The AI instructions your team wrote last year may now be making your results wor
By Briancnewman · October 29, 2026 · Curated by George's Blog
The AI instructions your team wrote last year may now be making your results worse.
That is the argument in a piece AI Business Insights shared this past week. The author spent months building large folders of context files, style rules, and custom skills for Claude, then reversed course.
Boris Cherny, who created Claude Code, is quoted saying: "Every 6 months, delete your Claude MD. Delete your skills. Delete your hooks. See what the model does, and it might surprise you." The author also says Anthropic's own prompting guide warns that skills written for older models "can degrade output quality."
The reasoning makes sense to me. Older models needed detailed hand-holding. Newer models read intent better, so old rules can turn into noise that pulls the model toward problems you already solved.
Why should telecom leaders care?
Operators are building AI into customer care, network operations, and field work. Nokia and NVIDIA this week talked about extending governed AI agents across network operations. Each of those deployments carries prompts, rules, and playbooks written for a specific model at a specific time.
So, I would add three habits to any operator AI program:
1. Review prompts and agent rules on a schedule, at least every six months or with each model upgrade.
2. Test with less. Run the same tasks with a stripped-down setup and compare results honestly.
3. Ask the model to restate the goal before it starts. A misunderstanding is cheapest to fix before the work begins.
I want to be careful here. This advice comes from one practitioner's experience with one vendor's models, and I have not seen hard data on how much old instructions hurt results. Yet the cost of testing is small, and the cost of an agent following stale rules in a live network is not.
How often does your team revisit the instructions behind its AI tools?
#Telecom #AI #AIAgents #NetworkOperations #GenAI