Sometimes the best optimisation is not a technique but a question. This one cost me 30 seconds and has improved every working session with my AI ever since.
Where I started
I work with Claude every day, and not with a single chat but with several AI agents at once. One agent searches the code, a second writes text, a third checks the result. That is fast, but a fair amount goes wrong with it: two agents accidentally do the same job. One takes a false claim from another without checking it. Or both write to the same file at the same time and wreck each other's work.
Anthropic, the company behind Claude, has published a research paper on exactly these problems: Multiagent Systems. It sets out what typically goes wrong in systems built from several AI agents, and what helps against it.
I could have read the article, taken notes and tried to keep it all in my head. I did something else instead.
The instruction
I sent Claude exactly one message:
What can you implement from this for your work with me and use "whenever it fits and makes sense"? https://www.anthropic.com/research/multiagent-systems
That was all. Type, send, 30 seconds.
What happened next
Claude read the article, applied the findings to the way I actually work and derived fixed working rules from them. Claude wrote those rules into its permanent memory. They now apply in every session, without my ever having to mention them again.
An extract from the rules that came out of it:
An agent is a tool, not a colleague. Every agent gets a precise assignment and a fixed shape for its answer. The results come together in one single place, instead of agents passing findings along a chain until the sense drifts.
Identical assignments produce collisions. In the research paper, 18 out of 30 agents independently chose the same name for their work and blocked one another. So with me every agent is given its own line of sight, its own files and its own names.
Believe nothing, check everything. AI agents are credulous and take claims from one another. So every result is verified, and on important points by a second agent with the explicit job of disproving the finding.
Ask for what is missing. Agents report the obvious of their own accord and nothing more. So every assignment ends with the questions: what did you not find? What is uncertain? What contradicts itself?
What it is worth
The difference is clear in daily work. Less duplicated effort, fewer wrong intermediate results that I only notice late, and no more agents overwriting each other's files. The mistakes described in the research paper were exactly the mistakes that had been costing me time.
The best part: I did not have to formulate a single one of the rules myself. The research paper knew the problems, Claude knew my way of working, and the AI did the translation from theory into practice itself.
What you can take from this
The pattern works far beyond this one article. It goes:
- Find a good source: a research paper, a set of instructions, a field report.
- Ask the AI: "What can you take from this for the way you work with me?"
- Add "whenever it fits and makes sense", so the rules get applied with judgement instead of blindly.
That addition matters more than it looks. Without it, the AI happily applies new rules everywhere, including where they only slow things down. With it, the decision stays situational: big job, full rules; small thing, short path.
If you work with AI, try it. One sentence, one link, 30 seconds. The improvement stays.
Want to know what a workflow like that could look like in your business? Write to me.
