I'm drafting a very old sci-fi novel of mine with Claude Code. Four scenes in. More excited about a creative project than I've been in years — and the reason isn't the speed. It's that the workspace is built to refuse. Setup: a folder called `writing-room`. Eight stages, from premise to compilation, each one a markdown directory the AI loads only when it's relevant. Compass, world, characters, structure, voice, writing, revision, compilation. The first rule, hardcoded in `CLAUDE.md`: > Before generating prose, always load `voz.md` and `padroes-prosa.md`. Without these two, refuse the writing task and ask the author to do Stage 05 first. Translation: the AI cannot draft a scene until I've locked in the voice. And `voz.md` was reverse-engineered from scenes I wrote by hand. The voice is mine. The AI only gets to extend it. There's also a file called `padroes-prosa.md` — 9 anti-AI-slope techniques. Verbalized sampling. Fragmentation. Character voice. Rare vocabulary. Every generated scene must apply at least 3, and the reviser uses the same file as a checklist. What this changes in practice: - I don't fight AI prose. I gate it. - Each stage loads minimum context. The AI doesn't drown in 200k tokens of worldbuilding to draft one scene. - After every scene, a `cronista` skill updates a canon file. Continuity stays cheap. - I'm the bottleneck on voice. I'm fine with that. The transferable bit, if you build with AI: The most useful thing your workflow can do is sometimes say no. Refusing to act without the right inputs forces you to produce those inputs — and that's where your taste enters the system. Without that gate, the AI averages you out. Toward the median sentence. The median plot beat. The median version of you. A friend of mine said that "in order to have a second brain, you need to have a primary working brain". I laughed: true enough. I wanted to build the gate first. Then let it write. And I'm loving it.

Today was a massive win. I had my meeting with the Head of AI for our global group, and it went beyond anything I had imagined. The Pitch: 132 Orders and a "Broken" System I had the chance to present a real-world challenge: Manually processing 132 sales orders in April. The workflow is a nightmare: Open each order, find the amount, cross-check it with an Excel sheet, invoice it, and repeat. To make it worse, there is a known bug in our D365 environment where the amount column simply shows "0" in the grid, meaning I can’t just export a list. It requires manual clicks. In a busy finance department, this takes days because of constant interruptions. I presented my workflow and explained how this concept isn't just for one task—it’s a framework for almost every repetitive monthly task we have. I knew from my previous Rebill Project that if I can automate the "friction," I can win back my time. The Result: Skipping the Queue When I told the Head of AI that this could turn a 3-4 day job into about 1 hour, his eyes lit up. Even though Claude Code is still stuck in corporate governance (it's currently with our CEO to decide on a global rollout), he didn't want me to wait. He immediately assigned me a Microsoft Copilot Studio license. These are highly restricted—usually, there’s a long waiting list, and if you don't use it for 30 days, you lose it. He bypassed the entire queue to get me started right away. Moving the Needle with IT To get "Copilot Cowork" talking to D365, I had to submit a technical IT ticket to enable the Model Context Protocol (MCP). I made sure to CC both the Head of AI and my own manager. The Head of AI jumped straight into the ticket with this comment: "I talked to Allan today. He has an idea to speed up a process in finance and save days of work... The use of the MCP server for this would help him very much. Open for a call if needed or any other help for the team."

I want to render cool visualizations of building and office floorplans...any idea what a good workflow or resource would be?

Locked in for the next 5 days only. Ends May 5th at 10:00 AM EST. No exceptions. 🎉 Premium: $27 → $14/mo 🎉 VIP: $97 → $67/mo The closest you'll get to our original launch pricing. We're doing this because the community has shown up for us, and we want to show up back. 🤝 🔥 Already a member? Read this carefully. To lock in the new rate, you need to: 1. Cancel your current plan 2. Resign under the new price That's the only way the system can apply the new rate. We have way too many members for manual refunds, so we can't refund anyone who just signed up at current pricing. But the savings stack month over month, so if you plan to stick around (and you should 😁), the math works out fast. 🚫 A few ground rules: Please do not DM myself or Jake about pricing, exceptions, or extensions. We love you, but we're a small team and we need to stay focused on building. Everyone gets the same window. Everyone gets the same deal. If you miss it, you miss it. We'll do more things for the community down the road. ⏰ The clock: 🟢 LIVE NOW 🔴 Locks May 5th, 10:00 AM EST - Premium gets you The Vault and Afternoon Tea calls. - VIP gets you The Drawing Room, High Tea, and bespoke folder builds from Jake himself. If you've been on the fence, this is the moment. 🚀 Tag a friend who needs to be in here. Let's make Cinco a movement. 🎊 🌶️🌶️🌶️

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.