7 feedback loops for self-improving AI content workflows

You’re already giving your contented workflows feedback. Every clip you edit a draft, hole the aforesaid awkward transition, aliases reword a vague heading, you’re providing corrections that an loop loop tin capture, truthful the adjacent tally starts person to what you’d approve.

I tally these loops crossed articles, LinkedIn posts, video scripts, and landing page copy. When an edit shape shows up 3 times crossed abstracted pieces, the strategy proposes an update to its instructions. I o.k. it, aliases I don’t. Either way, I nary longer manually update supplier docs each clip output drifts successful the aforesaid direction.

Here are 7 loops, from little improvement done post-publish performance. I usage Claude Code, but these structures activity successful immoderate supplier framework. You don’t request each of them. If you’re building your first, commencement pinch the value gross (loop 3). Otherwise, commencement wherever your workflow keeps breaking.

1. The upstream select loop

Most loop happens aft generation. This loop runs earlier penning begins.

It’s worthy the other measurement because a anemic perspective is the astir costly nonaccomplishment successful the pipeline. By the clip it reaches a vanished draft, you’ve spent a afloat pipeline tally positive your ain reappraisal clip discovering what a strategist supplier could person told you upfront. 

I tally excavation connected angles I’m considering pitching to extracurricular publications, wherever a killed perspective costs nothing, and a bad transportation costs an editor’s trust.

The strategist supplier evaluates the little aliases perspective against defined criteria earlier thing is written and issues 1 of 3 verdicts:

• Pass: Proceed to penning aliases pitching, depending connected the workflow.
• Revise: Something circumstantial needs to alteration first. The perspective is excessively adjacent to a portion you’ve already published, the thesis is excessively wide to support, the taxable fits, but the assemblage is wrong, aliases the statement needs a impervious constituent you haven’t gathered yet.
• Kill: The perspective can’t beryllium fixed pinch revision. There’s nary original constituent of view, aliases the root to support it doesn’t exist. The supplier documents why, and the rationale is logged.

The termination log is wherever this loop pays off. After capable runs, it shows which perspective patterns consistently neglect without anyone reviewing individual verdicts.

Before you build this, define:

• Evaluation criteria: Original constituent of view, thesis strength, and assemblage fresh requirements.
• What triggers each verdict.
• Where verdicts and termination rationales get logged.

Dig deeper: How to build a Claude Code-powered 2nd encephalon for agency work

2. The retrieval refinement loop

In a modular pipeline, a investigation supplier retrieves sources, the writer uses them, and problems aboveground astatine the extremity erstwhile an editor flags claims that the sources don’t support. 

By then, the hole is expensive: An editor tin emblem an unsourced claim, but can’t nutrient the missing source. This loop adds a checkpoint betwixt investigation and writing.

In my article pipeline, the investigation supplier retrieves sources for the planned piece. Before the writer runs, a mapping supplier sounds the outline alongside those sources and asks 1 mobility per section: Does this grounds support the claims this conception needs to make? 

It scores each section’s sourcing spot connected a 1-10 scale. For immoderate conception beneath your threshold, it writes the follow-up hunt queries itself because it knows precisely what’s missing. Only past does the writer run.

The quality shows up successful the draft. A writer moving from sources that don’t rather support the planned claims produces hedges and generalizations. A writer moving from validated sources produces specific, defensible claims.

3. The value gross pinch a revision cap

One-shotting contented produces AI slop. Adding a value gross is the simplest fix. Instead of generating a portion successful the aforesaid discourse model and calling it done, a 2nd supplier reviews the draught against defined criteria, classifies what’s wrong, and sends it backmost to the writer for revision. The writer makes the corrections and returns the draught to the reviewer, who checks it again.

When I tally this loop, I springiness each supplier a cleanable discourse model and group a revision limit. A draught that won’t walk aft 2 rounds has a structural aliases sourcing problem that revision can’t fix.

The reviewer doesn’t person to beryllium 1 agent. I primitively had my editor grip fact-checking too, but combining the 2 jobs meant neither sewage done well. So I divided them.

A dedicated fact-checker now runs successful its ain cleanable discourse window, takes the draught positive each root it cites, and checks each 1 to corroborate the draught accurately describes what the root says, not conscionable that the nexus exists. Giving each supplier a azygous occupation made some amended astatine it.

To build your ain value gate, specify what each verdict intends for your content:

• Pass: Every declare is sourced, the portion matches your sound guide, and the building serves the argument.
• Flag: Fixable issues, for illustration an undefined term, a anemic opening, aliases a declare that needs a stronger source.
• Escalate: Something revision can’t fix, for illustration a bladed perspective aliases missing research.

Route thing that hits the revision headdress to a quality alternatively of letting it loop. Once the gross works, adhd definitive re-entry points truthful you tin driblet a coworker’s draught straight into the reviewer without moving the afloat workflow.

Dig deeper: How to move Claude Code into your SEO bid center

4. Rubric-based scoring and ensemble selection

A value gross tells you whether a draught passed. A scoring loop tells you why it didn’t and what would hole it.

Start pinch a rubric that your supplier will usage to cheque the content. The criteria dangle connected what you’re creating and the goal.

For example, my LinkedIn station rubric scores 10 criteria, including specificity and concreteness, original constituent of view, a azygous clear insight, and whether each statement avoids platitudes.

I besides built a rubric into an grant submission researcher utilizing the submission guidelines. It’s been astir useful for comparing submissions and pinpointing precisely what to fortify successful each one.

Score output against each criterion connected a defined scale, specified arsenic 1-10. For each criterion beneath your threshold, person the scoring supplier nutrient a circumstantial test alternatively of a vague judgment. Send that accusation backmost to the writer supplier for revisions.

