Centralized AI prompt library (L2)

The steps

1. 01

   Select and configure your source of truth

   The most common failure in prompt management is "tribal knowledge" hidden in private chats. You must choose a centralized, searchable platform. If your team is non-technical, use Notion or Airtable. If you are developer-heavy, use a Git repository or a dedicated tool like PromptLayer or Portkey. In Notion, create a 'Gallery View' database named 'GTM Prompt Library.' • Add the following mandatory columns: Prompt Name (Title), Category (Sales, SDR, CS, Marketing), Version (e.g., v1.2), Model (GPT-4o, Claude 3.5), Status (Production, Testing, Deprecated), and Last Verified Date. • Who: RevOps Lead. • Time: 1 hour. • Pitfall: Using a Slack channel or a Google Doc; these lack the metadata filtering required to scale past five prompts. • Definition of Done: A live, empty database with the defined schema accessible to the whole team.

2. 02

   Standardize the prompt entry schema

   A prompt is useless without context. For every entry, you must document the 'System Instructions,' the 'User Template,' and the 'Variable Placeholders' (e.g., {{company_description}}). • Use the 'Variable' format consistently so users know what to swap out. Example User Template: 'Draft a three-sentence outreach email for {{prospect_name}} at {{company_name}} focusing on their recent {{trigger_event}}.' • Assign a 'Prompt Owner' for every entry,this person is responsible for the prompt's accuracy. • Who: Department Managers (e.g., Head of SDRs). • Time: 30 minutes per prompt. • Pitfall: Copy-pasting a long block of text without explaining which parts the user needs to change. • Definition of Done: A completed entry including high-level intent, specific model settings (Temperature: 0.7), and clear variable markers.

3. 03

   Attach test cases and gold standards

   To ensure consistency, every prompt must include 'Gold Standard' test cases. This consists of a specific input and the expected output. In your library, create a sub-section for 'Test Bed' where you list 2-3 links to specific LLM conversations (using ChatGPT's 'Share Link' or Claude's 'Share' feature) that show the prompt working perfectly. • Include a 'Negative Test' example,basically, what the prompt should *not* do (e.g., "Ensure it doesn't sound overly formal"). • Who: The Prompt Owner. • Time: 1 hour per prompt. • Pitfall: Assuming a prompt 'works' after one trial. Without test cases, you won't know if a model update broke your prompt later. • Definition of Done: Each prompt entry has at least two links or screenshots showing a successful output based on a specific input.

4. 04

   Establish a contribution and review workflow

   Open your library to the team but restrict 'Edit' access to Prompt Owners. Create a 'Prompt Submission Form' (using Notion Forms or Typeform) where any employee can suggest a new prompt or an improvement. • The form should ask: What problem does this solve? What is the prompt? What model did you use? • Marketing/RevOps should then 'Vibe Check' the submission before moving it from 'Review' to 'Production' status. • Who: RevOps + Team Leads. • Time: 2 hours to set up the workflow. • Pitfall: Allowing everyone to edit the main library, which leads to version chaos and accidental deletions. • Definition of Done: A submission workflow is active, and the core library is 'Read-Only' for the general staff.

5. 05

   Quarterly audit and ruthless deprecation

   AI models evolve every 3-6 months. A prompt that worked for GPT-4 might behave differently on GPT-4o. Set a recurring calendar invite for the first Monday of every quarter. • During this review, the Prompt Owner must run the 'Gold Standard' test cases again. If the output quality has dropped, the prompt must be updated or moved to 'Deprecated.' • If a prompt hasn't been used (track this via a simple 'Check-in' button or view counts), delete it. A cluttered library is a neglected library. • Who: RevOps Lead. • Time: 2-4 hours per quarter. • Pitfall: Keeping 'Legacy' prompts because they "might be useful." This creates noise and causes reps to use outdated tactics. • Definition of Done: A 'Last Reviewed' timestamp on all production prompts that is less than 90 days old.

Next playbooks

• AI literacy bootcamp for GTM (L2)
• Claude for Work shared projects (L2)
• Notion AI + Claude for company knowledge ops (L2)
• AI council + intake process (L3)