Automation, Repurposing and Custom GPTs

Most content teams I talk to have the same shape of problem. They’re producing decent primary assets, one or two a week, and they know the value sits in the fifteen derivatives nobody has time to make. So they buy a tool, paste a blog post in, get back a LinkedIn carousel that opens with “In today’s fast-paced digital landscape,” and quietly stop using it.

The failure isn’t the model. It’s that nobody encoded what good looks like for their particular audience before automating anything. AI content repurposing automation works when you spend the first two weeks building constraint, not output: a voice spec, a set of banned constructions, worked examples of your best derivative work, and a decision about which assets are worth the compute in the first place. After that, the automation part is genuinely boring, which is the point.

This page covers the parts I’d actually build if I inherited a five-person content team tomorrow, in rough order of payback.

Work Out Which Assets Deserve Repurposing

The instinct is to repurpose everything. Don’t. Repurposing a weak asset multiplies the weakness across five channels and costs you the same effort as repurposing a strong one.

Pull your last six months of content into a sheet with four columns: organic sessions, average engaged time, assisted conversions (last-touch is useless here), and inbound mentions or shares. Then rank. In most B2B programmes I’ve seen, the top 15% of assets carry 60-80% of the value. Those are your repurposing candidates. A post doing 40 sessions a month with 22 seconds of engaged time does not need a video script.

A practical threshold: only repurpose assets that clear both 90 seconds average engaged time and either 500 monthly sessions or one attributable pipeline touch per quarter. That usually leaves you with 8-15 assets from a year of publishing, which is plenty. Each one can plausibly yield a LinkedIn text post, a carousel, three short-form video scripts, a newsletter section, two answer-style FAQ blocks for the page itself, and a slide for sales.

Second filter, less obvious: repurpose by durability. A piece tied to a Q3 product release has maybe eight weeks of shelf life. A piece explaining how your buyers actually evaluate a category has three years. Build your automation around the durable set, because you’ll re-run it. I re-run the derivative pipeline on evergreen assets every four to five months with a “what’s changed since [date]” instruction, and the refreshed versions typically outperform the originals on LinkedIn because the examples are current.

The Voice Spec That Makes Everything Else Work

Before you touch a workflow tool, write a document the model can read. Not a brand guideline PDF. A working spec, 800 to 1,500 words, structured like this:

Sentence-level rules. Mine includes: no sentence longer than 30 words unless it’s doing list work. No “moreover,” “furthermore,” “in conclusion.” Numbers as digits from 10 up. British spelling, Oxford comma off. Second person for instruction, first person plural only when describing the company’s own actions.

Banned openings. This is the highest-leverage section. Ban “In today’s,” “As we all know,” “Imagine a world where,” “Let’s dive in,” “The truth is,” and any sentence beginning with a gerund that’s just throat-clearing. Also ban the triplet rhythm (“faster, smarter, cheaper”) because models reach for it constantly and readers have learned to skim past it.

Three paired examples. For each output type, give one real example you’d publish and one plausible-but-wrong version, with a line explaining the difference. “Wrong because it states the benefit instead of showing the mechanism.” Paired examples do more work than any amount of adjective-based description. Telling a model to be “punchy and conversational” produces nothing; showing it two LinkedIn posts and saying “this one, not this one, because the first names a specific number in line two” produces a lot.

Evidence rules. Mine: every claim needs either a number, a named tool, or a described scenario. No claim may be sourced from the model’s own knowledge. If the input asset doesn’t contain the evidence, the output says [NEEDS SOURCE] rather than inventing one. That single instruction cut my fact-checking time on derivatives by roughly two thirds, because the gaps are flagged instead of hidden behind confident prose.

Keep the spec in one canonical place and reference it from everywhere. When it lives in five prompts, it drifts in five directions.

Building the Repurposing Chain

Here’s the shape I use, with the tools I’ve actually run it in.

