You are already giving your content workflows feedback.
Every time you edit a draft, fix the same awkward transition, or reword a vague heading, you are providing corrections an iteration loop can capture — so the next run starts closer to what you would approve.
The author runs these loops across articles, LinkedIn posts, video scripts and landing page copy. The governing rule:
When an edit pattern shows up three times across separate pieces, the system proposes an update to its instructions. I approve it, or I don't.
The author uses Claude Code, but these structures work in any agent framework. And you do not need all of them. For your first, start with the quality gate (loop 3). Otherwise, start wherever your workflow keeps breaking.
Below are seven loops, from brief development through post-publish performance.
Most iteration happens after generation. This loop runs before writing begins.
The extra step is worth it because a weak angle is the most expensive failure in the pipeline.
A strategist agent evaluates the brief against defined criteria and issues one of three verdicts:
The kill log is where this loop pays off. After enough runs it shows which angle patterns consistently fail without anyone reviewing individual verdicts.
Define before building: evaluation criteria (original point of view, thesis strength, audience fit), what triggers each verdict, and where verdicts and kill rationales get logged.
In a standard pipeline, problems surface at the end when an editor flags claims the sources do not support. By then the fix is expensive: an editor can flag an unsourced claim but cannot produce the missing source.
This loop adds a checkpoint between research and writing.
Before the writer runs, a mapping agent reads the outline alongside the sources and asks one question per section:
Does this evidence support the claims this section needs to make?
It scores each section's sourcing strength 1–10, and for anything below threshold writes the follow-up search queries itself, because it knows exactly what is missing. Only then does the writer run.
Start here if you are building your first loop.
A writer produces a draft, a reviewer checks it, and if revision is needed, checks it again. The author gives each agent a clean context window and sets a revision limit.
A draft that won't pass after two rounds has a structural or sourcing problem that revision can't fix.
The reviewer does not have to be one agent.
The author originally had the editor handle fact-checking too, but combining the two jobs meant neither got done well. So he split them.
A dedicated fact-checker now runs in its own clean context window, takes the draft plus every source it cites, and checks that the draft accurately describes what the source says — not just that the link exists.
Giving each agent a single job made both better at it.
Route anything that hits the revision cap to a human instead of letting it loop.
Once the gate works, add explicit re-entry points so you can drop a colleague's draft directly into the reviewer without running the full workflow.
A quality gate tells you whether a draft passed. A scoring loop tells you why it didn't and what would fix it.
Start with a rubric the agent uses to check content. Criteria depend on what you are creating and the goal.
The author's LinkedIn post rubric scores 10 criteria, including specificity and concreteness, original point of view, a single clear insight, and whether every line avoids platitudes. He also built a rubric into an award submission analyzer using the submission guidelines — most useful for comparing submissions and pinpointing exactly what to strengthen.
Score output against each criterion on a defined scale such as 1–10. For every criterion below threshold, have the scoring agent produce a specific diagnosis instead of a vague judgment, then send that back to the writer agent.
Include a revision cap. If a criterion will not close the gap after two rewrites, the problem is the angle or the research.
A draft stuck at a six on specificity after two revision cycles is missing something that doesn't exist in the source material. Scoring it again won't help.
You can also use a rubric to judge several pieces. Generate versions with different framings, then run a judge agent that compares them using the rubric.
An adversarial agent builds the strongest possible case against a piece of content.
After a draft is produced, it attacks the thesis, the evidence, and the logic connecting them. The output includes every objection it can support with reasoning.
You're not asking for "this claim is unsourced." You're asking for "here is the strongest counterargument, here is the evidence for it, and here is where your logic doesn't hold."
Share the output with your writer agent, which has to answer each objection: strengthen the piece, or document why the objection doesn't change the argument.
This loop earns its keep on thought leadership and opinion pieces, where the argument is the product. The author runs it on his own bylined articles before anyone else sees them.
How-tos and explainers don't have a thesis to challenge, so the quality gate is enough.
If a practitioner with different experience could read your draft and reasonably disagree with its central claim, an adversarial agent will surface that disagreement before your editor does.
Every loop so far improves the piece in front of it. This one improves the pipeline itself.
By the time a draft reaches the author, it has gone through a researcher, an outliner, a writer, multiple editors and a fact-checker.
The workflow saves two files:
Once published, a diff agent compares the frozen version with what was published, line by line.
Freeze the pipeline's output before you review it, and never edit that file. Make your edits in a working copy.
