Most people customise Claude in one place: The prompt. The two surfaces that actually shape every session, you ignore. How Claude talks to you, and how you read what's happening. Both are configurable. Both compound across hundreds of hours. The framework I run, and how to set it up: Surface 1: The voice Claude has defaults. They're fine. They're not yours. I'm dyslexic. Walls of text are invisible to me. So my Claude has a non-negotiable response format, written into ~/.claude/CLAUDE.md: Every reply ends with a bold marker: → DECISION, → QUESTION, → NO DECISION NEEDED. I know what's required of me in under three seconds of skimming. Header tier = importance, not nesting depth. # is THE headline. ## is a main section. #### is an inline label. The size matches how load-bearing it is to read. `mono font` is reserved for paths, values, commands. Never for emphasis. The font swap is a signal. --- between major pivots like status to decision. The horizontal rule is the unmissable section break. Bold keywords inside sentences. Dyslexic eyes skim for bold anchors. Absolute paths only, so cmd+click opens them from inside the terminal. If you also use Codex, mirror the same rules to ~/.codex/AGENTS.md. Two harnesses, one voice. Surface 2: The interface The Claude Code statusline runs once per turn, costs zero tokens, can render colour. Mine shows: 47% · opus-4.7 · WORKSPACE · main* · 2🛠 · 02:41 Six fields: context %, model, workspace name, branch + dirty flag, dispatch worker count, session clock. Colour-coded by zone. Context % goes green under 50%, yellow at 50-75%, red at 90%. Branch yellow when dirty, red on conflict. Workspace gets a brand colour mapped from the path I'm in. The 50% threshold matches the rule already written into my voice config: write a session handoff at 50% context. Now I can see the threshold without asking. How to set this up Customise the voice. Open ~/.claude/CLAUDE.md. Write your own rules. Mine took an afternoon to draft and three weeks of iteration to lock down. Yours will start ugly. That's fine.

Even before the meetings happen, preparation is everything. I’ve been brainstorming how to take the principles we learn here and scale them for a global corporation (14,000 employees, 72 countries). This is the "North Star" for my journey toward becoming an AI Finance Manager. It’s a technical skeleton that ensures security, scalability, and—most importantly—the K.I.S.S. (Keep It Simple Stupid) model for the end-user. The Infrastructure (The Engine Room) Using the folder structure we’ve mastered, I’ve mapped out how a global hub would actually look "under the hood":Plaintext/ ├── CLAUDE.md # The Master Router (Direct OData Logic) ├── /identity # ACCESS CONTROL GATE │ ├── global_admins.md # Full access (Admin level) │ ├── DK.md # Validated users for Denmark │ └── [70+ countries].md # Global validation files ├── /library # THE COMMAND CATALOG │ ├── /global # Standard tools (e.g., /extract_ledger) │ └── /local # Country-specific tools (e.g., /dk_vat_report) ├── /core-logic # THE ENGINE ROOM │ ├── date_parser.md # Translates "MMYY" into OData parameters │ └── odata_templates.md # Master OData entity mapping ├── /countries # LOCAL CONTEXT & AUDIT │ └── /DK │ ├── audit-log.md # Record: Date | User | Command | Status │ └── local_settings.md # Local Legal Entity IDs / Currencies ├── /technical # SYSTEM RELIABILITY │ ├── mcp_config.md # OData read-only parameters │ └── health_check.md # System status triggers └── /schemas # EXCEL TEMPLATES └── ledger_standard.xlsx # Raw data structure The User Journey: From Teams to "Done" I want a user (let’s call him Lars) to get what he needs in under 60 seconds without ever touching a complex ERP menu. 1. The Request: Lars types /dk_vat_report 0126 in Microsoft Teams. 2. The Gatekeeper: The system checks /identity/DK.md. Lars is validated. 3. The Health Check: The system pings the MCP (Model Context Protocol). Connection is green. 4. The Data Pull: The AI uses the date_parser to see that 0126 means January 2026. It reaches directly into D365 via OData. 5. Delivery: The AI maps the data into a clean Excel template and uploads it to the chat. 6. Audit: The system silently logs the action for compliance.

