identity. A workspace is a working environment.

The agent doesn't need to "be" everything at once. It needs to know where to look for what it needs right now.

When you stop stuffing the prompt and start designing the room, three things happen.

The agent gets sharper — less context contamination from material it's holding "just in case."

The work gets composable — a stage hands off to the next stage through files, not through a running conversation.

And the non-technical human gets a seat. Paul edits markdown files. He doesn't prompt-engineer. The system meets him where he lives. That last one mattered more than I expected.

## What I'm taking ##

Externalizing the persona was the first crack. Once it opens, you can't close it.

You realize the prompt was never the right place for the workflow. Or the brand. Or the knowledge base. Or the templates. Or the policy on what the agent isn't allowed to make up.

All of that belongs in folders.

The system prompt should be the smallest part of the system.

0 comments