WAYS OF WORKING · DEMO 09 NINO CHAVEZ

The chiropractor's four questions.

A sports chiropractor I know emailed me four plain questions about selling rehab programs from his website. I handed them to an agent with one instruction: research it properly. What came back was confident, tabulated, sourced — and its evidence was marketing all the way down. This is the series' first client-facing session: the deliverable's reader can't check the work, which makes the evidence discipline the whole product.

Q1

How does the shopping part integrate into WordPress — via an app?

Q2

What processor do they use for the money part, e.g. PayPal?

Q3

Do they charge a monthly fee, transaction fee, etc?

Q4

Any links to shops using this, so I can see what it looks like?

or scroll to advance · to go back

02 THE FIRST PASS

The research looked finished. The sources were vendors ranking themselves.

Ten web searches, a platform comparison table, fee math, a clear recommendation, real example shops. Genuinely useful shape — the agent even reframed the question the client didn't know to ask (one-time program vs. recurring membership, which determines everything downstream). Then I asked one question:

"are the platforms you considered well reviewed? did you check reviews? prioritize them by number of reviews and rating? anything else to consider to justify including a platform to the list?" operator message, verbatim — the question that broke the first pass

The agent's own answer: no. The comparison had been assembled from vendor blogs — the recommended platform's post ranking itself against its competitors, a video host's post ranking itself among video hosts, a Gumroad rival's "best Gumroad alternatives" page. Marketing copy dressed as comparison, and it read exactly like research. Search summaries flatten who is speaking; the byline is the first thing they drop.

03 THE SECOND PASS

The "correction" swapped vendor marketing for competitor marketing

Round two pulled review-platform ratings and complaint patterns — better instincts, same disease. Three claims shipped into the client's brief wearing confidence they hadn't earned:

Shipped in the brief

Primary pick tagged verified directly — 4.8/5, 750 reviews (G2)

✗ re-derivation: G2 blocks direct page reads (HTTP 403). The number came from a search snippet referencing G2. Nobody had read the page — nobody could. The green tag was self-attestation.

Shipped in the brief

"The closest real-world example runs on the platform I'm recommending."

✗ live fetch of the shop: checkout runs on ClickBank. The "matching" URL was an affiliate marketer's funnel. Bonus finds: the "$67" price is a permanent fake discount off $138, and the sales page guarantees outcomes no licensed clinician could put their name on.

Shipped in the brief

"Roughly 1 in 3 reviews mention delayed or frozen payouts" — the stat that killed a platform.

✗ domain check: the figure traces to a competing payment processor's blog. A rival's stat about its rival is the same failure as the vendor's stat about itself — round two had just switched whose marketing it believed.

04 THE AUDIT

A second model, told to re-derive — not re-read

Mid-session I switched models and gave the new one the whole thread: "review the analysis and conclusion from sonnet's previous turns and review the delivered artifact." The instruction that matters isn't the model swap — it's that the auditor treated every "verified" in the draft as a hypothesis and went to the sources. Three probes, three different outcomes:

Refuted → rewritten

The example-shop claim died on a live fetch. The card was rewritten: copy the shape (one program, one price, instant delivery), explicitly not the marketing.

Unverifiable → downgraded

The G2 rating couldn't be verified by anyone. So the label changed to match the evidence: verifiedsearch-corroborated — plus an honest note in the brief telling the client the source blocks direct reads and the recommendation doesn't hinge on the number.

Checked deeper → strengthened

The first pass had read a FAQ about the payment processor's restricted-business list. The audit read the list itself — and found telehealth sits in the restricted tier, which upgraded a soft positioning suggestion into a hard boundary for the client.

The conclusions survived; the evidence labels didn't. That's the honest outcome most audits land on — and exactly the case where an unaudited draft is most dangerous, because nothing about the recommendation looked wrong.

05 DESIGN PROVENANCE

Claims need sources. So does a color palette.

The brief's visual design failed the same way its research did — twice. First pass: indigo accent, monospace labels everywhere. Second pass: warm cream, terracotta accent, bookish serif. Both are recognizable AI-default design clusters — free-picked "taste" drifting into the same few trained-in looks. My correction was blunt: "screams vibe coded in design choice for colors and typography."

