You Still Have to Put the Effort In: My Standards for AI-Assisted Blogging

Readers Can Tell When Nobody Chose

You know the feeling. You click a link, read two paragraphs, and something goes flat behind your eyes. The grammar is clean. The structure is tidy. There is a list of benefits and a conclusion that repeats them. Nothing is wrong, but nothing in it needed to be written either.

I've been writing this blog with AI in the loop since January 2025. It has made publishing much more accessible to me, especially the parts that used to consume all my executive function before I reached the actual writing. It has also made it dangerously easy to produce more.

That is what I mean by rapid fire: treating the number of published posts as the goal. Pick a keyword, generate a plausible article, dress it for search, publish it, repeat. It is the old content-farm model with a much faster machine attached.

The problem isn't that AI touched the prose. The problem is that nobody had to bring an experience, make an argument, or decide whether the post deserved to exist.

Transform, Don't Manufacture

The distinction that keeps me honest is simple: I use AI to transform material, not manufacture it.

The best posts here begin with something I already have: terminal output, a failed configuration, notes from a talk I attended, an argument I cannot stop thinking about, or a system I built and then watched break. The model helps me turn that material into sentences and structure.

That matters because producing sentences is often the part I find hard. I can know what happened, why it matters, and what I think about it while still being unable to make the words come out. AI removes that bottleneck. It turns writing into a production and editing problem that I can work through.

A bare request such as "write a post about code review" has none of that material. The model can still return competent prose, but it has no access to the particular failure, disagreement, or hard-earned preference that would make the post mine. It fills the empty space with something plausible and familiar.

Bring the substance. Let the model help with the shape.

What Goes Wrong Without the Substance

I learned this while building a bot that reads a client's inbox. I asked it to list active leads from an empty data store, and it returned plausible names with plausible statuses. The bot was trying to satisfy the shape of the request without the facts needed to answer it.

I have seen the same failure in my writing drafts:

• Anecdotes that never happened. A draft once said a friend shrugged at me. Nobody shrugged. The opening wanted a personal beat, so the empty slot got filled.
• Attendance I didn't have. Conference notes became prose that implied I had attended sessions I missed.
• Paraphrases promoted to quotations. A rough note acquired quotation marks and the authority of a transcript.
• Precise-looking references from memory. Titles, dates, affiliations, and links looked right until I opened them.

Those are my failures, not evidence of a model misbehaving. I asked for finished prose before supplying enough truth, then failed to challenge the result.

This post caught itself doing it. An earlier draft attributed the phrase "rapid fire" to a named speaker even though I could not remember the exact quote. The attribution made the opening sound more authoritative, but I could not support it, so it came out. The same draft said rules in CLAUDE.md "run" like tests. They don't. They remind the agent what I care about; they cannot prove the result follows those rules.

That is the editing work: not making generated prose prettier, but noticing where confidence outran the material and correcting it.

The Workflow I Actually Use

1. Start with evidence. I bring the failure, commands, configuration, notes, or argument. If I have only a topic, I am not ready for a draft.
2. Use the model for shape. It can organize the material, propose a structure, and get sentences moving when I am stuck.
3. Challenge the substance. I ask what is unsupported, generic, overstated, repetitive, or missing. Then I correct it from what actually happened.
4. Edit across a batch. I review ten or fifteen posts in the same mode, making separate passes for argument, accuracy, tone, repeated habits, links, and images.
5. Make the useful checks durable. Instructions in CLAUDE.md keep expectations visible. Mechanical checks catch broken links and placeholders. Neither replaces reading the post and taking responsibility for it.

The batch editing is important. One post opening with "Look, I've been..." can sound like my voice. Six posts opening that way reveal a template. One tidy three-part list may be useful. A dozen of them reveal that nobody questioned the first structure offered.

The repository history shows the process better than a polished final page does. My commits touch groups of posts because I make critical passes across the whole set. Tone becomes visible in comparison. Repeated structures become annoying enough to notice. A claim that sounded convincing alone looks suspicious beside the source material.

AI Saves Me Work

My older description of this workflow emphasized speed. The tooling still holds up, but speed is no longer the part I find interesting.

AI genuinely reduces the work. Frontmatter, file naming, HTML structure, folder selection, and staring at an empty sentence used to consume energy I often did not have. Automating those parts means more of my ideas make it to the page.

What it cannot remove is the minimum judgment required to make a post worthwhile. I still need an experience worth examining, a position on what it means, verification that the details are true, and enough taste to reject prose that is merely competent.

The total effort is lower. The remaining effort is concentrated where I am useful.

The Slop Pass

One editing pass is devoted to patterns that make the writing feel assembled instead of considered:

• the same contrast construction in every section;
• everything arranged into groups of three;
• paragraphs marching down the page at the same length and rhythm;
• abstract words such as crucial, pivotal, landscape, and testament doing work that a concrete detail should do;
• bold lead-ins attached to every bullet;
• a conclusion that summarizes the summary.

Wikipedia's Signs of AI writing is the best catalogue I have found. I don't use it to prove whether a model wrote something. I use it to catch habits that make prose predictable and easy to abandon.

Readers cannot always identify which tool produced a sentence. They can tell when nobody made meaningful choices about it.

The Writers Who Set My Bar

A model can help reproduce patterns, but it cannot give me standards I never developed. The prerequisite for editing is reading people whose work makes me want to write better.

• Sean Goedecke shows me what a strong engineering argument looks like when it comes from shipping real systems.
• Simon Willison sets the bar for documenting what tools can actually do while the knowledge is still useful.
• Xe Iaso is a reminder that infrastructure writing can have an unmistakable voice.
• Cory Doctorow connects technical systems to ownership, power, and the right to modify the things we depend on.

I am not listing them to claim their work as validation for mine, or to argue that AI-assisted writing supersedes the craft that taught me what good blogging looks like. I use this workflow because it enables me to work well. Their writing helps me recognize the distance between a clean draft and a post with a point of view.

The Test

One question settles most cases:

Would this be worth reading if a human had typed every character?

If the answer is no, disclosure does not rescue it. If the answer is yes, the next question is whether I can stand behind every experience, claim, and choice on the page.

The workflow is available to adopt: bring real material, use AI to give it shape, challenge the result, edit it beside your other work, and make repeated lessons durable. The goal is not to publish as often as the machine permits. It is to publish things that would have been worth saying without it.

Header photo by Amsterdam City Archives on Unsplash.