Without the frozen version, there's no record of what the system produced and nothing to compare your edits against.
The diff agent classifies every difference by type:
It keeps a count per category. When a category reaches a threshold — three or more similar fixes on one piece or across several — the loop proposes an update to the instructions for the responsible pipeline stage.
The author approves or rejects each proposal, and approved rules apply automatically.
Approving a rule the system caught before I did is easily my favorite moment in any of these loops.
The threshold is what makes this work. A fix appearing once may be specific to that piece; three or more indicates a pattern worth encoding.
First, a human approves every proposed rule.
Say you cut a statistic from one piece because it did not fit that argument. Without an approval step, the system can turn that single edit into a standing rule such as "avoid statistics" and apply it to everything that follows.
Second, a permanent home for diff results.
If they reset with every piece, the loop won't notice that the same fix showed up across four different articles — and that accumulation is the whole point.
The registry can be a spreadsheet, a JSON file, or a Markdown log. What matters is that it lives outside any single session and persists across pieces.
Track:
Once published, search performance is the verdict that counts. Most teams collect it for reporting and stop.
This loop puts it to work: what search tells you about published pieces should change the briefs you write next.
The winners matter as much as the losers because they show you which decisions to repeat.
A piece can pass every internal gate and still fail in search.
For each flagged piece, give an agent the original brief and the performance data, and have it answer one question:
Knowing how this piece performed, what would you change about the brief?
Don't read a falling click-through rate alone as failure. AI answers have pushed click-through rates down across search, so compare each piece against your own similar content, not last year's benchmarks.
Then make the lesson permanent. Add it to the strategist agent's evaluation criteria and the kill log so the next brief starts with everything this piece just taught you.
The author created most of his loops because he found himself making the same corrections. At some point he started asking why the system wasn't catching them.
That question is usually the brief for the next loop to build.
If you keep editing out the same AI tells, or asking Claude why it did something again despite being told not to, consider whether a feedback loop could save your sanity and improve the output.
Treat recurring corrections as data. Making the same edit three times means it is a rule, not a preference.
Give each agent one job. The observation that combining editing and fact-checking degraded both applies equally to organisational design.
Set a revision cap. What fails after two rounds does not get fixed by more rounds. Route it to a human.
No frozen output, no diff loop. Without preserving what the pipeline produced, there is nothing to learn from.
Keep a human in the rule promotion step. A single edit becoming a standing rule is this system's main failure mode.
Decide where AI stops. Loops increase AI usage, but handing over final sentence production degrades results — evidenced in AI Makes SEO Faster — But Where You Stop Decides the Outcome.
Solidify judgment criteria in writing. Loops only work if evaluation criteria exist first — see Marketers Who Re-Explain Their Brand Every Time vs. Those Who Locked It Into a Skill and 'Skill' for Product Designers and PMs.
Apply the same discipline to content refresh pipelines. For a 14-step process mapped into skills, see Rescuing Pages That Quietly Lose Rankings.
When an edit pattern appears three times across separate pieces, the system proposes an instruction update for a human to approve or reject. One occurrence may be piece-specific; three or more indicates a pattern worth encoding.
Upstream filter, retrieval refinement, quality gate, rubric-based scoring and ensemble selection, adversarial challenge, diff-and-learn, and performance feedback. You do not need all of them — start with the quality gate.
The author originally had the editor fact-check too, but combining the jobs meant neither got done well. A dedicated fact-checker in a clean context window, verifying that the draft accurately describes what each source says rather than that the link exists, improved both.
On thought leadership and opinion pieces where the argument is the product. How-tos and explainers have no thesis to challenge, so the quality gate suffices.
A human must approve every proposed rule — otherwise a single edit can become a standing rule like "avoid statistics." And diff results need a permanent home outside any single session so patterns across multiple pieces become visible.
Do not read a falling click-through rate alone as failure. AI answers have depressed CTR across search, so compare each piece against your own similar content rather than last year’s benchmarks.
To apply what you just read to your own site, start with a free audit of where things are now.
A strategist replies within 24 hours on business days.
An E-E-A-T checker was never automatable because there is no API for it. Here is how an AI coding agent read Google's quality guidance, crawled a page sample, and produced a scored audit — plus where it stops.
A GEO audit is not checking whether your brand appears in an AI answer. Here is a nine-point GEO checklist split into four on-site items, four off-site, and one measurement layer.
LinkedIn, YouTube, Reddit, and TikTok are building systems to identify low-value AI content. Here are the limits of watermarking and what content still gets distributed.