Plan libraries before you run workflows
- Build libraries before running workflows whenever you already know key people, organizations, or prompt content.
- Treat libraries as shared inputs, not workflow‑specific fixes.
- Start with high‑confidence entries to avoid introducing noise early.
Use the right library type
- Person libraries. Use for known individuals who are likely to appear repeatedly in the data.
- Organization libraries. Use for known companies, firms, domains, or other organizations.
- Prompt libraries. Use for approved, reusable prompt text that should remain consistent across workflows.
Keep entries clean and normalized
- Use a single, canonical version of each person or organization name.
- Avoid duplicate entries that represent the same entity.
- Remove outdated or incorrect entries during periodic reviews.
- Prioritize clarity over completeness. Large, unreviewed libraries can reduce accuracy.
Update libraries during review cycles
- Add new people or organizations discovered during workflow review.
- Feed corrected workflow output back into libraries where appropriate.
- Update prompt libraries when prompt guidance changes or improves.
- Treat library maintenance as part of the review process, not a one‑time setup step.
Reuse libraries across workflows and matters
- Share libraries across workflows that involve the same participants or objectives.
- Reuse prompt libraries to enforce consistent instructions and outputs.
- Avoid creating duplicate libraries for the same purpose unless required by data separation needs.
Version and govern prompt libraries carefully
- Store only approved or validated prompt content in prompt libraries.
- Make prompt changes deliberately and document the reason for changes outside the library if required.
- Avoid experimental or workflow‑specific prompts in shared prompt libraries.
Review libraries regularly
- Schedule periodic reviews to:
- Remove duplicates
- Correct errors
- Retire unused entries
- Validate that libraries still reflect current matters and workflows.