Include a revision cap. If a criterion won’t adjacent the spread aft 2 rewrites, the problem is the perspective aliases the research. A draught stuck astatine a six connected specificity aft 2 revision cycles is missing thing that doesn’t beryllium successful the root material. Scoring it again won’t help.

You tin besides usage a rubric to judge respective pieces. Generate respective versions pinch different framings, past tally a judge supplier that compares them utilizing the rubric arsenic a guide. 

5. The adversarial situation loop

An adversarial supplier builds the strongest imaginable lawsuit against a portion of content.

After a draught is produced, the adversarial supplier attacks the thesis, the evidence, and the logic connecting them. The output includes each objection it tin support pinch reasoning.

You’re not asking for “this declare is unsourced.” You’re asking for “here is the strongest counterargument, present is the grounds for it, and present is wherever your logic doesn’t hold.”

Share the output pinch your writer agent, which has to reply each objection: fortify the portion aliases archive why the objection doesn’t alteration the argument.

This loop earns its support connected thought activity and sentiment pieces, wherever the statement is the product. I tally it connected my ain bylined articles earlier anyone other sees them. How-tos and explainers don’t person a thesis to challenge, truthful the value gross is enough.

If a practitioner pinch different acquisition could publication your draught and reasonably disagree pinch its cardinal claim, an adversarial supplier will aboveground that disagreement earlier your editor does.

Dig deeper: How a ‘client brain’ gives AI the discourse SEO activity needs

6. The diff-and-learn loop

Every loop truthful acold improves the portion successful beforehand of it. This 1 improves the pipeline itself.

My article generator runs this loop. By the clip a draught reaches me, it’s gone done a researcher, an outliner, a writer, aggregate editors, and a fact-checker. The workflow past saves 2 files: a Markdown type that stays stiff and a DOCX I edit and upload to WordPress. Once the portion is published, a diff supplier compares the stiff type pinch what I published, statement by line.

For this to work, frost the pipeline’s output earlier you reappraisal it, and ne'er edit that file. Make your edits successful a moving copy. Without the stiff version, there’s nary grounds of what the strategy produced and thing to comparison your edits against.

Once you’re done editing, the diff supplier classifies each quality by type:

• Language simplification.
• Tone shift.
• Structural reorder.
• Factual correction.
• Heading rewrite.

It keeps a count for each category. When a class reaches a period — for me, 3 aliases much akin fixes connected 1 portion aliases crossed respective — the loop proposes an update to the instructions for the pipeline shape responsible.

I o.k. aliases cull each proposal, and approved rules use automatically. Approving a norm the strategy caught earlier I did is easy my favourite infinitesimal successful immoderate of these loops.

The period is what makes this work: A hole that appears erstwhile whitethorn beryllium circumstantial to that piece, but 3 aliases much appearances bespeak a shape worthy encoding.

Two guardrails support this loop from going wrong. First, a quality approves each projected rule. Say you trim a statistic from 1 portion because it didn’t fresh that argument. Without an support step, the strategy tin move that azygous edit into a opinionated rule, specified arsenic “avoid statistics,” and use it to everything that follows.

The different guardrail is simply a imperishable location for diff results. If they reset pinch each piece, the loop won’t announcement that the aforesaid hole showed up crossed 4 different articles, and that accumulation is the full point.

The registry tin beryllium a spreadsheet, a JSON file, aliases a Markdown log. What matters is that it lives extracurricular immoderate azygous convention and persists crossed pieces. Have it track:

• Per fix: Which piece, which category, what the pipeline produced, and what you changed it to.
• Per category: Total count, really galore abstracted pieces contributed, and whether the shape is still being watched aliases has already go a rule.

Dig deeper: 6 contented audit workflows to build successful Claude

7. The performance-feedback loop

Once a portion is published, hunt capacity is the verdict that counts. Most teams cod that verdict for reporting and stop. This loop puts it to work: What hunt tells you astir published pieces should alteration the briefs you constitute next.

Set up a scheduled regular aliases supplier that pulls capacity signals play and flags pieces moving successful either direction:

• Signals: Rankings, click-through rate, impressions, and traffic, pulled done the Semrush MCP aliases API, aliases from Google Search Console via a BigQuery connector.
• Cadence: Weekly, truthful you drawback activity while there’s still clip to respond.
• Flags: Pieces underperforming your ain akin content, rankings that ne'er materialized, positions a portion utilized to clasp and lost, and pieces outperforming expectations. The winners matter arsenic overmuch arsenic the losers because they show you which decisions to repeat.

A portion tin walk each soul gross and still neglect successful search. For each flagged piece, springiness an supplier the original little and the capacity data, and person it reply 1 question: Knowing really this portion performed, what would you alteration astir the brief?

A portion that ne'er ranked, while akin pieces did, usually had an perspective problem: It entered a speech wherever you had thing caller to say. A portion ranking for queries it ne'er targeted answered a different mobility than the 1 the little asked.

A be aware arsenic you group this up: Don’t publication a falling click-through complaint unsocial arsenic failure. AI answers person pushed click-through rates down crossed search, truthful comparison each portion against your ain akin content, not past year’s benchmarks.

Then make the instruction permanent. Add it to the strategist agent’s information criteria and the termination log truthful the adjacent little starts pinch everything this portion conscionable taught you.

Build for the nonaccomplishment mode you’re seeing

I created astir of my loops because I recovered myself making the aforesaid corrections. At immoderate point, I started asking why the strategy wasn’t catching them. That mobility is usually the little for the adjacent loop to build.

If you find yourself perpetually editing retired the aforesaid AI tells aliases asking Claude why it did thing again contempt you telling it not to, see whether a feedback loop could prevention immoderate of your sanity and amended output.