Im running a prompt to generate an email opener (using the prospect first name) . Currently using sonnet for the gpt engine (as have been told it writes better than chatpgt). An issue im facing is that it's generating "Sarah" for like 250/800 rows, depsite not being present in the First name field NOR present in the prompt. Any ideas? Clay table: https://app.clay.com/workspaces/443643/workbooks/wb_9tHV7fsip34Z/tables/t_GgbQbsYhv7hb/views/gv_55uFmuM3QJ8F
When i force the prompt to run again, the name doesn't change - it's so frustrating haha
This issue might be related to how your prompt is configured or how the AI model is interpreting your data. Here are a few potential causes: 1. The prompt might contain "Sarah" as an example somewhere, causing the model to default to it when uncertain. 2. There could be an issue with how your output fields are defined in the prompt configuration. Make sure they're correctly mapped to pull the actual first name from your table. 3. Check if you've set the first name field as "Required to Run" in your Prompt Editor. Without this setting, the AI might default to a placeholder name when the field is empty. Try reviewing your prompt template and ensuring it's properly referencing the dynamic first name field from your table. You might also want to adjust the creativity settings if they're set too high.
This issue is likely happening because your prompt may have "Sarah" as an example somewhere in the template. When using AI-generated email openers with Clay, the model can sometimes default to example values if: 1. The first name field isn't properly mapped in your prompt configuration 2. The first name field isn't set as "Required to Run" in your Prompt Editor 3. There's a reference to "Sarah" in your prompt template that's being used as a fallback When you rerun the prompt without changing these configurations, the same output will be generated because the AI is following the same instructions with the same inputs. Try reviewing your prompt template, checking your field mappings, and ensuring the first name field is properly connected to your data source.
The prompt does not contain Sarah, see prompt below:
#CONTEXT# Youāre writing the first sentence of a cold email to a finance leader in a small NZ company. This is the punchy one-liner and must be a question, so that the recipient can identify themselves. #OBJECTIVE# Rephrase the provided Personalized Finance Problem Problem Statement into a short, clean, natural opening line. It should reflect the reality of what this person is stuck doing ā and what itās costing them. #INSTRUCTIONS# 1. Read the Personalized Finance Problem Problem Statement carefully. 2. Write a single sentence, no more than 15 words. 3. This line should: - Sound like a calm observation or peer comment - Acknowledge a real bottleneck or frustration - Sit naturally at the top of an email, especially when viewed on mobile 4. You may reference their First Name or Company Name if it adds to the impact, but itās optional. 4. You may use **industry-native abbreviations or shorthand** if they feel natural and specific to the task. These include: - āEOMā instead of āend of monthā - ārecsā instead of āreconciliationsā or āreconcilingā - āAPā for accounts payable - āboard packā (common NZ shorthand) Only use these if they sound like something a finance lead would say in a message to their colleague or team. #FORMATTING INSTRUCTION# Return a single sentence with no trailing or leading line breaks or empty lines. Output should be a single string, not a paragraph block. - DONāT include a new line or return at the end of the output. - DONāT start the sentence on a new line. #AUDIENCE CONTEXT# - This person is a finance lead at a New Zealand company with 11ā50 employees - Theyāre doing reporting, reconciliations, compliance, and month-end themselves or with limited help - Strategic work keeps getting postponed because thereās not enough capacity - Theyāre stretched, not lazy ā donāt shame them, just reflect the tension #ANTI-AI GUARDRAILS# - Avoid rhetorical questions or sales hooks - No techy verbs (āoptimize,ā āscale,ā āstreamlineā) - Avoid symmetry (donāt write āstuck doing X when they should be doing Yā) - No poetic language or startup clichĆ©s (āwearing all the hats,ā āburning the candleā) - No fancy punctuation (no em-dashes, ellipses, quotation marks, or excessive commas) - Avoid unfamiliar or forced abbreviations ā use only terms a finance lead would naturally say - Donāt explain or spell out the abbreviations ā use them as the reader would #DOs# - DO sound like a peer describing what they see - DO write plainly and clearly - DO anchor the sentence in one tangible thing (e.g., āstill reconciling,ā āburied in templates,ā ācanāt start planningā) - DO include the recipients first name and always in the format of Name,. Never Name:. #DONāTs# - DONāT pitch or hint at your solution - DONāT force cleverness ā quiet realism beats slick language - DONāT use abstract phrasing (āstuck in the weedsā / āspinning platesā) - DON'T reference anywhere other than New Zealand (ie. Pacific, Australia etc). #EXAMPLES# Example input ( Personalized Finance Problem Problem Statement š āYouāre constantly juggling last-minute requests for detailed financial reports on network upgrades and capital projects, often having to pull data together yourself after hours because your team doesnāt have enough hands.ā Example outputs: - Still reconciling month-end reports while the infrastructure work keeps backing up? - Capex trackingās on fire and thereās no time to even look at ROI. - Compliance reporting eats the day before anything strategic even starts. - Francois, your forecasting workās buried behind reconciliation and board pack prep again.
This issue might be happening because the AI model is defaulting to "Sarah" as a placeholder when it can't properly access or process the first name field from your data. To fix this problem: 1. Check that your first name column is properly mapped in the prompt configuration 2. Make sure the first name field is set as "Required to Run" in your Prompt Editor 3. Try using a different AI model - you mentioned using Sonnet, but you could test with another model 4. Clear any cached results and force a complete refresh of your data If your prompt includes examples (like the one with "Francois"), the model might be getting confused. Try removing all example names from your prompt instructions.
Hey there Alex thanks for reaching out, sorry for the delay here, taking a look at this the cause of this "Sarah" issue is occurring due to the mention of the name Sarah inside of the example outputs as shown below. Essentially what is happening here is that the AI is getting confused by this section of the prompt and is defaulting to the name Sarah instead here.