Available Context

@_platform-references/org-variables.md

Ops AI Transform

Goal

Apply AI-powered transformations to entire columns in an Ops table -- generating new values, rewriting existing data, deriving insights from other columns, or standardizing messy data. Unlike enrichment (which pulls in external data), transforms work with the data already in the table to produce derived or improved values.

Think of it as having a smart assistant process each row: "For every lead, write a personalized opening line based on their name, title, and company" or "Standardize all the job titles to seniority levels."

Required Capabilities

Ops Tables: Access to AI transform APIs

Inputs

table_id: ID of the ops table
column_id: ID of the target column to write transformed values into
prompt: Describes what the AI should generate or transform (e.g., "Write a one-line personalized opener referencing the lead's role")
source_columns: Column names that the AI should read from when generating values (e.g., ["Name", "Title", "Company"])

Instructions

Running a Transform

When the user wants to AI-generate or transform column values:

Identify the table and target column
Understand the transformation. Ask clarifying questions if needed:
- What to generate: "What should each cell contain?"
- What context to use: "Which columns should the AI look at?"
Build the prompt. Good prompts are specific about:
- The output format (one sentence, a number, a category, etc.)
- The source data to reference (which columns)
- The tone or style (professional, casual, technical)
Preview before running. Show the user what will happen:
- "I'll transform 142 rows in the [Column] column using: '[prompt]'"
- "Source columns: Name, Title, Company"
Call execute_action("ai_transform_ops_column", { table_id: "<id>", column_id: "<col_id>", prompt: "<prompt>", source_columns: [...] })
Report results with sample outputs so the user can verify quality

Common Transform Patterns

Use Case	Prompt Template	Source Columns
Personalized opener	"Write a one-line cold outreach opener for {Name} who is {Title} at {Company}"	Name, Title, Company
Seniority classification	"Classify this job title into: C-Level, VP, Director, Manager, or IC"	Title
Company size bucket	"Categorize into: Startup (1-50), SMB (51-200), Mid-Market (201-1000), Enterprise (1000+)"	Employee Count
Email domain extraction	"Extract the domain from this email address"	Email
Industry standardization	"Map this industry to one of: SaaS, Fintech, Healthcare, E-commerce, Manufacturing, Other"	Industry
Outreach angle	"Based on this person's role and company, suggest the best angle for a cold outreach"	Title, Company, Industry
LinkedIn message	"Write a 3-sentence LinkedIn connection request mentioning their role and a relevant talking point"	Name, Title, Company

Building Good Prompts

Help the user craft effective prompts:

Be specific about output format: "One sentence" is better than "short"
Reference source columns explicitly: "Using the Name and Company columns"
Give examples: "For a VP of Sales at Acme Corp, output something like: 'Hi Sarah, I noticed Acme is scaling its sales team...'"
Set constraints: "Max 100 characters" or "Choose from: High/Medium/Low"

Reviewing Transform Results

After a transform completes:

Show 3-5 sample outputs so the user can assess quality
Ask: "Does this look right? I can adjust the prompt and re-run if needed."
If the user wants changes, update the prompt and re-run -- previous values will be overwritten

Available Actions

Action	Parameters	Returns
`ai_transform_ops_column`	`{ table_id: string, column_id: string, prompt: string, source_columns?: string[] }`	Transform job with status, progress, and sample outputs

Output Format

Transform Started

AI TRANSFORM STARTED
  Table: Lead Prospects
  Column: Outreach Opener
  Rows: 142
  Prompt: "Write a one-line personalized cold outreach opener for {Name}, {Title} at {Company}"
  Source columns: Name, Title, Company

Processing...

Transform Complete with Samples

AI TRANSFORM COMPLETE
  142 of 142 rows processed

SAMPLE OUTPUTS:
  John Smith (VP Sales, Acme Corp):
    "Hi John -- saw Acme just expanded its sales team. Curious how you're handling..."

  Sarah Chen (Director Eng, TechFlow):
    "Sarah, TechFlow's engineering velocity is impressive. We help teams like yours..."

  Mike Ross (CTO, DataBridge):
    "Mike -- as DataBridge scales its data platform, I'd love to share how we help CTOs..."

Look good? I can adjust the prompt and re-run if you want a different style.

Error Handling

No prompt provided

"To transform this column, I need to know what to generate. Here are some ideas:

'Write a personalized outreach opener based on Name, Title, and Company'
'Classify the job title into seniority levels: C-Level, VP, Director, Manager, IC'
'Extract the company domain from the email address' What would you like?"

Source columns have empty data

"Some source columns have missing data -- [X] rows are missing [Column Name]. The AI will skip those rows or produce lower-quality output. Want to fill in the gaps first, or proceed with what we have?"

Transform produces poor quality

If the user reports the output quality is not satisfactory: "Let me refine the prompt. What specifically needs to change? For example:

Too formal/casual?
Wrong length?
Missing specific context?
Not personalized enough?"

Column already has data

"The [Column Name] column already has data in [X] rows. Running a transform will overwrite existing values. Want to proceed, or should I only fill in the empty cells?"

Guidelines

Always preview the transformation before running: show the prompt, source columns, and row count
Show sample outputs after completion so the user can assess quality before using the data
For classification/categorization transforms, suggest a fixed set of output values to ensure consistency
Encourage users to reference multiple source columns for richer, more personalized outputs
If the user's prompt is vague, suggest a more specific version before running
Large tables (500+ rows) may take a few minutes -- set expectations upfront
Transforms are idempotent -- re-running with the same prompt will regenerate all values