Using GitHubs Actions & Prompt Bundling to Provide Compliance Guardrails

Why it matters

• Prompt text is business logic. Untracked changes can introduce risk or inconsistent behavior.
• Most compliance frameworks require that production code (including prompts) is:
  • Immutable after deployment
  • Peer‑reviewed before release
  • Traceable with a full audit trail
• Bundling a prod Prompt Template into an artifact that lives in Git provides these guarantees with almost zero additional tooling.
• This also protects your application in the unlikely event that you're application cannot connect to the Freeplay platform

Prerequisites for this guide

• GitHub repository
• GitHub Actions enabled
• Freeplay CLI
• A Prompt Template already promoted to the prod environment

Step 1  ·  Bundle the production prompt

Run during your release pipeline:

freeplay download \
  --env prod \
  --output bundled_prompts

This writes an immutable JSON artifact containing:

• Prompt text
• Model/provider selection
• All request parameters

Step 2  ·  Open a pull request if the bundle changes

Add the following workflow file to .github/workflows/prompt-bundle-guard.yml:

name: Update Bundled Prompts
on:
  push:
    branches: [main]

jobs:
  prompt-bundle-guard:
    runs-on: ubuntu-latest
    permissions:
      contents: write          # push temp branch
      pull-requests: write     # open PR
    steps:
      - uses: actions/checkout@v4

      - name: Install Freeplay CLI
        run: python -m pip install --upgrade freeplay

      - name: Re‑bundle prod prompts
        run: |
          freeplay download \
            --env prod \
            --output bundled_prompts

      - name: Open PR if bundle changed
        env:
          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
        run: |
          if git diff --quiet --exit-code -- bundled_prompts; then
            echo "No prompt updates detected." && exit 0
          fi

          git config user.name  "Freeplay Prompt Bundler"
          git config user.email "[email protected]"
          branch="auto/prompt-bundle-$(date +%s)"
          git switch -c "$branch"
          git add bundled_prompts
          git commit -m "chore(prompts): update prod bundled prompt"
          git push --set-upstream origin "$branch"

          gh pr create \
            --title "Update production bundled prompts" \
            --body "Automated PR generated by Prompt Bundler workflow." \
            --label "prompt-bundle" \
            --base main \
            --head "$branch" \
            --draft

What happens?

1. Bundle – Regenerates the prod bundle.
2. Diff – If anything changed inside bundled_prompts/, the workflow:
   • Pushes the change on a new auto/prompt-bundle-* branch.
   • Creates a draft PR that a human must review & merge.
3. Once merged, the new bundle is locked in Git history.

Tip: Protect main with “Require PR approval” to enforce peer review.

Step 3  ·  Pin the Bundled Prompt for production systems at runtime

from pathlib import Path
from freeplay import Freeplay
from freeplay.resources.prompts import FilesystemTemplateResolver

# Point the resolver at your committed bundle directory
fpClientLocal = Freeplay(
    freeplay_api_key=freeplay_key,
    api_base=freeplay_api_base,
    template_resolver=FilesystemTemplateResolver(Path("bundled_prompts"))
)

# Invoke the production template as usual

The SDK reads the local bundle, so the application can only execute prompts that passed the PR gate.

Compliance mapping

FAQ & Troubleshooting

Q : What if we maintain multiple prod environments (e.g., per‑tenant)?
A : Run freeplay download --env prod-<tenant> (or similar, given your environment naming conventions) for each environment and store each bundle under its own path.

Q : Can I use Bitbucket Pipelines or GitLab CI instead?
A : Yes mirror the same logic: re‑bundle, diff, and open a merge request when changes are detected.

Q : How do I invalidate a bad prompt quickly?
A : Revert the bundle commit or promote a previous prompt template version in the Freeplay dashboard and re‑run the workflow.