Extract Reference Feature Bag
# Extract Reference Feature Bag
Use this prompt when converting a public AI/MCP reference into a normalized
`xuunity.reference-watch.feature-bag.v1` JSON object.
## Inputs
- source id from `Modules/AIReferenceWatch/reference_sources.yaml`
- source URL
- reviewed files or pages
- reviewed timestamp in UTC
- focus areas to inspect
## Instructions
1. Extract only capabilities supported by source evidence.
2. Prefer code registry, tool schema, manifest, and repo-local docs over README
marketing text.
3. Mark each capability as one of:
- `implemented`
- `claimed`
- `unknown`
- `contradicted`
- `not_a_base_goal`
4. Use `claimed` for public docs claims that were not verified in code.
5. Use `unknown` when reviewed sources do not make a clear claim.
6. Use `contradicted` when docs and code disagree.
7. Do not infer feature parity from a broad phrase such as "editor tools".
8. Decompose grouped tools before treating them as capability evidence.
9. For each capability, decide whether it is a direct analog:
- `directAnalog: true` when the reviewed reference capability directly maps
to the target capability.
- `directAnalog: false` when the reference has related implemented code but
is broader, narrower, or materially different.
- Use `analogTarget` and `analogNotes` to explain the mapping.
10. For each capability, include provenance:
- `confidence`
- `evidenceType`
- `focusAreas`
- `sourceFiles`
- `sourceLines` when available
- `extractionMethod`
- `lastReviewedAtUtc`
- `reviewer`
- `notes`
## Evidence Checklist
For every capability, answer these questions in the capability metadata or
notes:
- What exact capability is being evaluated?
- Is this a direct analog to the target capability?
- Which file proves registration, schema, or manifest presence?
- Which file proves implementation?
- Is the capability only documented but not code-confirmed?
- Is a broad grouped tool being decomposed into narrower capabilities?
- What related capabilities were searched but not found?
- Should the status be `implemented`, `claimed`, `unknown`, or `contradicted`?
## Output
Return one JSON object matching:when to use it
Community prompt sourced from the open-source GitHub repo FoxsterDev/ai-research-hub (no explicit license). A "Extract Reference Feature Bag" 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.
tags
roleplaycommunitygeneral
source
FoxsterDev/ai-research-hub · no explicit license