"The haters, the doubters are all drinking champagne in the top deck of the Titanic… and we are the iceberg." — Casey Neistat A couple of days ago I rediscovered one of my favorite YouTube videos, posted 9 years ago by Casey Neistat. The premise is simple: everyone will tell you what you can't do. In life, in business, in relationships. In our case, with AI. F*&% that. When I joined this community a month ago, I didn't know what a markdown file was. Couldn't build a workspace or a system. Had no idea what "IDE" stood for, let alone how to build and deploy skills in Claude. That being said, I did the thing anyway. And now I'm building and solving things in ways I never imagined. If you're new: welcome. Do what you can't. If you've been here a while: keep doing what you can't. This is one of the best communities on the internet. Grateful to be in it.

​After a high-level meeting with the Head of AI where everything felt possible, I’ve just hit the "Corporate Wall." ​In my last post, I shared how the Head of AI personally backed my IT ticket to enable the D365 MCP server. But now, reality is sinking in. In a global corporation of 14,000 people, even a "Go" from the top has to pass through the gatekeepers of the legacy systems. ​The "Soon" Trap ​The manager responsible for our ERP system is hesitant. The argument? We are eventually moving from D365 to Business Central. ​But here’s the kicker: The rollout happens in waves, and our turn isn't until Q2 2027. ​That’s over a year away. In corporate time, "soon" is a dangerous word. It’s an excuse to stay in "manual hell" for another 12+ months. I refuse to let my team—and teams across 72 countries—waste thousands of hours clicking buttons that an AI agent could handle today. ​90% Done: The Solution is Already Built ​What makes this frustration even more real is that I’ve already done 90% of the work in Copilot Cowork. The logic is mapped, the agent is configured, and it’s ready for its first dry run. The only thing standing between a 4-day manual task and a 1-hour automated process is a single IT toggle. We aren't waiting for development; we are just waiting for a door to be unlocked. ​Expanding the Frontline ​The ERP manager sees this as a "finance request." He doesn't yet realize that I’m not just building a script; I’m building a global AI infrastructure for Finance. ​I’ve already started reaching out to colleagues in other countries to build a coalition. The feedback from Germany was clear: ​"It's definitely a really cool idea. I hope they react quickly so you can start development. If you need any support or testers, feel free to reach out. We'd love to help." ​ The Next Move ​On Monday, I’m meeting with the Head of Finance for the Nordics to get even more weight behind this. Then, I’m going back to the ERP manager. ​My mission is to show him that this isn't an "IT project" that adds to his workload. It’s a transformation that makes the entire organization ready for the future. Whether we use D365 or Business Central, the AI logic and the agentic workflows I’m building now will be the foundation.

Hi all. Quick background - The organisation I work for recently signed an agreement with Anthropic and last week we got access to their most recent models (Haiku, Sonnet and Opus). We’re predominantly a Microsoft house so have been using OpenAI this whole time and I have a Copilot licence. With everything I’ve learnt in this community, it’s been the biggest frustration not having access to tools that I can leverage to build ICM structure. I was trying to build my own if I’m honest. This week that changed as getting the Anthropic models also gives me access to the Copilot Cowork ‘Agent’ (frontier, and not just an agent 🤦). I now have something that can interact with my folder structure and today I completed testing of a workflow for creating proposals off the back of receiving an RFP - a process that usually takes 2-4 people a working week (5 days) of collaboration, research, meetings, etc, but took me just over 1 hour to have a fully branded working draft, off of the prompt “A new RFP has just come in - see attached”!!! (Screenshot for proof) We have a pretty rigorous process when we receive RFPs, so to be clear exactly what’s going on: 1. Validate the RFP against a number of criteria first to establish its something worth our time responding and bidding for. 2. Go through a research stage to gather information about the client, the industry, competitors (known or potential), case studies etc. 3. Analyse all research with additional context of our relationships and previous opportunities or projects if they exist. This additional data is all store in a CRM that has limited access to only senior people. Artefacts are created here with matrices of competitor comparison analysis, value wedges and our approach based on the case studies. 4. Map out our value proposition comparatively to potential competitors and develop the story we want to tell. (All this before we start building out the proposal and embed this information into it) 5. Build out the proposal. There is a template but it’s extensive and requires shaving down quite often. This is also done by junior members more often who are just pulling information in without understanding what it all means. 6. Check that branding guidelines are in place throughout and consistent. 7. First draft is now ready for senior review and changes