We spent part of this week doing something that felt wrong: deleting instructions from the AI systems we build for clients. Not adding more guardrails, more skills, more scripted steps. Cutting them. The system prompt is the first thing operators reach for when an agent misbehaves, and it is often the wrong instinct.
The pattern we noticed in how Claude Code gets built
Every time Anthropic ships a new Claude model, the team behind Claude Code goes back into the product's default system prompt and starts removing lines, not just adding them. The reasoning is simple. A lot of that prompt exists to correct for a specific model's blind spots. When a newer, more capable model ships, those corrections stop being necessary and start getting in the way. Instructions that made an older model behave well can make a newer model worse, because you are telling it how to do something it already knows how to do.
This lines up with what Anthropic has published on context engineering for agents: a system prompt should carry the minimum set of tokens needed for the task, not a maximal one. Every instruction you add is not free. It competes for attention with every other instruction, and it can crowd out the model's own judgment on tasks you never anticipated. Anthropic's guidance on context engineering and on building effective agents both make the same point from different angles: simple, composable instructions beat dense, prescriptive ones once the underlying model is strong enough to fill in the gaps itself.
Product overhang: the capability you already paid for
There is a concept worth stealing for how you think about your own tools: product overhang. It describes the gap between what a model can already do today and what you are actually letting it do, because your instructions, your workflow, or your habits are still tuned for a weaker model from a few months ago. You are sitting on capability you already have access to and not using.
The test for this is uncomfortable but cheap to run. Take your AI system's configuration, whether that is a CLAUDE.md file, a system prompt, a set of skills, or a stack of automation rules, and strip it back to almost nothing. Give the model a task slightly harder than what you'd normally trust it with. Describe the task, the guardrails, and what done looks like, then let it run before you intervene. If the result surprises you, your instructions were the ceiling, not the model.
Why this is not blanket advice
The people telling you to delete everything are usually building harnesses for a living: frontier labs, agent infrastructure teams, people running massive codebases through models every day. That is not most operators, and it is not us most of the time either. When we are producing a client deliverable, a report, or a piece of content, the parts of our instructions that encode facts, brand voice, file locations, and compliance boundaries are still doing real work. Deleting those does not unhobble anything. It just produces a worse, less-branded output.
The useful distinction is between context and choreography. Context is the stuff a model cannot infer: where your data lives, what your brand sounds like, which numbers are real and which are placeholders, what done means for this specific client. Choreography is the step by step script for how to get there: do this, then this, then this, formatted exactly like that. Modern models are good enough that choreography is frequently the thing holding them back, while context is still what makes an output usable.
How we apply this to client agent systems
Every system we build, whether it is a Second Brain, an automation layer, or a full AI Operations Agent, runs on a system prompt and a set of skills, exactly like Claude Code does. The temptation for anyone building these systems is to keep adding rules every time the agent does something unexpected, until the prompt is a wall of edge cases. That is how you end up with an agent that technically follows every rule and still produces stiff, unusable output.
Our rule now: instructions that encode business context stay. Instructions that encode a process the model can reasonably work out on its own get cut and re-tested. That includes how to structure a summary, how to sequence a research task, or how to draft a first pass at a follow-up email. We keep the brand voice, the client facts, the escalation rules, the guardrails around what the agent is not allowed to touch. We stop scripting the how. Clients keep 100% ownership of the code and files we build, so nothing about this audit locks anyone into us either.
This matters more once the agents run unattended. Our systems are built to work 24/7 without someone babysitting every step, which means the gap between what the model can do and what we let it do costs real hours every week it goes unnoticed. Most of the operators we work with are already losing 10+ hours a week to admin that an agent could absorb. Over-scripted agents recover only part of that time, because they choke on the tasks nobody wrote a rule for.
The quarterly system prompt diet
Here is the version of this we now run on our own systems and recommend to clients, on a quarterly cadence or any time a materially better model ships:
- Duplicate the live system prompt, skills, and automation rules before touching anything, so you can roll back in minutes.
- Strip the copy down to business context only: facts, file locations, brand voice, hard constraints. Remove step by step instructions for how tasks get done.
- Run your five hardest real tasks from the last month through the stripped-down version and compare output side by side with the current production version.
- Where the stripped version wins or ties, keep it stripped. Where it loses, add back only the specific instruction that closed the gap, not the whole original block.
- Re-test after the next model upgrade. What earns its place today may not in three months.
This is the same audit we run inside a luup assessment, a fixed €999 engagement that gets credited straight into the build if you move forward, and it is one of the first things we check when a client's existing bot or workflow feels stiff instead of useful. If you want a broader read on where the operational leaks are before you rebuild anything, our revenue leak heatmap is a faster starting point than guessing.
Where instructions still earn their keep
| Keep in the prompt | Let the model figure out |
|---|---|
| Client and brand facts (names, tone, pricing rules) | How to structure a first draft |
| Data locations and system boundaries | The order of steps within a task |
| Hard guardrails (never send X, never touch Y) | How to phrase a routine follow-up |
| Escalation rules for edge cases | Which research angle to pursue first |
The left column is the part of a system prompt Anthropic's own Claude Code memory documentation describes as durable project context, the kind of thing you would tell a new hire on day one. The right column is the part that should shrink every time the underlying model gets stronger, because a model capable enough to run an agentic coding harness unattended is capable enough to sequence a research task without a numbered script.
What to build this quarter
If you are running any kind of AI agent for your business, whether that is a support bot, a content system, or a full client deployment we have shipped, here is the highest-value move this quarter. It probably is not adding a new capability. It is auditing what you are already paying to run and finding out how much of the instruction set is quietly capping it. We have started every new client build this way, because the first version of a system prompt is almost always written for a weaker model than the one that will actually run it by the time the system ships.
None of this means walking away from structure entirely. It means being honest about which parts of your instructions are protecting a fact and which parts are just protecting a habit. The first kind should stay. The second kind should get tested every time a new model shows up, because the gap between what your agent could do and what it is doing is the cheapest capability you will find all quarter.