Thechrismarrano
Share
The workflow function is crazy. It built out a plan to launch 63 sub agents to audit a number of different functions within my agent workforce to find any gaps, vulnerabilities improvements, database issues, etc.. what it found was actually really really valuable. The question is was the 4.2 million tokens worth it? I’ll report back next week once we make these updates.
Opus 4.8 just came out and I’m already seeing a massive difference in both the reasoning, the explanations, the rationale and the results that I’m getting back. The future is looking crazier and crazier as these new models continue to roll out.. if you have any questions, shoot me a DM
05/09/2026
Prompts aren't docs. They're code. Treating them like markdown almost broke prod.
True story.
For a year, every agent prompt in my codebase lived in PROMPTS.md files. Loaded at runtime. Versioned in git. Looked clean.
Friday. 5pm. Production. Mia silently started missing campaigns.
Cause: someone edited PROMPTS.md, dropped a section header. The loader returned an empty string. No type error. No build failure. Just bad output, all weekend.
That's when it clicked:
If a model reads it at runtime, it's code.
Code needs:
✓ types
✓ tests
✓ PR review
✓ version pins
Markdown gets none of that.
I migrated every prompt to TypeScript:
export const MIA_PROMPT = `
You are a senior media buyer...
` as const;
Imported. Typed. Diffable. PR-reviewed. Breaks at build time, not 5pm Friday.
The rule I now live by:
→ If a model reads it, it's code.
→ If a human reads it, it's docs.
→ Never the same file.
This one rule killed an entire class of bugs.
4 more lessons from running an AI agency — plus the post-mortem template I now use after every incident — are in my Skool.
🔗 Link in bio.
Click here to claim your Sponsored Listing.
Category
Contact the public figure
Website
Address
506 SW Federal Highway
Stuart, FL
34994
