Prompt Template Versioning

Prompt Template Versioning

Prompt Builder templates drift because authors iterate in Setup without a change-log or rollback plan. This skill defines a three-stage lifecycle (Draft → Candidate → Active) stored as custom-metadata pointers, promoted with a signed-off checklist, and rolled back via the same CMDT.

When to Use

Any Prompt Builder template that (a) is referenced by a production Flow/Agent or (b) has user-visible output. Not needed for ad-hoc developer experiments.

Typical trigger phrases that should route to this skill: how do I version a prompt builder template, rollback a prompt in production, a/b test a prompt variant, audit which prompt was live last week.

Recommended Workflow

1. Create a CMDT Prompt_Template_Binding__mdt with fields DeveloperName, Target_Slot__c (e.g. 'SalesEmail'), Version__c, Active__c.
2. Author new prompt versions in Setup with suffixed DeveloperNames (SalesEmail_v3, _v4).
3. Update consumers (Flow/Apex/Agent) to read the DeveloperName from Prompt_Template_Binding__mdt at invocation — never hardcode the template name.
4. Run the fixture test set against _v4; record metrics (accuracy, latency, token cost) in a release note.
5. On promotion: flip Active__c on CMDT from _v3 row to _v4 row; on rollback, reverse. Both are deploy-validated CMDT changes — reviewable in git.

Key Considerations

• Prompt Builder has no native version history — once you save, the old text is gone. The CMDT pointer + named versions give you git-backed history.
• Flow-referenced prompts are bound by DeveloperName; if you rename, the Flow breaks. Use the CMDT layer of indirection.
• Retest whenever the underlying model changes (model upgrades happen transparently).

Worked Examples (see references/examples.md)

• CMDT-backed template binding — Sales email prompt updated weekly.
• Canary rollout via user-hash bucketing — Shipping _v4 to 10% of reps before going full.

Common Gotchas (see references/gotchas.md)

• Setup UI 'Save' loses the prior text — You can't compare v3 vs v4 without an external copy.
• Model silently upgrades — Identical prompt, different output next week.
• Bound variable schema change — v4 needs a new {{record.Field__c}} that doesn't exist in your sandbox.

Top LLM Anti-Patterns (full list in references/llm-anti-patterns.md)

• Hardcoding template DeveloperName in Flow/Apex — rollback requires redeploy.
• Editing prompts in place — no diff, no rollback.
• Skipping fixture tests 'because the change is small' — prompts amplify small changes.

Official Sources Used

• Agentforce Developer Guide — https://developer.salesforce.com/docs/einstein/genai/guide/agentforce.html
• Einstein Trust Layer — https://help.salesforce.com/s/articleView?id=sf.generative_ai_trust_layer.htm
• Invocable Actions (Apex) — https://developer.salesforce.com/docs/atlas.en-us.apexref.meta/apexref/apex_classes_invocable_action.htm
• Agentforce Testing Center — https://help.salesforce.com/s/articleView?id=sf.agentforce_testing_center.htm