35 AI refactorings with an eval before and after

Pull one responsibility out of the main prompt and give it its own call.

Remove a call that does not change the result and fold its work into the neighboring step.

Replace a pile of rules with a few well chosen examples that show the behavior.

Move a tool's usage rule into its description, right where the model makes the choice.

Gather scattered prompt fragments into one template with a name, version, and variables.

Clearly mark what is a system instruction and what is content to be read.

Take examples out of the code into a versioned set, picked at request time.

Turn an IF this, do that written in the prompt into an explicit transition in code.

When the path is always the same, fix the order of steps in code.

Give a large task its own agent with a clean context and only the tools it needs.

Cap steps, time, and money for each run and define what happens at the limit.

Run calls that do not depend on each other at the same time.

Classify the input first and send it to a specialized prompt or agent.

Merge agents that only pass text along into one agent with the right tools.

Define a schema for the step's output and validate it before moving on.

Replace old turns with a summary and with facts stored outside the context.

Stop pasting documents into the system prompt and fetch only the relevant passage.

When the material fits in the window, send it whole and use prompt caching.

Make every tool return only the useful fields in a stable format.

Merge similar tools and remove the ones nobody uses.

Move calculation, formatting, and validation into a function the model calls.

Give each tool only the access it needs for its use case.

Break a generic tool (run_sql, call_api) into named, safe operations.

Publish the integration as an MCP server so any agent uses the same contract.

Pause the run and show the action to a human before paying, sending, or deleting.

Read outside content with a tool-less model and pass along only structured data.

Check generated code, SQL, or commands and run them in an isolated environment.

Define what happens when the model, the tool, or the provider fails.

Separate transient errors from permanent ones and retry only what can succeed.

Keep credentials inside the tool code, where the model never sees them.

Build a test set and measure the current version before any tweak.

Record every call, tool, prompt version, and cost in a browsable trace.

Reuse prompt prefixes and answers for equivalent questions.

Try a cheap model first and escalate only when confidence is low.

Persist state at every step so the run can resume after a crash, a wait, or an approval.