Claude drafts each article on this blog from two documents: oAT's brand voice guide and a file of verified company facts. Software then checks the draft for structure, banned phrasing, and broken links. A claim-by-claim verification pass follows. A named person at oAT reads the finished piece and decides whether it goes live.
We're writing this one down because you deserve to know how the thing you're reading was made, and because the method is the interesting part.
Who Actually Writes These Articles
The division of labor
A person picks the question. Claude produces the plan and the draft. Software enforces the mechanical rules. A person verifies, edits, and publishes. Nobody at oAT pretends the first draft came out of a human head, and nobody pretends a model made the final call about what oAT believes.
That split matters because the two jobs are different. Generating fluent prose about a topic is now cheap. Deciding that a claim is true, that a framework is one we'd actually stand behind, and that a sentence sounds like us: that part stays with the person whose name is on the account.
Why the byline says "of All Trades"
The byline reads "of All Trades" because inventing a human author would be a lie, and a small lie about authorship makes every other claim in the piece worth less. When Britt writes in her own voice, she publishes under her own name on the "how We Do" Substack. When the blog publishes a process piece like this one, the company owns it.
The Process, Start to Finish
how We Do: the seven stages
Each stage has one owner. A draft can't skip a stage, and a stage that fails sends the draft backward, never forward.
| Stage | Who or what performs it | What has to be true to move on |
|---|---|---|
| 1. Question selection | A person at oAT | The question is one a real founder asks out loud |
| 2. Plan | Claude, reviewed by a person | Sections, tables, and FAQ answer the title's question |
| 3. Draft | Claude, constrained by the voice and context files | Voice matches the calibration essays; no facts outside the context file |
| 4. Mechanical gate | Software | Zero violations, or a logged override |
| 5. Claim verification | Claude, then a human spot check | Every factual claim maps to a source that opens |
| 6. Human read | A named person | The piece is one oAT would defend in a client conversation |
| 7. Publish and log | A person, with the record kept | The approval and any override are documented in writing |
Where the Words Come From
The draft model doesn't get to improvise about oAT. It gets two source files, and it's constrained to what they contain.
- The brand voice file. Derived from Britt's published essays, distinct from the site's marketing copy, with a calibration set of two live pieces the writer reads before drafting. It sets sentence register, the use of contractions, the italics-for-emphasis habit, and the terminology we use and refuse.
- The company context file. Public facts about what oAT does, taken from weofalltrades.com [1]: embedded generalist operators who join a client's team, work in the client's tools, and cover core functions so the leader isn't carrying all of it personally.
- An internal block that never gets published. Margin targets, contract values, and capacity mechanics sit in the context file so a writer doesn't contradict them, marked as unpublishable. If a draft reaches for one of those numbers, that's a failure, and it gets cut.
- A silence rule. Where a source says nothing, the file says "not established." A drafting model that finds no fact writes no fact.
The practical effect: an article can't quietly acquire a client name, a price, or a statistic that nobody at oAT ever said. oAT publishes no rates, so no article contains one.
What the Automated Checks Actually Check
The mechanical gate runs before any human reads the draft, because human attention is the scarce resource and machines are better at counting.
| Check | What causes a failure |
|---|---|
| Phrasing blocklist | Em-dashes, filler openers, and a list of words that read as machine-written |
| Sentence shapes | Contrastive-parallel constructions, rhetorical-question transitions, triple-adjective strings |
| Repetition | Any three-to-five word phrase used more than three times in one piece |
| Structure | A stray H1, a section with no table or list or subsection, a missing FAQ, a word count outside range |
| Citations | Stacked markers, a marker with no matching reference, a reference nothing cites |
| Links | Any cited URL that fails to resolve, including pages that look fine and aren't |
A draft that trips any of these comes back with line numbers. It isn't a suggestion the writer can shrug off.
Why the Link Check Gets Its Own Explanation
The problem with a URL that looks alive
A fabricated citation is the single most damaging thing a language model can put in an article, because it's the hardest for a casual reader to catch. Checking that a URL exists sounds trivial until you meet the soft 404. Plenty of sites answer a request for a nonexistent page with a friendly "we couldn't find that" page and a success status code instead of the 404 that HTTP defines for missing resources [2]. Google treats these as soft 404s and documents them precisely because they mislead crawlers about whether content exists [3].
How we catch it
So the checker does more than read the status code. It fetches the cited URL, fetches a known-good page from the same domain, and compares them. When a deep link comes back looking suspiciously like the homepage, or like a generic error template, the citation is treated as unresolved and the claim it supports has to find a real source or come out of the article. A reference list here is short because every entry survived that test.
What Happens When a Draft Fails
- The gate reports every violation at once, with locations, so fixing one problem doesn't hide the next.
- The draft is revised and the gate runs again. This loop repeats. Most drafts go around more than once, and the second pass usually catches phrasing the first pass introduced.
- An override exists. A human can push a piece through with a violation standing, because occasionally a rule is wrong for a specific sentence.
