Prompt Versioning

GPTClaudeDeepSeek··▲ 741 copies·updated 2026-07-14

prompt-versioning-3.prompt

# Prompt versioning

Prompts and model choices are configuration, not code (ADR-006).

## Shape

A prompt config is a row: task, model id, parameters, a system template, and a
user template. Templates use `{{variable}}` placeholders filled at render time
(`src/ai/templates.ts`). One config is active per task.

## Lifecycle

A change is an INSERT or UPDATE, captured by a trigger into
`prompt_config_history`. The AI client loads the active config for a task at
request time and never sees a hardcoded model id or prompt string. Rolling back
is re-activating a prior version, not a code revert.

## Why it matters

- The exact configuration behind any past output is reconstructable.
- The eval harness (ADR-010) grades a specific version, so a quality change is
  attributable to a specific config.
- A wording tweak does not require a deploy.

In this reference implementation the active config ships in code
(`src/reports/prompts.ts`, `evals/judge.ts`) so the path is self-contained; the
table is the production source, and the shape is identical.

fill the variables

This prompt has 1 variable. Pro fills them into a ready-to-paste prompt for you — no manual find-and-replace.

{{variable}

Unlock with Pro →

when to use it

Community prompt sourced from the open-source GitHub repo chrisrollins-labs/rag-report-platform (MIT). A "Prompt Versioning" style prompt — adapt the placeholders and specifics to your task. Imported as-is and not independently retested here, so check the output before relying on it.