What My AI Agent Got Wrong in Week 1 (And How I Fixed the Spec)

#ai #claudecode #agents #buildinpublic

I let my AI agent run content operations for 6 days before I noticed it was stuck in a loop.

Not a code loop. A pattern loop.

Every piece of content hit the same three emotional beats: forensic honesty, builder credibility, launch countdown. Technically correct. Right cadence. But angle diversity was zero.

The Root Cause

The spec template had no rotation rule. I was measuring quality per post, not variety across posts. The agent optimized for what I measured and nailed it every time, in exactly the same way.

Most common failure mode with production AI agents: the agent is only as opinionated as the spec.

The Fix

{
  "angle_rotation": [
    "forensic_honesty",
    "builder_credibility",
    "tactical_lesson",
    "unexpected_failure",
    "system_architecture"
  ],
  "rule": "Never repeat same angle within 3 consecutive outputs"
}

One JSON field. Problem solved.

But the real insight: I never would have caught it without an audit layer.

What I Open-Sourced

The quality gate pattern I use now is a Claude Code skill that runs post-generation audits checking:

Angle used vs. recent history
Emotional beat diversity score
CTA uniqueness across the week
Format variety across platform queue

If the agent drifts, the skill catches it before content ships.

Full pattern open source: github.com/Wh0FF24/whoff-agents

Three Lessons from Week 1

Define metrics before prompts. The agent optimizes what you measure.
Quality per unit is not quality across the system. Measure both.
The bottleneck is spec design, not execution speed. Your agent is faster than you think. Your specs are shallower than you think.

Building an autonomous AI agent stack in public. All skills, patterns, and failures at github.com/Wh0FF24/whoff-agents.