← LetMeShowYouSomething

The whole loop

An agent asks, a person answers on one offline page, and the agent checks the answer and acts on it. Every piece below is real: the page, the answers it exported, the checker's verdict and the agent's report.

  1. The agent

    asks, as one page

    A proposed change to a small booking app: when someone cancels more than 24 hours before, they get store credit instead of their money back. The agent wrote it as a review and rendered it as one offline page — the screens to click through, the flow, and the question.

  2. A person

    answers, on that page

    They press Start, judge each step of the booking and the cancellation, leave notes, ask the agent to explain one step differently, add something nobody asked about on Return, and download the answered page. Thirty seconds, no account, no install.

  3. The answer

    comes back as a file

    Readable without the page: every answer carries its step's title. What is still open is listed as gaps, and the question to the agent as a request. An excerpt of the exported file:

    {
  "review": "booking-flow-2026-09",
  "summary": {
    "total": 11,
    "answered": 9,
    "unset": 2,
    "added": 1
  },
  "gaps": [
    "join-waitlist",
    "ask",
    "send-question",
    "added-1"
  ],
  "requests": [
    {
      "itemId": "ask",
      "title": "Taps Ask a question",
      "kind": "explain"
    }
  ],
  "responses": [
    {
      "itemId": "join-waitlist",
      "title": "Taps Join the waiting list for 10:00",
      "verdict": "disagree",
      "note": "Thirty minutes to book is too short when the email arrives at night."
    },
    … 10 more, one per step
  ],
  "addedItems": [
    {
      "id": "added-1",
      "title": "Show the cancellation deadline on the booking",
      "body": "People should see \"free cancellation until Friday 10:00\" before they try, not find out when it is too late."
    }
  ]
}
  4. The agent

    checks it before it trusts it

    The checker recomputes the summary and the gaps and refuses a file that does not add up. The agent does not act on a file that fails.

    $ node bin/check.mjs pair examples/flow-booking.review.json examples/flow-booking.feedback.json --root .
  ! examples are unverified: "A studio a colleague worked with" has no source. Add one, or the page shows it as unverified; agents invent convincing examples

PASS — 42/42 checks, 1 warning(s)
  5. The agent

    acts on it — gaps first

    In the order the skill prescribes: what is still open, with a recommendation and a question; what the reviewer asked for; the reviewer's own additions in their words; then everything else.

    The reviewer answered 9 of the 11 steps and added one thing of their own. The review didn't record their name, so I'll call them the reviewer. Four things are open.

    Gaps first

    1. Taps Join the waiting list for 10:00 — the reviewer disagrees.

    "Thirty minutes to book is too short when the email arrives at night."

    My recommendation: keep the 30 minutes during the venue's opening hours, and when a slot frees up at night, start the clock at 8:00 the next morning. The next person still gets a fair chance, and nobody loses a slot while asleep. For the reviewer: does that fix it, or should the window be longer at all times (say 2 hours)? I'll send it back as a one-item follow-up review.

    2. Taps Ask a question — no answer yet, and the reviewer asked me to explain it. The explanation is below under Requests. The step stays open until they have read it and answered.

    3. Taps Send — no answer yet. It is the second half of asking a question, so it depends on (2). My recommendation: keep it as suggested and ask again together with (2), in the same follow-up.

    4. Something the reviewer added — see What the reviewer added below. It has no verdict yet because it is new; I have to answer it, not the other way round.

    Requests

    The reviewer asked me to explain "Taps Ask a question". Here is what that step is and why it is in the flow.

    Today customers who have a question before booking, such as "Is there parking nearby?", email the venue, and nothing in the app handles it. This step is marked suggested: nobody asked for it; I added it because the flow had no way to ask anything before booking. Tapping Ask a question opens a short form, Ask the venue. Send delivers it to the venue by email, and the person sees "Question sent: the venue replies by email, usually within a day". Nothing about the booking changes.

    The question for the reviewer: should this be in the product (agree), left out for now (disagree), or done differently, for example as the venue's phone number instead of a form (partly agree, with a note)?

    Choices

    This review had no choose-one section, so there was no pick to report. The question I asked in the brief — store credit instead of money back when someone cancels more than 24 hours before? — was answered in the cancellation step below: store credit is accepted, with one change.

    What the reviewer added

    Show the cancellation deadline on the booking
    "People should see "free cancellation until Friday 10:00" before they try, not find out when it is too late."

    I didn't ask about this, and it is right: today the 24-hour limit only appears when someone tries to cancel too late. I'll add the deadline to the Booked screen, under the date, in the next round.

    Everything else

    • Confirms Cancel booking — partly agree.
      "Store credit is fine, but say "store credit" on the button itself, not only in the text above it."

      So the answer to my question is yes, store credit. I'll rename the button from Cancel booking to Cancel for €20.00 store credit, and the dialog text can stay as it is. Agreed, built as shown: Taps Book 10:00 · Taps Cancel booking · Taps Keep it · Taps Pick another time · Taps Find a time · Taps Pick another day · Taps Show times.

    What happens next

    As the review said: the seven agreed steps are built as shown. The cancellation button gets the new label, and the Booked screen gets the deadline. The waiting list, Ask a question and Send go back to the reviewer as one short follow-up review, with the explanation above.

How this was made: the answers are demonstration answers, given by a scripted reviewer in a real browser and recorded (the circle in the video marks each tap); the file comes from a longer run of the same page and is what it exported, unedited; the checker output is its real output; the report was written by an agent following SKILL.md, from that file. All of it is in examples/.

Privacy and terms: no cookies, no tracking, nothing sent.