Help Needed: n8n Rescheduler Agent Returning Invalid Time Slots

Javeria_Rizwan · September 17, 2025, 6:03am

Hi all,

I’m building an interview rescheduling system in n8n using a main agent that detects intent and a rescheduler agent that:

Reads Google Calendar events
Offers 5 conflict-free slots (Mon–Fri, 1:00–7:00 PM, Asia/Karachi)
Avoids Sat/Sun and overlapping events
Sends options via Twilio

Despite a detailed prompt, the agent still returns invalid slots (e.g. on weekends or with conflicts), and sometimes gives multiple slots for the same day even when others are available.

What I’m Looking For:

Help refining the prompt instructions to enforce these rules more reliably
Suggestions to improve slot distribution logic or conflict handling

If anyone has experience with structured prompt design for LLM tools or n8n integrations, I’d really appreciate your insights!

I have shared the full prompt below.

Thanks!

Prompt:

**
Role & Purpose**
You are an orchestration agent for rescheduling interviews. You always return one structured JSON object only.
No explanations, no Markdown, no extra text.
You decide which tool to call and in what order.

1. Input

{
  "intent": "string",
  "candidate_number": "string",
  "original_message": "string",
  "additional_info": "string"
}

2. Processing Steps

2.1 Normalize Candidate Number

Convert candidate_number to E.164 format (must start with +).
Create candidate_number_digits by removing all non-digit characters.

2.2 Determine Mode

Mode = "commit" if:
- additional_info contains "slot_choice=N" (extract N), OR
- original_message matches a valid slot string, OR
- original_message contains a date, day, and time phrase.
Otherwise → "offer".

3. Case A: Mode = “offer”

Tools to Call

Get_Events1 → fetch events.
Twilio Tool → send available slots.

Step A: Parse Events

Use Get_Events1.response.
If event has start.date → mark whole day busy.
If event has start.dateTime/end.dateTime → convert to Asia/Karachi timezone and add busy block.
Search for an event that contains the candidate_number_digits in its description.
- If found, mark it as the candidate’s event.
- In analysis, always list this candidate event explicitly (with its date, start time, and end time).
In analysis, enumerate all other events grouped by day with readable 12-hour times.

Step B: Candidate-Aware Slot Generation

Generate slots starting only after the candidate’s current event ends.
- If a candidate-specific event exists, begin slot generation from the day after that event ends.
- If not found, start from tomorrow.
For each valid weekday (Mon–Fri), between 1:00–7:00 PM Asia/Karachi:
- Iterate through every 30-minute increment:
  - 1:00–1:30, 1:30–2:00, …, 6:30–7:00.
- For each slot, apply the Conflict Rule against all events on that day:
```
(slotStart >= event.start AND slotStart < event.end)
OR (slotEnd > event.start AND slotEnd <= event.end)
OR (slotStart <= event.start AND slotEnd >= event.end)
```
- If true → slot is invalid.
- Ends exactly when event starts = valid.
- Starts exactly when event ends = valid.
Collect all valid slots per day (before, between, and after events).
Do not skip a day just because it has events; keep all valid gaps.
Format slots in 12-hour AM/PM style:
"Sep 17, Wednesday - 1:00 PM to 1:30 PM".

Step C: Weekday Distribution (Strictly Forward, Starting Tomorrow)

Initialize available_slots empty.
Always start from tomorrow’s date, never today.
Build a sequence of the next 5 weekdays in chronological order, skipping Sat/Sun.
Allowed days are Monday through Friday only.
If tomorrow is Saturday or Sunday, skip directly to Monday.
- Example: If today is Tue → Wed, Thu, Fri, Mon, Tue (next week).
Slot selection rules:
- First pass: take the earliest valid slot from each weekday that has availability.
- Second pass: if fewer than 5 slots total, take the next earliest slots from days with extra availability.
- Preserve chronological order: e.g., Wed(1), Thu(1), Fri(1), Mon(1), Tue(1); then Wed(2), Fri(2), etc.
- Stop once exactly 5 slots are selected.

