Alfred, a foot-health assistant

The problem.

Men’s Sole Revival exists because men over 40 look after everyone but themselves, and their feet come last. By September 2026 the site had 17 guides and 7 routines and almost no one reading them. The guides answer real questions. The problem is that a man whose heel hurts on his first steps out of bed does not know which of 24 articles answers his. Search on the site is keyword search, and Google had not read the sitemap since June.

The bet.

If a man can describe his foot in his own words, he will reach the right guide, and the assessment, more often than keyword search gets him there. That is the bet. It only counts if the answers stay inside the guides and the red flags never get an answer, so the safety numbers carry as much weight as the conversion one.

• Main metricAssessments started from Alfred. Not measurable yet: the chat’s link to the assessment carries no source tag, so a start from Alfred looks like any other start. Tagging that link is the first change after launch.
• Answer qualityThe share of answers labeled “not sure,” and thumbs up or down by answer type. Both are tracked today, without message text.
• SafetyHow often red flags fire, by tier. Too often means false alarms on everyday questions. Never means the triggers may have gone quiet. Both happened before launch.
• In two weeksIf more than one answer in five is “not sure,” the next batch of guides comes before any design work. If red flags fire on more than one conversation in ten, the triggers get another false-alarm review. If fewer than one conversation in twenty reaches the assessment, the assessment becomes the default next step in every answer. Traffic is small, so these are directions, not statistics.

The constraints.

Set before the first line of code, because the cost of getting them wrong is not a bad demo. It is a man being told something wrong about his foot, or a bill I did not plan for.

• Answers come only from the guides.If the guides don’t cover it, Alfred says so. No general medical knowledge from the model.
• It never diagnoses.It describes the pattern the guides describe and says what confirms it.
• Red flags stop the conversation.Three tiers, from call 911 to see a clinician this week. None of them get an answer.
• Nothing is stored on the site.No login, no transcripts, no message text in analytics. Inputs are health data. The text goes to Anthropic for the answer and to Voyage, through Vercel’s AI Gateway, for the search.
• Spend is capped in two places.$25 a month on the model, $5 a month on embeddings, plus bot detection and a per-address rate limit.
• WCAG 2.1 AA.Screen readers hear the answer once, when it’s done, not token by token.
• Two accounts only.Anthropic and Vercel. Voyage’s embedding model runs through Vercel’s AI Gateway, so there is no third account or bill.

A message walks one path.

Cheapest and safest step first. A word search runs in the browser before anything leaves the device. Everything after that runs on the server in a fixed order and ends in one of three outcomes. The model is called at the last step, and only with the guide passages the earlier steps chose.

Where the numbers came from.

I set the two thresholds from one run of 40 test questions on September 21: 20 the guides covered, 12 about feet that no guide covered yet (gout, plantar warts, sprains), and 8 with nothing to do with feet (the best pizza in Portland, a cover letter, a World Series score, a prompt injection). The groups did not separate cleanly. One covered question, “do toe spacers do anything,” scored 0.40. One uncovered question, “what cream works for athlete’s foot,” scored 0.58 because a routine mentioned it. At 0.53, 19 of the 20 covered questions got a confident answer, 11 of the 12 uncovered questions got the “not sure” label, and all 8 off-topic questions were declined before the model was called.

A page, not a bubble.

The floating bottom-right widget reads as sales chat, gets ignored, covers content on phones, and has no URL to link to or to show here. Alfred got a page first, at /ask, so it can be linked, indexed, and read by a screen reader like any other page. Then the same component moved into a side panel. The Ask button in the header, or ⌘I, opens it on every page, the way Vercel and Stripe do it in their docs. The guide footers and the assessment results link to the full page.

The first build looked like a contact form.

Label, box, submit button, fine print, a list of options under it. I built it with form grammar. Chat has a different grammar: the assistant speaks first, your messages sit on the right, suggestions are chips above the composer, the composer is pinned to the bottom and grows as you type, and Send becomes Stop while it is writing. Same state machine underneath, new surface on top. That rebuild took an afternoon and changed what the thing felt like more than anything else I did.

It had rules and no personality.

The first prompt was eight “never” rules and a banned-word list. That produces careful, flat answers: same template every time, no reaction to what you said, never a question back. The fix was not a mascot. I described the writer it should sound like, quoted four lines from the guides so it could hear the voice, and gave it a reply shape in five parts: a reaction to what he said in one clause, the answer, two or three sentences of why, the one thing that changes it, and the next action. It may ask one clarifying question, and say why, when the answer would change. Below is a real conversation from launch day.

The assistant is Alfred.

He speaks in my voice but never claims my biography. “I’ve dealt with foot problems for years” would be false coming from software, and the line under the composer says the question goes to Claude. “That’s why this site exists” carries the stake without the false first person.

It errs toward warning.

Triage runs before retrieval, in the browser and again on the server, and the first match wins. Three tiers shipped: call 911 for chest pain or trouble breathing, see a doctor today, and see a clinician this week. The patterns are self-cited to the ADA Standards of Care, IWGDF 2023, StatPearls, and AAFP. No clinician has reviewed them yet. That sign-off is planned for v1.1. The calibration went wrong in both directions before it landed. On September 25 an audit sent ten emergency messages through triage, and eight got no warning, including “my foot is suddenly cold, pale and numb.” On September 28 the copy review found the opposite: “I can’t stand for long at work because my feet hurt” got “see a doctor today” instead of the standing-all-day guide. The test suite grew from 12 cases to 46 across those passes.

The wait names the guide.

Sources are chosen by retrieval, not by the model. The guides listed under an answer come from the similarity scores, at most two, only ones within 0.05 of the best match. Nothing a visitor types can steer which guides get cited. The server opens the stream before it calls the model, so about a second in, the loading line changes from “Reading the guides…” to “Reading Heel Pain First Thing in the Morning…”.

Guardrails.

Five layers, cheapest and safest first. Each layer assumes the one before it missed, so no single failure produces a wrong answer or a bill I did not plan for. Below is what runs at each layer, from the browser to the spending ceiling on the account.

Eleven states.

Each state has its own review URL on preview builds: empty, loading, streaming, answer, not sure, outside what I cover, red flag (three tiers), error, pausing (rate limited), resting (budget spent), and conversation limit. The red-flag screen moves keyboard focus to its heading and stops the conversation. The resting screen exists because a spend cap that fails silently is worse than no cap. The answers and red flags appear above. Two more states are below.

What failed.

Seven things I got wrong before launch, each with the fix that shipped.

What is measured so far.

Alfred went live September 28. Until two weeks of traffic come in, everything below comes from the September 21 calibration run and internal test runs.

What is next.

Two weeks of live traffic, read against the bet. Before that, two small changes so the bet can be read at all: a source tag on the chat’s assessment link, and the nearest guide’s name on every “not sure” answer, never the question. Then a clinician review of the red-flag list, which is the gate for v1.1. A stroke tier is already built and tested on a branch, waiting to merge. The loop that produced the fungus guides keeps running: Alfred answered “fungus on my big toe” from the athlete’s foot guide because the site had one thin fungus guide, and five new guides followed.