buildbetter-review-improve-smart-tag
Installation
SKILL.md
Review and improve a Smart Tag
Requires the BuildBetter MCP release that exposes list-tag-groups,
list-tag-group-tags, and the durable Smart Tag workflow tools below. Do not
install or distribute this skill against an older production MCP inventory.
Workflow
- Call
list-tag-groupsandlist-tag-group-tagsto resolve the canonicalmemberId, then callget-smart-tag-workflow. Do not create a separate tag or group. - If the live tag is published, call
create-smart-tag-revisionor resume its existing candidate. The candidate keeps the same member, tag, group, and target while the published revision continues live. - Poll any pending evaluation or improvement with
get-smart-tag-operation; do not edit the candidate or start concurrent work. - To run a fresh regression, disclose the bounded date range, example limit, and maximum credits, obtain approval, generate an idempotency UUID, and call
run-smart-tag-evaluationwithincludeAlreadyProcessed: truewhen recently classified records should be eligible. - After the requested run completes, present every case with its prediction and evidence. Ask the user to mark
should matchorshould not matchwhen their intent is not already explicit. Capture a concrete exclusion reason for false positives and a positive pattern for false negatives. - Call
review-smart-tag-evaluationwith the complete set of judgments and current revision. Report true/false positives and negatives, precision, recall, pre-filter recall, F1, corpus balance, and readiness. - If pre-filter recall is weak, revise approved terms and run a new evaluation. Instruction-only improvement cannot recover examples filtered out before classification.
- For a fully reviewed but unready corpus, let
Nbe the locked-corpus size. Show the approved ceiling: one improvement iteration costsNcredits and three iterations can cost3N. Obtain approval, generate an idempotency UUID, and callimprove-smart-tagwithmemberId, current revision,maxCreditBudget, and at most five iterations. - Poll the operation until terminal. Report exact credits, instruction diff, rationale, metrics, and remaining errors. Stop when the user accepts the quality, the target is met, the cap is reached, or quality stops improving.