How I give Claude Fable the right context to solve problems

Claude Fable is the flagship of the Claude 5 family: Anthropic's biggest, smartest, most expensive model. They're calling it a step change over the previous generation, which was already very good. Whether it really is a step change, we'll see. What's not in question is that a model this size carries an enormous amount of world knowledge, and with it, better intuition on hard problems.

The catch: capability follows context. Ask a brilliant model a context-free question and you get a generic answer, beautifully phrased. The model isn't the bottleneck anymore. What you feed it is.

Andrej Karpathy calls the work around this context engineering. He describes it as filling the context window with “just the right information for the next step.” Thariq, who works on Claude Code at Anthropic, put the prompting side even more plainly:

“thin prompts, thick artifacts + context, thin skills”

Thariq, Claude Code at Anthropic

That is a good description of what I'm doing here. The prompt itself is five sentences. The weight sits behind it: a year of conversations about the business and my life, plus a tool that can find the relevant pieces when I ask.

So this post is about how I feed it. Here's how I use MemoryPlugin to give Claude Fable 5 the context it needs to solve my actual problems, not the average internet user's problems.

Why context is the whole game

Not every query needs rich context. "Convert this to a for loop" doesn't care who you are. But many queries produce completely different answers with and without the right background, and the difference isn't subtle. Strategy questions, health questions, anything touching your business, your finances, your history: the answer that fits you depends on facts about you.

And here's the problem: the right context might span back years and touch a dozen areas of your life and work. There is no way you're sitting in front of Claude typing out a 3,000-word essay to build the perfect prompt. I wouldn't either, and I do this for a living.

The thing is, you've almost certainly already discussed most of that context with your AI tools at some point. It's sitting in old conversations. MemoryPlugin makes that history available to all your AI tools, including Claude Fable 5 through our MCP server. So instead of writing the context, I point the model at it.

Context engineering for the rest of us

Karpathy is talking about the whole system around a model: instructions, examples, tools, retrieved documents, state, and history. Most people using ChatGPT or Claude are never going to build that system. They still feel the difference when it is missing.

For me, the split is simple. The prompt says what I want now. MemoryPlugin carries what happened before. Claude uses recall to pull in what matters. That is context engineering, done for you. I can write a thin prompt because the thick context is already there.

If you're wondering why a million-token window does not solve this, I wrote the technical version here: Why a Bigger Context Window Is Not Memory

The prompt

Here's a real prompt, the kind I actually send:

How do I grow MemoryPlugin to $10K MRR and beyond? Use MemoryPlugin to recall details about the performance over the last year, marketing activities, changes to the product, changes to the market, things I've tried that worked and ones that didn't, how things have evolved over time. Clearly separate hypothetical ideas discussed in the abstract from things concretely true. Gather context about my current life across all relevant areas, then come up with a thorough report that I can implement to grow to my target as quickly as possible while work stays calm and enjoyable.

Five sentences. But each one is doing a specific job, and leaving any of them out degrades the answer in a predictable way.

Taking it apart

The goal comes first. "How do I grow MemoryPlugin to $10K MRR and beyond?" Everything the model recalls and reasons about afterward gets filtered through the lens of this goal. State it plainly, state it early.

Name the tool, name the action. "Use MemoryPlugin to recall." Both words are load-bearing. MemoryPlugin's chat history search is a tool called recall, and naming it steers the model to actually reach for it instead of answering from its training data. Models have lots of tools available; nudging them toward the right one is often necessary, and it costs you three words. This is the "thin skills" part of Thariq's formula. The instruction is tiny because the tool knows where the thick context lives.

Sketch the context you think matters. Performance over the last year, marketing activities, product changes, market changes, what worked and what didn't. I'm not writing the context, I'm writing the categories of context. The model takes it from there: it fires off recall queries, expands into related topics, maybe mixes in web searches. I gave it a map, not the territory.

Make it separate fact from hypothetical. This is the sentence most people wouldn't think to include, and it might be the most important one. Your chat history is full of ideas you discussed in the abstract, plans you floated and abandoned, hypotheticals you gamed out. An unsolved problem for every AI memory tool, ours included, is that recalled hypotheticals can surface looking like things that actually happened. On a high-stakes query, that leads your AI genuinely astray. Telling the model to clearly separate the two makes it noticeably more careful: it double-checks, it hedges where it should, and the answer comes back grounded in what's real.

Ask for thoroughness, once more with feeling. Fable is expensive. You want the full answer in one reply, not spread across ten follow-ups. So I say thorough, and you'll notice I also ask a second time to gather context "across all relevant areas". Repeating an instruction you care about isn't bad prompting, it's emphasis, and models respond to it.

State the secondary goals. "As quickly as possible while work stays calm and enjoyable." Without that clause I get the hustle-culture answer: run paid ads, post daily on five platforms, launch three funnels. Technically responsive, practically useless, because it ignores how I want to work. Secondary goals and preferences are the difference between an answer that looks impressive and one you'll actually implement.

What comes back

With the context recalled and the guardrails set, what I get isn't generic advice. It's a report that cites what I tried last spring, knows which channel actually converts for me, remembers what I said I'd never do again, and fits the way I want to run my work. The difference between that and the blank-slate answer is the difference between advice from a longtime friend and advice from a stranger at a conference.

The prompt is five sentences. The answer draws on a year of context.

This works everywhere Claude does: claude.ai on web and mobile, the desktop apps, and Claude Code, all through the same MemoryPlugin MCP connection. Same memory, every surface.

Key takeaways

• Context engineering is the real work - The model is no longer the bottleneck. A frontier model with no background still gives generic answers. This five-sentence prompt works because it can reach a year of history.
• Don't rewrite your history. Point at it - “Use MemoryPlugin to recall” tells Claude where to look and what to do. Recall beats retyping what you've already said.
• Give it a map, not an essay - Name the areas that matter, then let the model find the relevant details.
• Separate fact from hypothetical - Your chat history contains both. Tell the model to distinguish them, or abandoned ideas can come back sounding like decisions.
• State the life the answer has to fit - The best answer is not just technically correct. Include how you want to work, or the model will optimize for someone else.

If you want context engineering done for you, MemoryPlugin gives you one shared memory across 21+ AI tools, including ChatGPT, Claude, and Gemini. The context you've built follows you, whichever model you're talking to.