Well currently it's helping me to con my obsidian vault to evolve my rag system, Make agent routes better to acquire knowledge for less token expenditure, investigate and make better the knowledge base, and make tags for them Also I'm trying it for outlook automatization. It's checking my routines, Important mails to establish some automation rules. Nevertheless, I'm still trying it. Will let you know how it goes. But it definitely looks promising. So far so good

Here is the complete, expanded breakdown of Anthropic’s exact prompting and architectural rules for Claude Fable 5: * **Target the absolute top of your difficulty range:** Fable 5 belongs to Anthropic's ultra-capable *Mythos* tier. Testing it on routine workloads or simple one-shot tasks is an explicit waste of budget and capabilities. Reserve it entirely for highly ambiguous, long-running, multi-step, end-to-end problems that would take a human hours or days to complete. * **Prompt for goals and autonomy, not dialogue:** The golden rule of Fable 5 is to stop over-scaffolding and micro-scripting every sub-task. State the final objective, define the strict boundaries of what it *cannot* do, establish clear success criteria, and then give it the autonomy to plan and execute its own path. * **Calibrate effort levels to enforce budget control:** Fable 5 introduces an explicit effort parameter (low, medium, high, xhigh). Never max it out by default. For everyday components or routine scripting, drop the effort to *low* or *medium* to prevent the model from over-deliberating, over-planning, or unnecessarily refactoring surrounding code. Save *high* or *xhigh* purely for capability-sensitive architecture. * **Enforce strict brevity and outcome-first summaries:** This tier is prone to excessive narration at higher effort levels. Steer it by embedding a short constraint: *"Lead with the outcome. Your first sentence after running must answer 'what happened' or 'what did you find' as a TLDR. Drop details that don't change subsequent actions."* Short brevity prompts are now highly effective. * **Deploy an external memory system:** Fable 5 shows a massive performance uplift when it can log lessons learned. Instead of carrying massive, token-heavy conversation histories forward, Anthropic recommends having the model read from and write to a simple, dedicated Markdown file to track pendekatan (approaches) and corrections across sessions. * **Build a client-side send_to_user tool:** For long, asynchronous agent runs, do not let the model end its execution turn just to give you a status update. Build a custom tool that takes a string input and prints it directly to your UI. This allows the model to stream progress reports without wasting massive token overhead on constant loop restarts.

@Guillermo L never tried it. Let me look into it and revert to you

You're not wrong at all. Cowork looks nicer, but Code is going to give you way more freedom. Cowork is great for quick stuff that looks clean, but when you're building something real like your Wealth Advisor, it starts feeling limiting. You're already doing a lot—checking screenshots, connecting to different apps, organizing data, making reports. That's where Code shines, even if it's not as pretty. Yeah, Code has a learning curve at first, but it's totally worth it. Once you get comfortable with it, you can fix things when they break, improve the workflow without fighting the interface, and scale everything without hitting walls. My take: go with Code. It's harder at the start, but you'll avoid a ton of frustration later. Cowork will eventually drive you crazy when you want to do stuff that's not in the menu.