Bernat Sampera
← Essays

Context is Everything

7 min read · May 21, 2026

We carry too much context in our heads. The practice of offloading what you know into structured, referenceable files changes how you work with AI and with your team.

no contextno memory
Chat 1

Investigate support ticket SUP-123

I don't have access to that ticket. What system is it tracked in? What's the issue?

It's in Linear. User hasn't received payment - probably Supabase. Here are the logs: [14:31:02] ERR payment.webhook timeout [14:31:03] ERR edge_fn: payment-process

Based on those logs, the webhook timed out - likely a Supabase edge function failure. Check payment-process errors around 14:31.

new chat
Chat 2

What's the status of SUP-123?

I don't have context for that ticket.

You open a new chat. You type "investigate support ticket SUP-123." The AI asks: what system is it tracked in? What's the issue? You paste the logs, explain it's in Linear, mention the Stripe integration, describe your Supabase setup. Twenty minutes later, you get a useful answer.

Next day, different chat. "What's the status of SUP-123?" The AI has no idea what you're talking about.

This happens constantly. Not because the model is bad, but because you're carrying all the context in your head and expecting a stateless system to keep up. Every session starts from zero. Every conversation requires the same setup. You become the repeated explainer, not the AI, not the tooling, you.

The fix isn't a better model. It's a practice.

Context offloading

We're good at accumulating knowledge. How your company works, what tools you use, how your support flow runs, what happened with that one billing bug last Tuesday. All of it lives in your head, scattered across Slack threads, Notion pages, and muscle memory.

The problem is none of it is accessible to the tools you use every day. When you sit down with an AI assistant, you start from scratch. You re-explain your stack. You re-describe your workflows. You paste the same docs. And you do this over and over, across every session, every tool, every team member.

Context offloading is the practice of systematically writing down what you know in a structured way so it can be referenced later, by you, by your team, or by any AI tool you work with.

It's not note-taking. Notes are messy, personal, and hard to navigate. Context offloading is deliberate. You're building a knowledge layer that's structured for both humans and machines.

What this looks like in practice

Think of a directory. Call it context/. Inside it:

code
context/
  company.md
  integrations/
    stripe.md
    linear.md
  workflows/
    support.md
  tasks/
    TASK_123.md

company.md describes who you are, what you build, your stack, your conventions. Not a wiki page, a concise document an AI can read in seconds and immediately understand your world.

integrations/stripe.md captures how you use Stripe. Not the Stripe docs, your specific setup. Webhook endpoints, checkout flows, the edge cases you've hit. Same for Linear, same for whatever else you rely on.

workflows/support.md describes what happens when a support ticket comes in. Triage steps, severity levels, which systems to check, how to draft a response. The stuff that lives in someone's head and takes weeks to transfer to a new hire.

tasks/TASK_123.md is the live context for a specific problem. A user reported a billing failure after upgrading. Here's the charge ID, their email, the error code, what you've tried so far.

Each file is small. Each file is specific. Together they form a structured map of everything you know about your work.

Why structure matters

There's a temptation to dump everything into one big document. Your company info, your integrations, your workflows, all in one mega-prompt. It works at first.

Then it grows. You add more context, more integrations, more edge cases. The document becomes a wall of text where everything connects to everything else. Finding what's relevant becomes harder. Updating one section risks breaking the coherence of another. What started as helpful context becomes noise.

This is what unstructured context looks like at scale, a tangle where more information actually makes things worse.

Structured context scales differently. Each file is independent. You can add integrations/postmark.md without touching anything else. You can update your support workflow without worrying about your company description. A new team member can read company.md and workflows/support.md and be useful on day one. An AI agent can pull exactly the files it needs for a given task and ignore the rest.

Modular context doesn't degrade as it grows. It compounds.

The real shift

Here's what changes when you start offloading context.

You stop repeating yourself. That explanation of your Stripe setup you've given five times this month? Write it once. Every future conversation, with AI, with a teammate, with yourself six months from now, starts with that knowledge already in place.

AI becomes useful across sessions. When your context/ directory exists, a new chat isn't a blank slate. Point any AI tool at those files and it already knows your company, your integrations, your workflows. "Investigate SUP-123" stops being a twenty-minute setup and becomes a one-line instruction. The AI checks Linear because it knows you track tickets there. It queries Supabase because it knows that's where your payment logs live. It knows the support workflow because you wrote it down.

Your team gets aligned faster. Context files aren't just for AI. They're the clearest, most concise version of how your company actually works. Not the aspirational Notion wiki that nobody updates, the practical, working documents that describe reality.

Knowledge survives turnover. When someone leaves, their context usually leaves with them. With structured context files, the knowledge persists. It's not locked in anyone's head.

Context offloading is a habit

This isn't a one-time project. You don't sit down on a Saturday and write all your context files and call it done. Context offloading is a habit, something that happens naturally as you work.

You explain your billing flow to an AI during a debugging session. That explanation, cleaned up slightly, becomes workflows/billing.md. You onboard a new integration and describe the setup. That becomes integrations/newrelic.md. A tricky task comes in and you gather all the relevant details in one place. That becomes tasks/TASK_456.md.

The best part: conversations themselves can build context. You talk through a problem, the important parts get captured, and your knowledge layer grows without extra effort. Over time, the context/ directory becomes a living map of your work, always current, always accessible.

What this isn't

This isn't about any specific tool or product. You can do this with a folder of markdown files on your desktop. You can do it in a shared repo. You can do it in whatever system makes sense for your team.

The point is the practice, not the implementation. If you're working with information, and in 2026, who isn't, you should have a structured place where knowledge lives outside your head.

Define the will

AI is a tool. A good one, but still a tool. It executes intent, it doesn't generate it. The direction, the decisions, the "what should we actually do", that's yours.

The problem is that intent is hard to act on when it's vague. "Handle support tickets" means nothing to an AI without knowing your triage process, your integrations, your conventions. The gap between what you want and what AI can do isn't intelligence, it's definition.

Context offloading closes that gap. When your knowledge is written down, structured, and accessible, your intent becomes actionable. "Investigate SUP-123" works not because the AI got smarter, but because you defined your world clearly enough for it to operate in.

The better you define your context, the more you can orchestrate in one go. Not because you stepped aside, but because you made your will precise.

Context is everything. Start offloading it.

· · ·