I’ve been using AI to write production content for months. Tweets, blog posts, drafts that go out under my name.
For a while, everything it produced sounded like AI wrote it.
You know the type. Sentences that sound smart but say nothing. Snappy triads like “fast, efficient, reliable.” Hedging everywhere. “It’s worth noting that…” No. It’s not worth noting. Just say the thing.
The obvious fix was better prompts. More detailed instructions. Longer system messages. “Write in my voice. Be conversational. Don’t sound like AI.”
Didn’t work. The output got marginally better, then drifted right back. Better prompting is the writing equivalent of “just work harder.” The real leverage was somewhere else.
The System Around the Model
After enough failed drafts, I started building a system around the model. Skill files with explicit rules, non-negotiables, and review workflows.
The first version was basically a prompt with voice notes. It fabricated quotes. It produced generic startup platitudes. It drifted from positions I actually hold. One draft had a fake scene at a “YC dinner” with a founder quote that never happened. Completely invented. Confident-sounding. Wrong.
So I started adding guardrails. Not abstract principles — specific rules written in response to specific failures. Thirteen-plus iterations on a single skill file, each one patching something that broke.
What I found: the guardrails that prevent slop aren’t complex. They’re just specific.
Rules That Survived
What stuck after months of iteration and pruning.
Provenance tracing. Every claim, quote, or attributed advice must trace back to known context. If it can’t, rewrite it as a hypothesis or ask for one clarifying detail. This single rule killed the fabricated-founder-quote problem overnight.
Anti-fabrication policy. Never fabricate founder advice, named attribution, or strong positions. Never emit unsourced scene+quote+number claims in a single line. “At a YC dinner, a founder said…”? Banned unless it actually happened.
Belief alignment. Less obvious one. The AI would produce confident claims I’d never endorse — positions that sounded reasonable but weren’t mine. So I added a gate: don’t output claims the user is unlikely to hold based on known voice, worldview, and context. If uncertain, frame it as a question.
This might be the most underrated guardrail in the whole system. Fabricating facts is bad. Fabricating your beliefs is worse. A wrong fact, you catch. A subtly wrong belief just ships.
Graceful degradation. When evidence is thin, don’t fake it. “I’ve seen this happen” instead of “This is what happens.” Match confidence to evidence rather than inventing authority to fill the gap.
Anti-slop pattern matching. An explicit kill list: em-dash overuse, “it’s not X, it’s Y” parallelism, snappy triads, hedging phrases like “it’s worth noting,” filler paragraphs, words like “delve,” “unpack,” “landscape,” “multifaceted.” If it reads like AI, kill it.
Staged review. Different quality contracts at different stages. Brainstorming gets loose. Final drafts get a four-pass review: structure, clarity, voice, integrity. The integrity pass checks for unsupported specifics and converts uncertain claims into scoped statements.
The Revert
Here’s the part most “how I built my system” posts leave out.
I built a 10-gate quality system. Reader and Promise Gate. Title and Subhead Gate. Structure Gate. Anti-Slop Gate. Evidence and Integrity Gate. Density Gate. Ten checkpoints a draft had to clear before shipping.
Built the whole thing and rolled it back the same day.
Too heavy. The friction wasn’t proportional to the quality. Writing went from a creative act to a compliance exercise. Output was technically cleaner but felt more sterile — like the gates were creating a different kind of slop.
The guardrails that survived were the ones embedded in the skill policies. Short, specific rules that fire during generation. Not a 10-step post-hoc audit. Over-engineering quality gates is its own failure mode.
What I’ve Seen Work
Not presenting this as a framework. But after months of building and pruning, a few patterns held up.
Inline rules outlast reference docs. A provenance policy embedded in the skill file works. A standalone checklist the model is supposed to consult mid-generation doesn’t. The closer the guardrail is to generation, the more likely it fires.
Specific bans outlast general principles. “Don’t sound like AI” never reliably worked for me. “Never emit unsourced scene+quote+number claims in a single line” did. The more concrete the rule, the more it holds.
Every rule that stuck exists because something specific broke. The belief alignment gate exists because the AI produced a take I’d never endorse and I almost shipped it. The anti-fabrication policy exists because it invented a founder quote with a scene and a number, convincing enough that I had to stop and think about whether it actually happened.
Light outlasts heavy. The 10-gate system failed. Five to seven inline rules with a lightweight staged review didn’t. More process is not more quality.
This post was written with AI assistance, using the same guardrail system described above. The provenance policy, the belief alignment gate, the anti-slop checklist — they all ran during drafting. Whether it reads like a person with real experience or like a system describing itself is yours to judge.