The agent that helped us write 200 blog posts

This is the 210th piece of content we've written about Avery NXR. The first ~30 were written entirely by hand. The last ~180 were drafted with help from an agent we built internally.

If you've been reading along, this disclosure might be obvious in retrospect. We've been transparent throughout that we use Avery NXR to build Avery NXR (post 180), including for content. But we want to talk specifically about what the content agent does, what it doesn't do, and what we've learned.

What the content agent does

Our blog drafting agent is called (internally) blog-drafter. It runs on-demand when I'm starting a new piece.

Workflow:

1. I give the agent a topic, angle, target audience, and rough outline
2. Agent reads our existing blog corpus (it has access to all our previous posts)
3. Agent identifies relevant prior posts to link to + drafting style to match
4. Agent generates a structured first draft of 800-1500 words
5. Agent flags places where it's unsure or where I should add specific expertise
6. Output goes to a draft file I open in my editor

The agent takes about 3-4 minutes to run. The output is usually 70-85% of the final post.

I then edit for ~45-90 minutes per post. The editing is mostly: → Rewriting sections to sound more like me → Adding specific anecdotes or examples → Refining the structure → Correcting any factual or positioning issues → Polishing the conclusion

What the agent does NOT do

To be clear:

→ The agent does not pick topics. I pick topics based on what I want to write about, what customers are asking, what's relevant to current launches.

→ The agent does not provide strategic direction. What we say about local-first vs cloud-first, what we say about competitors, what positions we take — these are deliberate choices I make.

→ The agent does not write what I'd disagree with. I review every draft. Anything off-tone, factually wrong, or strategically off gets rewritten.

→ The agent does not have my voice perfectly. Most posts need significant voice editing. The agent gets close, not all the way.

→ The agent does not publish anything. Drafts go to me, I edit, I publish.

→ The agent does not know information I haven't given it. It can't surface new customer stories I haven't told it about. It can't reference data I haven't shared.

The agent is a drafting tool. The judgment, voice, strategy, and final word are mine.

How the agent was built

The agent's been refined over ~20 iterations during these 180 posts. Roughly:

Initial version: Generic LLM prompt asking for a draft on a topic. Outputs were bland, off-voice, often factually wrong.

Iterations 2-5: Added voice guidance with examples. Outputs improved on voice but still generic in arguments.

Iterations 6-10: Added the "read existing corpus" capability so the agent could match our existing positioning. Quality jumped significantly.

Iterations 11-15: Added "flag uncertainties" capability so I knew where to invest editing time. Reduced wasted editing on parts that were already good.

Iterations 16-20: Tuning specific patterns — when to use lists vs prose, when to include numbers vs avoid them, how to structure CTAs.

The current version is good enough that drafting time has dropped from 3-5 hours per post to about 45-90 minutes of editing per post. Across 180 posts, that's saved 360-700 hours of writing time over the past year.

What the agent gets right

Patterns it does well:

→ Maintains structural consistency across our content (sections, callouts, CTAs) → References prior posts appropriately (it knows what we've said before) → Uses the right level of technical depth for our audience → Avoids LLM cliches (we explicitly trained it to avoid certain phrases) → Produces content that fits our standard length range → Includes appropriate CTAs (avery.software, Free Desktop tier mentions)

What the agent gets wrong

Patterns I still have to fix manually:

→ Tone drift in the middle sections (agent gets formal/generic when sections feel "neutral") → Specific examples (agent can't invent specific customer stories — I add those) → Strong positional takes (agent hedges; I make takes more pointed when warranted) → Personal anecdotes (agent can't add what didn't happen in our actual experience) → Specific competitor analysis (agent needs guidance to take positions vs. competitors) → Numbers (agent often uses round numbers; I use specific ones from our data)

These are recurring patterns I've learned to fix as part of editing. The agent doesn't improve at them automatically.

Why we've been transparent about this

A few reasons:

1. It's the right thing to do.

Readers deserve to know how content is produced. Some readers may discount AI-assisted content. That's their right. Hiding the AI use feels dishonest.

2. It demonstrates the product.

If we're going to claim AI agents help with operational work, we should be using them for our own operational work. The blog drafting is a real example. Talking about it openly is part of the product story.

3. It builds trust.

When companies hide AI use, it eventually comes out and damages trust. Being upfront prevents the damage. Some readers respect honesty.

4. It models good practice.

Lots of writers and content marketers are figuring out how to use AI ethically. Our public approach (use AI for drafting, edit substantially, disclose openly) is one model others can adapt.

What this means for the writing

Some readers might wonder: is AI-assisted content still "ours"?

Our view: yes, when the substantial work is human.

The structure of the content (which posts to write, what to argue, what positions to take) is human. The voice editing (making it sound like us, not generic) is human. The strategic direction (what we're communicating about the company) is human. The fact-checking (is this accurate) is human.

The drafting (the first 70-80% of the words) is AI-assisted.

In writing, the structure and voice and strategy and fact-checking are the parts that matter most. The first-pass drafting is the most automatable part. Letting AI absorb the automatable part doesn't change whether the content is "ours."

You could disagree with this framing. Some writers think any AI involvement means the content isn't authentically theirs. That's a reasonable position. We don't share it, but we respect it.

What's different about this disclosure vs. typical "AI-assisted" notes

Most companies that use AI for content either:

→ Don't disclose it → Add a vague "AI-assisted" line in fine print → Treat the AI use as embarrassing rather than ordinary

We're trying to do something different: write a full post about exactly how the AI is used and what it does/doesn't do.

The reason: the AI question is going to keep coming up in content marketing. Companies that figure out transparent, ethical, professional use of AI for content will model what good practice looks like.

We hope this post is useful as a reference for that conversation.

What we'd recommend other content teams do

If you're using AI for content and wondering how to handle it:

→ Define what the AI does and doesn't do. Be specific. "Drafts first version, human edits substantially" is clearer than "AI-assisted."

→ Be honest in your disclosures. Don't downplay or hide. Don't over-claim either.

→ Make sure substantial human work happens. If AI does 95%+ and humans rubber-stamp, that's different from AI doing 70-80% and humans editing meaningfully.

→ Talk to your team. Internal alignment on AI use matters. Disagreement is fine; surprise is not.

→ Stay consistent. Pick a policy. Don't change it case-by-case.

→ Be open to evolving your view. Best practices in AI-assisted content are still being established. What works in 2026 may need adjustment in 2028.

The bigger lesson

The agent that drafts our blogs is our most-used custom agent. The ROI on it (in time saved, in content volume produced) is dramatic.

It's also a microcosm of how we think AI agents should work in operational settings: AI does the high-volume scaffolding work. Humans do the judgment, voice, and strategic work. Together produces more than either alone.

If you've been reading our content over the past year, the agent has been part of how it got produced. We hope the content has been valuable regardless of how it was drafted. The transparency about how it was produced is, we think, part of treating readers with respect.

200 posts done. The pattern continues. The agent keeps running. We keep editing.

→ avery.software — Free Desktop tier. The agent that helped write this post. And the platform you can use to build your own.