The reliable version is a Google Doc or Notion page as input, a Make scenario as orchestrator, and Claude via API as the model. Make over Zapier for this specific job because multi-step branching with conditional retries is much cheaper: a scenario doing 10,000 operations a month costs about £14 on Make’s Core plan, against considerably more in Zapier task pricing once you’re past a few thousand multi-step runs.

The chain runs in five steps, and the ordering matters:

1. Extract, don’t summarise. First call pulls out the structural elements: the central claim, every number with its context, every named example, the counterintuitive bit, and the strongest single sentence verbatim. Output as JSON. This step is cheap and it’s what prevents the derivatives from all being flat restatements of the intro.
2. Angle generation. Second call takes the extraction and proposes six distinct angles, each with a named audience and a one-line reason someone would stop scrolling. You review this. It’s the only human checkpoint in the middle, and it takes about four minutes.
3. Format-specific drafting. Parallel calls, one per format, each with its own prompt, its own three paired examples, and the voice spec attached. Critically, each gets only the extraction plus the chosen angle, not the whole original. Feeding the full post makes the model paraphrase; feeding the extraction makes it compose.
4. Adversarial pass. A separate call whose only job is to find the generic sentences. Prompt: “List every sentence that could appear in any company’s content on this topic. For each, quote it and say what specific detail would fix it.” Then a second call applies the fixes. This two-step beats asking for a self-edit in the same call, because a model asked to critique and revise simultaneously tends to declare its own work fine.
5. Route to a review queue. Not straight to publish. Drafts land in an Airtable base with fields for status, channel, scheduled date, and a human_edited boolean. Nothing publishes with that boolean false.

Per-asset cost on Claude Sonnet, running all five steps with a 2,500-word source and producing eight derivatives, comes to roughly 90,000 input tokens and 12,000 output tokens once you account for the voice spec riding along on every call. That’s pennies. Use prompt caching on the voice spec and paired examples and you cut the input cost by around 90% on the cached portion, which matters once you’re running 40 assets a month.

The step people skip is step 4, and it’s the one that separates output you’d publish from output that reads like everyone else’s. Budget for it.

Where Automation Actively Hurts

Some things should not be automated, and knowing which saves you from expensive cleanup.

Never automate the first paragraph of a primary asset. It’s the one place where a specific observation earns you the rest of the read, and models generate openings by pattern-matching to other openings. Write it yourself, then let the model work from it.

Don’t automate comment replies or community interaction. Beyond the obvious risk, the signal you get from writing replies yourself is one of the few unmediated channels into what your audience actually misunderstands. That’s raw material for next quarter’s content.

Avoid automating anything where the output goes out under a named individual’s byline without them reading it. I’ve watched this damage two senior people’s credibility: a client asked a follow-up question about a LinkedIn post and the “author” couldn’t answer it.

And be careful with fully automated internal linking. Tools that insert links based on keyword matching produce contextually silly results (“automation” in a sentence about factory automation linking to your marketing automation guide). Semi-automate instead: have the model propose five link opportunities with the exact anchor text and surrounding sentence, and approve or reject each. Takes 90 seconds per post and the relevance is dramatically better.

Custom GPTs as the Team’s Shared Judgement

A Custom GPT solves a different problem from the automation chain. The chain handles volume on a defined input. A Custom GPT handles the ad-hoc requests that arrive constantly and would otherwise route to whoever writes fastest.

The ones that earn their keep in a small team, in order of how much time they save:

The brief builder. Takes a topic and a target keyword, asks four clarifying questions, then outputs a brief in your house format: search intent, the specific angle, the three claims that must be evidenced, the internal links to include, and the two competitor pieces to beat plus what they’re missing. Loaded with 15 of your best historical briefs as knowledge files. This turns a 45-minute task into an eight-minute one and, more importantly, stops junior writers from receiving vague briefs.

The editor. Not a rewriter. Configured to return structured feedback only: flagged generic sentences, unsupported claims, voice violations quoted against the specific spec rule, and structural issues. It never produces replacement prose, because the moment it does, writers stop thinking and start accepting.

