Script Your Skills, Not Your Prompts

Free-form prompts are a trap for anything you run more than once. The moment you start asking an LLM to repeat the same multi-step logic — parse a config, call an API, reformat output — you're paying for the same tokens every time while getting slightly different results.

Here's what happens when you skip scripting:

• Drift: A step that worked Monday breaks by Friday because the model reinterprets instructions each run.
• Token bleed: Re-explaining rules eats into your actual context window.
• Secrets in logs: Pasting credentials into prompts turns into accidental exposure.
• Brittle calls: No retries means a single 429 kills the whole task.
• Debugging hell: Failures buried in conversations instead of clear stack traces.

How to fix it

1. Identify repeatable steps — same input, same output, every time.
2. Write a real script in a language with a test framework (Python, Node, Go — your call).
3. Move credentials into .env and load them at runtime.
4. Add retries and backoff around every external API call.
5. Unit test the script so regressions fail fast.
6. Review once, run forever — treat it like any production code.
7. Keep the model for judgment calls — deciding which script to run, not executing it.
8. Audit existing workflows — replace repeated WebFetch calls with direct API scripts where possible.

Why it matters

• Determinism: Same input, same output — no more chasing phantom behavior changes.
• Speed: Scripts run in milliseconds, not model round-trips.
• Cheaper: Logic in code doesn't get re-explained in every prompt.
• Testable: Unit tests catch regressions before they ship.
• Safer: Reviewed scripts don't improvise under pressure.

The rule of thumb: if a step doesn't need judgment, it doesn't need a prompt. Script it.