#@Last Pootis Hey. please share some

1 messages · Page 1 of 1 (latest)

green elk
#

I tried with Claude 4.5 Haiku , Sonnet 3.7 both fail

Condition:
After ch_confirm_selection succeeds with selection_type "INSURANCE", immediately execute this condition and proceed with location selection. Do not pause or wait for user input.

@inner heart @trim jacinth

trim jacinth
#

Whats ch_confirm_selection?

#

Can you try with gpt-4.1 please

green elk
#

thats a tool i have, i want to switch immediately once the tool has been called!

#

Claude has this more conversational flow and hence was using it!

trim jacinth
#

I see. Try with gpt-4.1 if it doesnt work there is something wrong in the prompt setup. I have had trouble with haiku and tool calls before as well

green elk
#

Let me know if i am doing something fundamentally wrong!

trim jacinth
#

I will take a look

#

no worries

#

funnily we are building insurance agents as well

green elk
#

aye nice!

inner heart
#

maybe try wording it differently? executing a condition doesn't really make sense
something like: ch_confirm_selection succeeds with selection_type = "INSURANCE"

#

you execute a tool but a condition becomes true at some point if that makes sense

#

other option is just to wait for gemini 3 flash that will most likely figure it out loll

#

the PR is there, we're just waiting on it to be merged

green elk
#

ahh ah i see, i have my demo tomorrow :/

#

let me try your suggestions

inner heart
#

ohh nice congrats! sales call?

green elk
#

yup thats right

inner heart
#

letsgoo

#

if this can help here's how I structure multi step flows in my prompt:

## Flows
### Make Reservation Flow
If the user wants to make a reservation, follow this flow:
1) Collect the following details from the user one question at a time. If any information provided is unclear or appears incorrect, politely ask clarifying questions:
   a) **Date**. If the date falls on a closed day, say the restaurant is closed that day, share the open days and hours, and ask for a new date. **Do not ask for a time until an open day is chosen.**
   b) **Time** (make sure that the time is within opening hours).
   c) **Party size** (minimum party size of 1 and maximum of 20).
If the reservation is not within opening hours: Go back to [Make Reservation Flow step 1].
If party size is less than 1 or greater than 20: Go back to [Make Reservation Flow step 1].
Else: Go to [Make Reservation Flow step 2].
2) Call `demo_dev_check_for_availability`.
If the desired time slot isn't available: Go back to [Make Reservation Flow step 1].
3) Ask **naturally** where the user wants to sit based on the seating options returned by `demo_dev_check_for_availability`.
If the user wants to change the date, time OR party size: Go back to [Make Reservation Flow step 2].
If the caller picked a seating among the available options AND doesn't change the date, time or party size: Go **directly** to [Make Reservation Flow step 4].
4) Call `demo_dev_check_for_experiences` with the collected information.
If no experiences are available: Do NOT mention that no experiences are available and go **directly** to [Make Reservation Flow step 8].
green elk
#

neat i am wondering if the promp should also explicitly talk about entry and exit nodes

inner heart
#

im actually not using the workflow

#

i've tried it but I prefer the flexibility of the "one big agent"

#

also with better llms coming out that follow instructions better it's nice

green elk
#

i feel like i am might need to switch to that, but i am worried my prompt will be too huge!

inner heart
#

i've got a 300+ lines prompt and it's doing fine

#

with gpt oss 120b

#

with gemini 3 flash it'll fly

green elk
#

hmm yeah mine is much bigger and has a lot more steps

#

a lot of if/else conditions i will need to take care

inner heart
#

I've got 3 different flows with respectively 11, 3 and 12 steps where each step can have multiple conditions

#

for reference

trim jacinth
#

Workflows should work fine if used correctly. As said we have roughly 40k tokens in our agent if you count every node. No way to handle this in a single prompt

#

I see a couple of issues with your prompt. I will write something up and send it. Also have 5 b2b demos now

inner heart
trim jacinth
green elk
inner heart
#

im curious haha

trim jacinth
inner heart
#

very nice

#

using framer too haha it's goated

trim jacinth
#

true hahaha

#

Our main gtm channel is meta ads.

inner heart
#

then web app in flutter using the app subdomain?

trim jacinth
#

yes exactly.

#

But our customers mainly rely on the email notifications for new calls

inner heart
trim jacinth
#

thats great. What niche?

inner heart
#

restaurants

trim jacinth
#

Looks great!

#

nice website

inner heart
#

thanks!

#

web app in react also using the app sub domain

trim jacinth
#

nice! how many customers are you at?

inner heart
#

just started marketing at the start of the month so 0 rn and very few restaurants want to launch something at the end of the year but got 2 calls today and multiple meetings in jan

trim jacinth
#

Yeah q4 is tough!

inner heart
#

yeah especially restaurants with so much business near christmas haha but jan will be peak

#

they'll have the Christmas money

trim jacinth
#

I have heard every possible way of saying "eoy is hell, gtf away from me". (I just got timed out from the discord hahahah)

inner heart
#

how many are you guys in the team?

trim jacinth
#

@green elk so the agent is just not doing anything after triggering ch_confirm_selection?

trim jacinth
inner heart
#

sick and you're cto?

trim jacinth
#

3 engineer 2 gtm/business. I am one of the founders. We are both technical but by now i only do prompt and sales

#

So my cofounder would be cto

#

I am only doing prompting and agent stuff since he doesnt like it. He does all the app stuff

green elk
#

and now it finally seems to be switching

#

still sucks that i need to use 3.7 for this!

trim jacinth
#

is it slower?

green elk
#

yeah defintely

#

also you mentioned some issue with prompts, anything serious ?

trim jacinth
#

So there are 3 things I would try.

  1. be much more specific in the main prompt about the transition to the next step. We have most of the logic for transitions in the main prompt and not the condition. F.E this should be in the main prompt:

State Transition (Auto-Proceed Rule)

  • Immediate Action: As soon as ch_confirm_selection is called with selection_type: "INSURANCE", the "Insurance Selection" phase is considered COMPLETE.
  • No-Pause Execution: Do NOT wait for a user response after this tool call. You must immediately transition to Phase 2: Location Selection by asking the first question of that phase in the same response.
  1. Split up the nested conditionals in the main prompt into multiple workflow steps. What has worked really well for us is having 1 node handling a single responsibility (SRP also in agents not only in code) and then overwriting the prompt.

The idea here would be to have a single node for each of the three "ASKS" that way the logic the agent needs to handle at a single point is much simpler.

  1. Try it with gemini 3 flash or gpt-4.1

The main bottleneck for tool calling remains the model. Testing for us shows that Claude 4.5 Haiku and 3.7 Sonnet struggle with state management in this specific architecture. I recommend prioritizing GPT-4.1 superior reasoning for logic, then iteratively refining its conversational tone, rather than attempting to 'force' better logic out of more conversational models like Claude.

What i can offer you is testing out your agent iterating. For that i would need access to all the prompts or the platform.

green elk
#

oh wow thanks man, will look into this and see what i can do!!

neat ruin
# inner heart with gpt oss 120b

You prefer gpt oss 120b compared to gemini 2.5 flash ? I did a lot of testing some weeks ago and found nothing better than gemini 2.5 flash, and it was even better when I let gemini 2.5 flash using 50 tokens before answering ( ~0.7 to 1.5 time to answer with the thinking). Did you try ?

green elk
#

So the LLM on individual node is 2.5 flash and how about your system prompt LLM, what do you use for that

inner heart
#

they're pretty similar, but I'll switch to 3 flash soon anyways so it doesn't really matter

inner heart
neat ruin
#

Yess mdrr

neat ruin
inner heart