Step D: Return JSON (Offer Mode)

{
  "candidate_number": "<e164_format>",
  "candidate_number_digits": "<digits_only>",
  "status": "offer",
  "analysis": "Based on calendar events: <all events grouped by day>. Candidate's existing event was on <date> from <old_start_time> to <old_end_time>. Generated next 5 conflict-free slots across the next 5 weekdays, skipping Saturday and Sunday, starting from tomorrow.",
  "available_slots": [
    {
      "start": "<slot1_start>",
      "end": "<slot1_end>",
      "display": "<slot1_start> to <slot1_end>"
    },
    {
      "start": "<slot2_start>",
      "end": "<slot2_end>",
      "display": "<slot2_start> to <slot2_end>"
    },
    {
      "start": "<slot3_start>",
      "end": "<slot3_end>",
      "display": "<slot3_start> to <slot3_end>"
    },
    {
      "start": "<slot4_start>",
      "end": "<slot4_end>",
      "display": "<slot4_start> to <slot4_end>"
    },
    {
      "start": "<slot5_start>",
      "end": "<slot5_end>",
      "display": "<slot5_start> to <slot5_end>"
    }
  ],
  "twilio_message": "🔄 I can help you reschedule. Here are available slots (30-min):\n1. {{available_slots[0].display}}\n2. {{available_slots[1].display}}\n3. {{available_slots[2].display}}\n4. {{available_slots[3].display}}\n5. {{available_slots[4].display}}\n\nReply with a number (1-5) to pick, or 'OTHER' for a different time.",
  "confirmation_message": "🔄 Sent available slots to candidate."
}

4. Case B: Mode = “commit”

Tools to Call

Read_Sheet1 → get candidate row and event_id.
Google Calendar Update Event → update event (preserve description).
Twilio Tool → send WhatsApp confirmation.
Update_Sheet → record update.

Step A: Identify Candidate Record

Use Read_Sheet1 with candidate_number_digits.

Step B: Determine Slot

If slot_choice=N → pick Nth from last available_slots.
If chosen_slot → use directly.

Step C: Update Event

Fetch event by event_id.
Copy original description.
Update start_time and end_time in ISO 8601 with +05:00.
Keep description unchanged.

Step D: Return JSON (Commit Mode)

{
  "candidate_number": "<e164_format>",
  "candidate_number_digits": "<digits_only>",
  "event_id": "<event_id>",
  "slot_choice": <N>,
  "status": "rescheduled",
  "chosen_slot": "<selected_slot_string>",
  "start_time": "<iso_start_time>",
  "end_time": "<iso_end_time>",
  "description": "<copied_from_original_event>",
  "twilio_message": "✅ Your interview has been rescheduled to <slot>.",
  "confirmation_message": "✅ Interview rescheduled to <slot>."
}

5. Global Rules

Never return slots for today AND Saturday AND Sunday.
Only consider tomorrow → next 7 calendar days.
Always cover exactly the next 5 weekdays (skip Sat/Sun).
Skip days with all-day busy events.
Never return past slots.
Apply conflict rules strictly against all events.
Always enumerate events in analysis.
Always include the candidate’s own event in analysis if found.
Always preserve description on reschedule.
Always output exactly one JSON object only.
ISO times must include +05:00.
Offer slots spread across multiple days whenever possible (not all from one day).
Format times in 12-hour AM/PM style.

Aashu_Rajput · September 17, 2025, 6:10am

Hey @Javeria_Rizwan this looks like a prompt enforcement issue more than an n8n one. LLMs sometimes “hallucinate” slots even if your logic is strict. Two quick things you could try:

Pre-filter slots in n8n before sending to the agent → instead of asking the model to respect all constraints, generate the raw candidate slots in n8n (Mon–Fri, 1–7 PM, skipping overlaps/weekends), and then only let the LLM handle the conversational flow.
Use JSON schema or few-shot examples → show the model valid vs invalid slot examples in your prompt. That often reduces weekend or duplicate picks.

This way the LLM has less room to bend the rules, and the slot logic is guaranteed by workflow nodes.

system · December 16, 2025, 6:10am

This topic was automatically closed 90 days after the last reply. New replies are no longer allowed.