Prompt Management CMS
A custom-built prompt management interface integrated directly into the SaaS CMS admin panel. Instead of jumping between Langfuse, the codebase, and the editorial dashboard, this UI syncs all 50+ production prompts into a single control plane, with model routing, version tracking, and category filtering built into the same workspace where content is managed.
Tech Writer Prompt: v13 with tone, linking, and attribution rules
Highlights
- 50 versioned prompts across Article, Pipeline, Social, and Media categories
- Multi-model routing: GPT-4o for editorial, Gemini Flash for extraction and enrichment
- Structured editorial rules with anti-hallucination gates for linking and data accuracy
- Version history with production tagging. Roll back or A/B test any prompt instantly
Why Build This
Managing 50+ prompts across Langfuse, the codebase, and an editorial dashboard means constant tab-switching. Every prompt edit requires checking the Langfuse UI, verifying the version in code, then confirming the pipeline picks it up. This CMS panel eliminates that friction by syncing everything into the same admin interface where articles are published and pipelines are monitored.
How It Works
The CMS prompt manager syncs from Langfuse's API and presents prompts in a familiar admin table:
- Sync: prompts are pulled from Langfuse with their metadata (model, category, version, status)
- Browse: filter by category, search by name, sort by version
- Inspect: view the full template, model configuration, temperature, and version history
- Edit: modify prompts in the CMS editor, changes push back to Langfuse
Prompt Architecture
The registry spans 50 prompts across 5 domains, each with its own model routing:
| Domain | Examples | Model | Purpose |
|---|---|---|---|
| Editorial | finance-analyst, tech-writer, sports-analyst | GPT-4o | Vertical-specific rewriting with domain expertise |
| Article | article-format, article-constraints, seo-meta-optimizer | GPT-4o / Gemini | Structural formatting and SEO |
| Pipeline | extract-companies, key-points, actionable-insights | Gemini Flash | Structured data extraction |
| Social | repurpose-twitter, repurpose-tiktok, social-enrich | GPT-4o | Cross-platform content adaptation |
| Media | media-script-audio-summary | Gemini 2.5 Flash | Audio and video script generation |
Engineering Decisions
Version pinning over latest: Production pipelines always run a pinned prompt version. New versions can be drafted, tested, and promoted without affecting live traffic, following the same pattern as database migrations.
XML directives for structured rules: Each editorial prompt includes a <ToneAndVoice> block with hard-coded guardrails: banned words, linking hierarchies, and citation rules. XML was chosen over freeform instructions because models respect structured markup more reliably than prose.
Anti-hallucination as code, not comments: URL linking rules, data accuracy constraints, and "never invent" directives are enforced at the prompt level. The pipeline also validates outputs downstream, but the prompt is the first line of defense.
By The Numbers
- 50 active prompts across the registry
- v14 highest version count (sports-analyst, the most iterated prompt)
- 5 model configurations routing to GPT-4o, Gemini 2.0 Flash, and Gemini 2.5 Flash
- 9 prompts updated in the last 7 days
