All essays
The Manifest
November 18, 202610 min read

Build an AI Content Workflow Without Losing Your Voice

The draft is grammatically fine. It covers the topic, uses the right terminology, and reaches a reasonable conclusion.

You still would not have written it. The argument is softer than yours, the examples imply experience you never supplied, and every section ends with a tidy little lesson.

I would build an AI content workflow that protects the author's decisions before it optimizes production speed. That means capturing the actual point of view, checking the evidence, editing the argument and language separately, and approving the exact version that will be published.

TL;DR & Takeaways

  • Voice includes what you believe, what you challenge, and what you choose to leave out. A tone prompt cannot supply those decisions.
  • Give the workflow approved examples, explicit boundaries, and a source packet before asking it to draft.
  • Separate factual review from voice editing, and never allow an invented anecdote to become proof of expertise.
  • Tie approval to an identifiable final version. Topic approval and a positive comment on an earlier draft do not authorize publication.
  • Measure accepted work, review effort, and corrections rather than celebrating how many drafts the system generates.

Start the AI content workflow with a real thesis

Before research, ask the author what the reader needs to understand or decide. Capture the answer in plain language, including the nuance that would be lost in a generic treatment.

For an article about marketing automation, the thesis might be that repeatable work can be automated while consequential judgment needs an explicit review path. That is a specific argument, not merely the topic “AI in marketing.”

Ask what prompted the view and whether there is a real example the author is willing to use. If there is no approved example, use a clearly labeled hypothetical scenario or teach the mechanism directly.

Do not prompt the model to “add a personal story” and hope an editor catches the invention later. Put the story boundary in the brief: use only supplied experiences, and flag any missing detail.

Record which parts of the thesis are settled and which remain a working hypothesis. The draft should not turn “this may be worth testing” into “this always works” because the stronger sentence sounds cleaner.

A short author interview can supply more useful direction than pages of adjectives about tone. Preserve the original notes so the editor can check whether the draft still makes the intended argument.

Build a voice reference from actual decisions

Choose a small set of recent, approved pieces that reflect the author's current style. Older work can provide context, but it should not override deliberate newer preferences.

Annotate the examples. Explain how the author opens a problem, qualifies uncertainty, uses first person, and moves from explanation to action.

Include negative examples with a reason. “Avoid hype” is vague; “replace claims of transformation with the specific workflow and outcome” tells the writer what to do.

Separate voice rules from business facts. A preferred sentence rhythm should not live in the same unstructured paragraph as current pricing, offer boundaries, or a customer's permission to be named.

Keep one current reference for each purpose and preserve historical source material separately. If a business changes its offer, the workflow should not retrieve an old description and quietly restore it.

Do not treat a banned-word list as a complete voice system. An article can avoid every prohibited phrase and still make an argument the author does not believe.

Give research its own evidence standard

Ask the researcher to identify the reader's decision, inspect competing explanations, and collect primary support for material claims. Search results are a discovery route, not the evidence itself.

Open the underlying pages and save the relevant passage with the URL and retrieval date. Note whether a source is primary research, official documentation, a vendor claim, or an interpretation.

Use a claim register with the proposed wording, source, supporting passage, and limitation. If a source supports a narrower claim, change the sentence rather than asking the model to defend the broader version.

Check examples and calculations separately. Run numerical scenarios in a calculator or code, preserve the inputs, and label them hypothetical when they are invented for teaching.

For confidential material, use only approved tools and appropriate permissions. Redact unnecessary personal or customer information before sending it to an external model, and confirm the relevant data-handling terms rather than assuming all AI products behave alike.

Google's guidance on generative AI content calls for manual fact-checking and review for accuracy and trustworthiness before publication. It also applies that review to titles, descriptions, structured data, and image alternative text.

That guidance does not mean useful AI-assisted work is prohibited. It warns that generating many pages without adding value can violate scaled-content-abuse policies, which is a different issue from whether a tool helped organize the draft.

Use a brief that leaves fewer decisions implicit

This is a copyable input structure for a substantive article:

text Reader and decision: Approved thesis and important nuance: What the article must help the reader do: Author notes and approved experiences: Current business facts and offer boundaries: Voice examples and explicit non-examples: Source packet and claim register: Hypothetical example assumptions: Existing related articles to avoid duplicating: Required reader asset: Open questions that must remain flagged: Approver and permitted next action:

The final field matters. “Prepare a review draft” should not give the system permission to update the live site, send a newsletter, or post a social summary.

Use a constrained drafting prompt:

Draft from the approved thesis and evidence packet. Preserve the author's uncertainty and offer boundaries. Do not invent personal experience, quotations, customer results, or sources. Mark unsupported claims for review rather than filling the gap. Include the reader asset, and return a draft with a separate list of unresolved questions. Do not publish or distribute it.

