Stop Writing Job Descriptions For Your AI Agent

Anthropic just deleted eighty percent of the instructions it gives its own coding agent. If you spent last month writing a rulebook for yours, you might want to sit down.

The Instructions You Wrote Last Month Are Already Wrong

Tom runs a seven-person plumbing outfit outside Manchester. Three months ago he built an AI agent to chase unpaid invoices, because he was tired of doing it himself at 9pm.

He did it properly. He wrote down every rule he could think of. Day 7, friendly nudge. Day 14, firmer tone. Day 30, mention the account is overdue. Day 45, cc the accountant. Never mention legal action before day 60. Never sound robotic. Always check the job was actually completed before chasing payment. Forty-one lines, numbered, in a document he was proud of.

It worked for about two weeks. Then it sent a stiff, legal-sounding message to a customer who was three days late. It forgot to cc the accountant, twice. On one invoice it followed the day-30 rule and the never-sound-robotic rule at the same time, and produced something that satisfied neither.

Tom's instinct was to add more rules to patch the gaps. That instinct is exactly backwards, and this week Anthropic told the world why, using its own product as the example.

What Anthropic Just Admitted About Its Own AI Agent

On 24 July 2026, Anthropic shipped Claude Opus 5, its new flagship model, priced at $5 per million input tokens and $25 per million output tokens, the same as its predecessor Opus 4.8. It leads on coding benchmarks like Frontier-Bench v0.1 and CursorBench 3.2, and on Zapier's AutomationBench, which measures whether a model can finish a real business task start to finish, Opus 5 passes about 1.5 times as many tasks as the next-best model at the same cost.

But the benchmarks were not the part that broke the internet. The same day, Anthropic engineer Thariq (@trq212) posted a thread that has since pulled in more than 2.6 million views, 25,000 bookmarks and 12,000 likes, pointing to a new Anthropic blog post titled "The new rules of context engineering for Claude 5 generation models." Its opening line: "We removed over 80% of Claude Code's system prompt for models like Claude Opus 5 and Claude Fable 5 with no measurable loss on our coding evaluations."

Read their own transcripts and you find Tom's exact problem, just at a bigger scale. Anthropic writes that its old system prompt, CLAUDE.md files and skills were giving the model conflicting orders in a single request, things like "leave documentation as appropriate" fighting "do not add comments." The model was spending its attention untangling contradictions instead of doing the work. Sound familiar?

How Do I Write Instructions For An AI Agent For My Business?

Give it a short brief in plain language: who it is working for, what "done" looks like, and the one or two boundaries that genuinely matter. Hand it good tools. Let it use judgment on the rest.

That answer will annoy some people, because it is not what most advice currently says. Ask an AI search engine the same question today and you will likely get told to write your agent's instructions "like a detailed job description for a new digital employee," complete with a role, a step-by-step workflow, named tools, an output format and a guardrail for every possible failure case.

That advice was correct for the last generation of models. It is precisely the pattern Anthropic just spent a blog post walking back. Rigid, exhaustive rulebooks made sense when models needed strong guardrails to avoid worst-case mistakes. Claude Opus 5 and Claude Fable 5 have better judgment, so the guardrails that used to help now mostly get in the way, the same way Tom's forty-one rules got in the way of his invoice agent doing the one thing he actually wanted: get paid, without sounding like a debt collector.

The Six Rules That Just Flipped

Anthropic's engineering post lays out, point by point, what changed between the old approach to Claude Code and the new one. The same shift applies to any AI agent you are running for your business, whether it lives in Claude, a support inbox, or a workflow tool.

Anthropic even shipped a command for this, /doctor, that checks a Claude Code setup for exactly this kind of bloat and contradiction. You do not need Claude Code to use the idea. You need to ask the same question of your own agent's instructions: which of these rules are doing real work, and which ones exist because something went wrong once, eight months ago, and nobody ever removed the patch?

Rewrite Your Agent's Brief In One Sitting

Here is the version of this Tom should have written from day one, and the steps to get any agent's instructions there.

1. Write the outcome, not the procedure. Not "on day 30, send message X." Instead: "get the invoice paid, or a clear reply, within five days of your message."
2. Cut every rule that exists because of one bad day. If a rule only makes sense because of a single edge case that happened once, delete it and trust the model to handle that case with judgment when it comes up again.
3. Move tool instructions onto the tool, not the brief. If your agent has a "send email" or "look up invoice" tool, put the usage notes in that tool's description. Do not repeat them in the main instructions.
4. Split anything longer than a page. If you have detailed guidance for a rare situation, such as a dispute over completed work, keep it in a separate reference the agent only opens when that situation actually comes up. Claude calls this a skill; the same idea works in n8n, Zapier, or a custom build.
5. Run your own version of the doctor check. Paste your agent's full instructions back to it and ask: "Which of these rules are you finding contradictory or hard to follow, and which ones do you actually need?" Claude Opus 5, and most current models, will tell you.

Tom's rewrite went from forty-one lines to six: "You chase unpaid invoices for a seven-person plumbing firm. Success means the customer pays or replies within five days of a message. Keep the tone friendly and human; never mention legal action before day 60. CC the accountant on anything over 30 days late. Use your judgment on wording and timing." One brief, one boundary that matters, one tool note, and the agent handling the rest the way a sharp new hire would after a five-minute conversation, not a forty-page manual.

What You Have After This

A week in, your agent stops making the strange mistakes that came from two rules quietly fighting each other. You notice it, because you stop getting the "why did it do that" moment.

A month in, you are comfortable handing it judgment calls you used to write rules for, because it has been making the small ones correctly without you watching.

By the time the next model generation ships, and it will, rewriting a bloated brief down to its outcome and its real boundaries stops being a clever trick you read about once. It becomes the first thing you do with any new agent, the same way checking a plumbing job before invoicing it is just what Tom does now.

We wrote before about giving your AI agent a goal instead of micromanaging it. Claude Opus 5's launch is the proof, straight from Anthropic's own transcripts, that this is not a style preference. It is what actually works as the models get better.

Same business. Same agent. A brief you can actually read in one sitting.