Booking → Intake → AI service brief automation (Zapier + OpenAI) with T-60 delivery, fallback, and dead-letter logging.
TL;DR
This automation provides the service provider (mock interviewer) with instructions on how to conduct the service (a mock interview), based on the clients needs.
When a client books, they receive a pre-filled intake form (hidden fields carry booking metadata). On submission, an LLM compiles a clean, custom-tailored brief that’s delivered to the service provider ~1 hour before the session (or immediately if the intake arrives late). There’s a failure path that alerts admin, sends the service provider the raw intake as a fallback, and logs to a dead-letter sheet.
Demo (2min Loom): (https://www.loom.com/share/0c481ce4503143c2bb39f7766ea8330a)
Prompt & Samples
System prompt (template) System prompt (filled sample) Service brief (sample AI output)
Screenshots
Stage 1 screenshot Stage 2 screenshot
Intermediaries were spending ~1 hour per session prepping service direction from scratch (chasing details, writing notes). We needed a repeatable, low-touch way to collect context from the client and deliver a standardized brief to the service provider, without manual ops.
Stage 1 — Intake request
Trigger on new booking (SimplyBook). Filter for the relevant service-type and email the client a pre-filled intake link (hidden fields carry booking metadata).
Stage 2 — Direction generation & delivery
Trigger on intake submission (Google Sheets). Generate the actor brief (OpenAI), write it back to the sheet, and either send now (if < 1 hr to start) or delay so it lands T-60m. If the AI step fails, email admin, email the actor the raw intake as a fallback, and append to a Dead_Letter sheet.
View full-size flow diagram
- Automation: Zapier (Filter, Paths, Formatter, Delay, Email)
- Data / triggers: SimplyBook → Google Sheets
- AI: OpenAI (chat completions, system prompt enforces mappings/format)
- Notifications: Email (actor/admin).
The OpenAI step uses a strict, sectioned system prompt that enforces:
- Seven output blocks (Session/Coaching, Interviewer Direction, Company/Role, Resume, Sliders + Actions, 20-question set, Notes)
- Canonical labels, spacing rules, slider math (0–100 → 1–5), and validations (e.g., include “Interrupt 5×+” when metrics==Practice; client specified they want to practice being interrupted in intake)
- Guardrails for missing inputs and off-limits topics
See the full text: System prompt (template with placeholder variables) System prompt (filled sample)
Inputs
From booking (Stage 1): client/provider emails, service name, session start + TZ, booking code → embedded as hidden/pre-filled intake fields. From intake (Stage 2): goals, session_format, coaching_opt_in, five sliders (0–100), behavior toggles (Practice!/Neutral/Avoid), off_limits, resume_text, job_description, plus booking metadata.
Outputs
- Client email with the pre-filled intake link (Stage 1)
- Actor email with the LLM-generated direction, delivered T-60m or immediately if < 60m (Stage 2)
- Sheet updates storing the brief on the intake row
- Dead-letter row on failure (timestamp, booking_code, error, notes)
Service brief (sample AI output)
- Access & secrets. Platform-stored connections; no keys in plain text. Service account with least-privilege scopes.
- Timing. Formatter computes T-60; branch to send-now vs delay so delivery lands at the right moment.
- Failure handling. If the AI output is empty, Path B:
- Email admin (flags booking and context)
- Email actor the raw intake as a fallback so the session isn’t blocked
- Append a row to Dead_Letter
- Logging. Intake sheet stores the generated brief; Dead_Letter captures failures.
- Deliverability (demo note). Email by Zapier is used for this demo; production will switch to a domain-authenticated sender to avoid delivery to SPAM folders (GSuite/Outlook or SMTP/SendGrid with DKIM/SPF).
- Privacy (planned). Prefilled intake URLs will be tokenized so PII isn’t exposed in query params.
- Idempotency (planned). Gate by booking_code (sheet flag or Storage by Zapier) to skip duplicate sends.
- Client books → instantly receives intake link; completes it in their own time.
- Stage 2 generates a structured brief and stores it in the google sheets row.
- If > 60 minutes to start, email is scheduled for T-60m; otherwise it sends immediately.
- A forced failure results in: admin being notified, actor receiving the raw intake, and a Dead_Letter entry being created.
- Swap sender to domain-authenticated email; add Slack alerts on failure.
- Tokenize intake URLs; add a short-lived one-time code.
- Add a tiny JSON “sidecar” to validate sliders/actions before send.
- Track metrics: on-time delivery %, error rate < 1%, hours saved/session.
© Jorgan Thibodeau, MIT License.