This is a starting instruction, not a guarantee of compliance. The workflow still needs to inspect the output and reject unsupported material.

Edit the argument before polishing the sentences

The first editorial pass should ask whether the article delivers the promised decision. Check the logic, the important distinctions, and the example's limitations before adjusting cadence.

A fluent draft may omit the condition that makes the advice useful. For example, a recommendation to increase acquisition spending needs to account for economics and capacity, not merely a promising conversion rate.

Next, check every claim that a reader could verify or act on. Confirm product capabilities, dates, quoted language, numbers, and the exact scope of the evidence.

Then edit for voice. Replace generic openings with the actual problem, remove repeated conclusions, and change abstract advice into a mechanism or action.

Read first-person sentences carefully. “I would start here” expresses judgment; “I have seen this across hundreds of clients” asserts experience that needs real support.

Avoid adding quirks merely to make the piece seem human. A forced joke, manufactured frustration, or invented memory makes the article less faithful to its author.

Keep a short edit log of recurring problems and the approved corrections. Update the reference with durable lessons, such as preserving uncertainty, rather than accumulating a large pile of prompts that contradict one another.

Separate the workflow's approval states

A queue should make the next permitted action obvious. Use explicit states and require the evidence appropriate to each transition.

State Required evidence Permitted next action
Brief approved Author's angle and reader decision Research and drafting
Draft in review Complete article, sources, calculations, and checks Request edits or exact-version approval
Changes requested Specific feedback attached to the version Revise and repeat affected checks
Approved Identified final version and author approval Prepare the authorized release
Published, verification pending Release receipt Check the actual public page
Published, verified Body, metadata, links, and layout checked Record the completed release

For The Manifest, the editorial control separates topic approval, draft review, and final publication approval. That is a concrete process example, not evidence of a measured performance improvement.

Use a version identifier or a computed file hash to associate approval with the exact artifact. If substantive edits follow, obtain approval for the revised version rather than treating an older approval as permanent.

Give the publishing step only the permissions it needs. Where practical, keep drafting tools unable to publish directly, with a separate authorized release path.

After release, read the page itself. Confirm that the approved body arrived intact, links resolve, tables remain readable, and the title and description do not promise something the article never says.

Technical checks cannot prove that a page will rank, be indexed, or receive AI citations. Record those limits instead of turning a successful deployment into a broader success claim.

Measure the work that survives review

Draft count is easy to improve by generating more drafts. It says little about whether the author has less work or the reader has a better article.

Track accepted articles, author review time, factual corrections, substantial rewrites, and post-publication fixes. Keep time spent maintaining prompts, integrations, and source references in the operating cost.

Consider a hypothetical eight-article pilot. Suppose the existing process takes five hours per article, or 40 hours total.

The AI-assisted process takes two hours of preparation and drafting plus one hour of review per article, or 24 hours total. Add eight hours to set up and maintain the workflow during the pilot, and total effort becomes 32 hours.

The pilot saves eight hours, a 20% reduction from the 40-hour baseline. At an illustrative $100 per hour, that represents $800 of capacity value before software costs, not necessarily an $800 cash saving.

Now suppose review takes two hours per article instead of one. Article work rises to 32 hours, and the same eight setup hours bring total effort back to 40.

The apparent drafting speed advantage disappears at the workflow level. That is why review effort belongs in the calculation from the start.

The AI Workflow ROI Calculator can help explore time-and-cost assumptions. Treat its result as a scenario, and measure the actual workflow before claiming savings or increasing output commitments.

Pilot one content motion before expanding

Choose one repeatable format with a clear author, review standard, and source base. Run it through the full process, including the failure cases: missing evidence, a changed thesis, conflicting feedback, and edits after approval.

Inspect where the author still has to rescue the work. If the draft repeatedly misses the argument, improve the interview and brief before adding another stylistic instruction.

If factual review consumes the saved time, improve source capture and claim boundaries. If approved work stalls in release, fix ownership and permissions rather than generating more inventory.

At MakeReality, AI Growth Operator work can include selected content-system builds with documented handover. The useful outcome is a process the team can operate and audit, with the author retaining control of the perspective and final release.

Start with one article and keep the evidence packet beside it. Expand when the process produces work the author can stand behind without inventing the person whose name appears at the top.

More from the manifest.

Browse all essays

Ready to grow?

Book a call to discuss your business, industry, and unique opportunities to see how we can help you scale.

Not ready for a call? Reach out directly → hey@makereality.io

Book a discovery call
01Book a time on our calendar that works for you.
02Have a 30 minute call to discuss your business.
03We'll give you some immediate opportunities we see, live on the call that we can carry forward into a project together.