Recently, I rebuilt an AI workflow that had once worked extremely well. It was designed to ghostwrite for a specific executive. The system included examples of his strongest writing, details about his perspective, preferences about tone and structure, phrases to avoid, and instructions for turning rough ideas into finished articles and social posts.

For months, the results were excellent. Then they began to deteriorate. The writing became robotic. It followed the rules, but it no longer sounded natural. I tried coaching it through the problem. Each disappointing output led to another correction, another example, or another instruction.

The quality kept declining, and the workflow became painfully slow. Its context window had ballooned. Every response required the model to work through a growing collection of instructions, reference files, examples, and previous corrections. Something that had once felt fast and intuitive now took forever to produce worse work. Eventually, I stopped trying to repair it and rebuilt it.

I preserved the executive's actual knowledge, point of view, and best writing samples. But I removed most of the behavioral instructions that had accumulated around them. The rebuilt system used smaller, focused SKILL.md files containing reusable instructions for specific tasks. The AI could load them as needed without having to process every instruction each time.

The rebuilt workflow produced better writing and ran considerably faster, too.

I'm still not sure exactly why the old workflow started to deteriorate. Model updates may have played a role, especially since the instructions had been developed over several months. Either way, the experience made me realize how easily an AI workflow can become outdated as the technology evolves.

Old Instructions Can Outlive the Problems They Solved

While researching the issue, I found that Boris Cherny, the creator of Claude Code, has suggested a similar reset.

In an interview with Y Combinator, Cherny recommended that Claude Code users periodically delete their CLAUDE.md instructions and skills, then test what the current model can accomplish without them.

Anthropic has taken a similar approach with Claude Code, removing instructions that were originally written to address limitations in earlier models. Its engineers tested the simplified system to see whether those instructions were still necessary. For models including Claude Opus 5, Anthropic removed more than 80% of Claude Code's system prompt. The company reported no measurable decline on its internal coding evaluations.

Many of the deleted instructions had been written to compensate for weaknesses in earlier models. Once the models improved, those guardrails were no longer helping. Some were getting in the way.

My ghostwriting workflow had accumulated the same kind of baggage. Every individual correction had seemed reasonable. I wanted the writing to be less formal, the openings to feel less repetitive, the paragraphs to be shorter, and the arguments to have enough context.

Together, those rules created a system that had become increasingly difficult to manage. Some overlapped, while others pulled the writing in conflicting directions. Trying to coach the model by adding more instructions only made the problem worse.

More Context Is Not Always Better

Anthropic describes the discipline of deciding what a model should see as context engineering.

Every instruction, example, reference file, and previous message placed in a model's active context competes for attention. More context can provide useful information, but it can also make the important material harder to identify. It may increase latency and cost while reducing the consistency of the output.

For marketing teams building AI workflows around content production, campaign planning, audience research, and reporting, this is a practical concern. Instructions accumulate as teams refine their processes, accommodate new requirements, and correct disappointing results. Over time, those additions can make a workflow harder to maintain, even when each change seemed necessary.

The executive's opinions, biographical information, and strongest writing samples still mattered. So did security controls, approval requirements, and factual boundaries. The material that needed to be questioned was the scaffolding built to manage an earlier model's behavior.

Developer advocate Jim Bennett draws a useful distinction between instructions that explain how to use a tool and those that encode how a company operates. Tool instructions can become outdated quickly. Proprietary knowledge and business processes should be preserved.

Smaller SKILL.md files helped me maintain that separation. The rebuilt system could pull in the instructions needed for a particular task without forcing the model to process the entire history of the workflow every time it responded.

When to Rebuild an AI Workflow

If a project, skill, or agent becomes slower and its output gets worse, resist the instinct to add another paragraph of instructions.

Save the existing version. Remove most of the behavioral rules and test the task again using the current model. Preserve the knowledge, examples, and safeguards that are genuinely important. Make everything else prove that it still deserves to be there.

A full reset is annoying. It can be painful to dismantle something you spent months building. But continuing to patch an outdated workflow can waste even more time. Sometimes, in the era of AI, you have to take a system down to the studs and rebuild it around what the technology can do now.

The short-term work is frustrating. The improvement in speed, simplicity, and output can make you grateful you finally did it.