- The override is logged, with the reason and the person. An escape hatch that leaves no trace stops being an exception and becomes the process. Written down, it stays rare.
How Facts Get Verified
Claim extraction, then matching
Verification is a separate pass from the mechanical gate, and it asks a different question. The gate asks whether a sentence is well formed. Verification asks whether it's true. Every factual assertion gets pulled out of the draft as a discrete claim: a number, a date, a name, a quote, a description of what oAT does. Each claim is then matched against a source, either the company context file or an external reference that a human can open in a browser.
What happens to a claim with no source
It's cut. No hedge softens it, no appeal to "industry consensus" replaces it, and no vaguer verb keeps it in the piece. Same treatment applies to a source that turns out not to exist at the URL given. A shorter, thinner article is a fair price for one you can check. After the model finishes this pass, a person spot-checks it, because a verifier and a drafter that share a failure mode will agree with each other confidently.
Checked, Verified, Approved: Three States of a Draft
These three states are tracked separately, and passing one says nothing about the others.
| State | What it guarantees | What it does not guarantee |
|---|---|---|
| Mechanically checked | Structure, phrasing, citation format, and links all clear the gate | That anything in the piece is true |
| Fact-checked | Every claim maps to a source that resolves | That oAT agrees with the framing or would say it this way |
| Human-approved | A named person read the whole piece and chose to publish it | That the piece will never need a correction |
A draft sitting in the first two states is unpublished by default. Approval is the only state that puts an article in front of you.
What a Person Looks for Before Publishing
- Truth about oAT. Does the piece describe how we actually work, in terms a client would recognize from a real engagement?
- Voice. Does it read like Britt's essays, or like a competent stranger imitating them? A draft with no contractions passes every automated check and is still off-voice.
- Usefulness. Can a founder make one specific decision by the end? A piece that only describes a problem gets sent back.
- Restraint. Is there a pitch where there shouldn't be? Our best-performing essay mentions no oAT service at all, and that's the norm we hold to.
The version you're reading is rarely identical to the draft that arrived. Cuts, reordering, and rewritten openings are the usual outcome.
What We Will Not Do, and Why We Write This Way
- No invented human bylines. Company authorship or a real person, nothing in between.
- No invented quotes, statistics, prices, or dates. oAT's numbers come from oAT's own records or they don't appear.
- No citation we haven't opened. Every reference below resolves to a live page.
- No competitor named to be knocked down.
Writing this way happens to suit how people find answers now. Google's guidance is consistent: content produced mainly to manipulate rankings is targeted, and appropriate use of AI or automation is allowed under its rules [4]. What the guidance rewards is content demonstrating real experience and expertise, made for people first [5], and what it penalizes is scaled production with no regard for quality [6].
FAQ
Is the whole article written by AI?
The first draft is generated by Claude, working from oAT's brand voice guide and a file of verified company facts. A person at oAT selects the question, verifies the claims, edits the text, and decides whether it publishes. Published pieces usually differ from the draft in structure and wording.
How do I know a person actually read this before it published?
Publication requires a recorded approval from a named person at oAT, and a draft with no approval stays unpublished. Any override of an automated check is logged with the reason and the person who authorized it. The approval step cannot be performed by software.
Does using AI to write articles hurt search or answer-engine rankings?
Google has stated that its systems target content produced primarily to manipulate rankings, and that appropriate use of AI or automation is allowed under its guidelines [4]. What matters is whether the content is helpful, original, and made for people rather than for search engines [5]. Scaled content produced without regard for quality is treated as spam under Google's policies [6].
What happens if an article turns out to be wrong?
The claim is corrected or removed, and the source that let it through gets reviewed so the same error doesn't recur. Corrections are made visibly in the article itself, never quietly deleted. No verification process reaches zero errors, which is why the human approval step exists.
Can I tell whether a specific article was fact-checked?
Every published article on this blog has passed claim verification and human approval, since a draft missing either state stays unpublished. The reference list at the foot of an article shows the sources its claims were checked against. Every listed URL was fetched and confirmed to resolve to real content.
Why doesn't the article have a named human author?
Attributing a machine-drafted article to an invented person would be dishonest, and dishonesty about authorship devalues every other claim in the piece. Company-authored articles carry the "of All Trades" byline. Essays written personally by oAT founder Britt Gage are published under her own name.
References
- of All Trades. https://weofalltrades.com
- IETF. RFC 9110, "HTTP Semantics." https://www.rfc-editor.org/rfc/rfc9110.html
- Google Search Central. "HTTP status codes, network and DNS errors." https://developers.google.com/search/docs/crawling-indexing/http-network-errors
- Google Search Central. "Google Search's guidance about AI-generated content." https://developers.google.com/search/blog/2023/02/google-search-and-ai-content
- Google Search Central. "Creating helpful, reliable, people-first content." https://developers.google.com/search/docs/fundamentals/creating-helpful-content
- Google Search Central. "Spam policies for Google web search." https://developers.google.com/search/docs/essentials/spam-policies