Docs/AI Prompts

AI Prompts

Upshot's AI synthesis is fully configurable. Customize how data is interpreted, how reports sound, and what gets filtered out.

Two-layer prompt system

Upshot uses two types of prompts that work together:

Interpreter prompts (per-source)

Each data source has an interpreter prompt that tells the AI how to understand that source's data. For example, the GitHub interpreter knows how to read PR metadata, understand commit messages, and recognize deployment events.

You can customize interpreter prompts to teach Upshot about your conventions:

# Custom GitHub interpreter guidance
- PRs prefixed with [HOTFIX] are emergency fixes
- The "platform" label means infrastructure work
- Deploy tags follow the format: v{major}.{minor}.{patch}
- Commits mentioning JIRA-XXX should link to Linear issues

Output prompts (per-audience)

Each audience (Team, Leaders, Customers, Market) has its own output prompt that controls voice, tone, structure, and level of detail.

# Custom Leaders output prompt
- Use business language, avoid technical jargon
- Lead with impact: revenue, customers, velocity
- Include quarter-over-quarter comparisons
- End with strategic recommendations
- Keep to 500 words max

Customizing prompts

  1. Navigate to Settings → Prompts
  2. Select the prompt type (Interpreter or Output) and the specific source/audience
  3. Edit the prompt text — you'll see a live preview with recent data
  4. Save. New reports will use the updated prompt immediately.
💡 Tip: Start with the default prompts and make small tweaks. The defaults are well-tuned — most teams only need to add org-specific conventions (label meanings, branch naming, deploy processes).

Sensitivity filtering

Upshot automatically detects sensitive content and filters it from customer-facing outputs. This includes:

  • Security issues — vulnerability fixes, CVE patches, auth changes
  • Payment/billing bugs — pricing errors, payment processing fixes
  • Internal tooling — CI/CD changes, internal dashboards, developer ergonomics
  • Data incidents — data loss, migration issues, privacy-related fixes

Sensitivity detection is AI-powered and configurable. You can:

  • Adjust the sensitivity threshold (strict / balanced / permissive)
  • Add custom keywords or labels that should always be filtered
  • Override individual items during report review
  • Choose which audiences are affected (typically only Customers and Market)

Prompt versioning

Upshot keeps a history of prompt changes. If a new prompt produces worse results, you can roll back to any previous version from the prompt settings page.