The fix was the research fix, applied to design: ground it in something verifiable. A headless browser read the computed styles off the client's live site — buttons, headings, a color tally across the page — and came back with his actual brand: cardinal red on white, near-black section grounds, geometric sans. The brief was rebuilt in his identity instead of a template's.

Why it matters for client work

A non-engineer judges the deliverable by how it looks before reading a word of it. A brief styled in the client's own brand reads as made-for-them; a free-picked palette reads as exported. Same craft, different trust.

This page is the demo

The accent you're looking at right now is that harvested cardinal. One side effect worth keeping: once the client's red became the accent, "rejected" markers couldn't be red anymore — brand and verdict can't share a color.

Harvested beats tasteful. "This is your brand" outargues "this looked nice."

06 THE DELIVERABLE

Research a non-engineer can act on — labels included

The final brief is a single shareable page: his four questions answered first, then the parts he didn't ask about. What makes it applied AI rather than AI output:

Every number wears its provenance

Green: read at the source. Amber: search-derived, confirm before you commit. The legend explains itself, and the one number that couldn't be verified says so in plain text next to the recommendation it supports.

Rejections carry reasons

Six options ruled out, each with the cause — fee math, complaint patterns, wrong tool shape, or wrong sequencing ($495 up front before a first-time seller has validated anything). A recommendation is only as trustworthy as its visible rejections.

The gaps he didn't ask about

Liability positioning (education, not remote care — with the processor's restricted list as the hard line), where paid videos actually live (the plan's only real recurring cost — the first draft implied $0), and what to re-check himself before spending money.

The reasoning stayed attached: every section has a click-open note explaining why that call was made — including the notes where the honest answer is "this number couldn't be verified firsthand." The client gets the audit trail, not just the verdict.

07 PROMOTION

The failures became equipment before the session ended

Quality converged here because I drove three audit rounds by hand — reviews? design? verify? That's the honest limit, and it doesn't scale. So the same session promoted each failure into something standing (the pattern demo 03 is about):

A provenance rule, in the canonical prefs

Written into the working-style document every agent session loads: search snippets are hearsay; "verified" requires an in-session fetch of the source; vendor and competitor comparisons are marketing; if a source blocks reading, say so and downgrade.

A mechanical claim auditor

A deterministic pipeline stage in my content tooling: extracts prices, ratings, statistics, and verified-language from a draft, checks each against a provenance ledger, and hard-fails any "verified" without a fetched-at-source entry. No AI in the loop — regex and a ledger.

A brand-harvest script

The manual computed-styles pull, packaged: client URL in, brand tokens out (accent, grounds, fonts), ready for the design system. Ran it against the client's site to verify — it reproduced the manual harvest exactly.

An invocable audit pass

The full discipline as a named skill: extract load-bearing claims, trace provenance from the transcript rather than memory, re-derive at sources, check the deliverable's confidence taxonomy against itself, hunt the unpriced-cost gaps. Run before anything client-facing ships.

Committed and pushed the same afternoon. The next first pass carries what this session's third pass had to earn.

08 YOUR VERSION OF THIS

Ask what the source was — not whether it was checked

If you never touch code

When AI research hands you a comparison, ask who wrote the pages it read. A vendor ranking itself, a rival ranking its competitor — both arrive looking like neutral analysis. Then ask for labels: which numbers did it read at the source, which came from search summaries?

If you work with agents

Separate the drafting pass from the audit pass, and make the audit adversarial: every "verified" is a hypothesis until re-derived. A fresh context — different model, different session, or just different instructions — refuses the draft's framing in a way the author can't.

If you build deliverables

Make provenance visible to the reader: label every load-bearing number with how it was obtained, and let "this couldn't be verified — here's why the recommendation survives anyway" be a sentence you're allowed to ship. A confidence system that admits its limits is the only kind worth having.

who wrote the page the AI read? verified → fetched · amber → search-derived · say which the conclusions can hold while the labels lie
Colophon. One real session, July 2026: a friend's business email in, a shareable decision brief out, with the audit arc in between. Operator messages are quoted verbatim from the transcript; the client is anonymized (profession kept, name and practice withheld pending his OK — if he consents, this page gains a link to both his site and the brief itself). Claims about the session were re-derived from the transcript on drafting day. The self-attestation failure mode is demo 07's subject; the correction-to-environment pattern is demo 03; the browser-as-shell-command tool that harvested the brand is demo 02.