The repurposing assistant. The manual counterpart to the automated chain, for when someone needs a LinkedIn post from a sales call transcript at 4pm.

The build details matter more than the concept: how you structure knowledge files so retrieval actually surfaces the right example, how to write instructions that survive contact with a distracted user typing three words, and how to version a GPT that four people depend on. I’ve written that up properly in Building a Custom GPT for Your Content Marketing Team, including the instruction structure I’ve landed on after rebuilding these several times and the knowledge-file mistake that quietly degrades output quality.

One thing worth stating here: keep knowledge files small and curated. A GPT with 8 tightly chosen examples outperforms one with 60 documents dumped in, because retrieval over a large mixed corpus surfaces mediocre matches. Curate ruthlessly, and re-curate quarterly as your best work improves.

Measuring Whether Any of This Worked

Volume metrics will flatter you. Of course you published more. The questions that matter are whether the derivatives earned attention and whether the time you saved went somewhere useful.

Track four things:

Derivative engagement against your own baseline, per channel. In GA4, tag every automated derivative with a UTM containing a consistent identifier (utm_content=rp_[assetID]) so you can segment repurposed traffic from original. Compare engaged sessions per post, not impressions. If your automated LinkedIn posts run at 60% of the engagement of hand-written ones, that’s arguably fine at 8x the volume; if they’re at 15%, the pipeline is producing filler and you should fix the adversarial pass.

Edit distance. Rough but useful. If your editors are rewriting more than about 40% of every draft, the automation is costing time rather than saving it, and the fix is almost always in the paired examples rather than the model choice. Track it for a month by having editors note a 1-5 rework score in the Airtable record. Honest data beats precise data here.

Time reallocation. Log where the recovered hours went. In the teams where this has stuck, the hours moved into original research, customer interviews, and distribution work outside their own channels. In the teams where it hasn’t, the hours went into producing more of the same content, and results were flat.

Assisted conversion on repurposed touchpoints. Look at conversion paths in GA4 over a 90-day window and check whether repurposed assets appear as non-final touches. They usually do, and this is often the strongest internal argument for keeping the programme funded, because the direct-conversion numbers on a LinkedIn carousel will always look unimpressive.

Set a review date before you build, 60 days out, with a stated kill threshold. “If automated derivatives run below 25% of hand-written engagement at day 60, we stop and rebuild the prompts.” Teams that don’t write the threshold down keep bad pipelines running for a year because nobody owns the decision to turn them off.

A Realistic First Month

Week one: build the ranking sheet, identify your 10 repurposing candidates, and write the voice spec. Nothing automated yet. This week feels unproductive and determines everything downstream.

Week two: build the chain manually for two assets. Actually paste things between prompts yourself. You’ll discover three things wrong with your extraction step that you’d never have found by building the automation first, and fixing them in a text file is much faster than fixing them in a Make scenario.

Week three: automate what you just did by hand. Wire up Make or n8n if you’d rather self-host (n8n on a £5 Hetzner box is fine for this volume and gives you unlimited executions). Add the Airtable review queue. Run 5 assets through it and edit every output carefully, noting what you changed.

Week four: build the first Custom GPT, the brief builder, using the briefs you’ve written across the previous three weeks. Set your 60-day review date. Tell the team what the kill threshold is.

By day 30 you’ll have a pipeline producing somewhere around 60-80 reviewable derivatives a month from 10 source assets, at a model cost under £40 and a human review cost of maybe six hours. The reason that works is that the constraint got built before the throughput, which is the opposite of how most teams approach it and most of the reason most teams end up disappointed.

If you want a single place to start tomorrow, open your analytics, find your best-performing post from the last year, and write down the six angles you’d take from it if you had unlimited time. Compare that list to what a model gives you from the same post with no voice spec attached. The gap between those two lists is exactly the work this page is about.