gtm-account-scoring

Score Accounts

Recipe

1. Resolve the GTM context repo and target organization node.
2. State Using GTM context: <display name> — <N> ICPs visible.
3. Report a missing prerequisite without a band when that node has no ICPs.
4. Preserve each supplied label while validating it against the target node's own ICPs.
5. Map no-match directly to no-fit.
6. Compare account facts with the labeled ICP prose to assign one qualitative fit band without arithmetic.
7. Render compact scoring fields and a bulk band distribution when applicable.
8. Close No files, git history, or external systems changed.

Details

• Treat labels as inputs: never assign, change, repair, or reinterpret one through segmentation.
• Flag a label absent from the visible set as unknown; preserve it, assign no-fit, and mark review rather than substituting a visible ICP.
• Use exactly one band: strong-fit, good-fit, weak-fit, or no-fit.
• Use strong-fit for clear core-profile fit with all or nearly all named signals.
• Use good-fit for core-profile fit with meaningful signals, notable absent signals, and no disqualifier.
• Use weak-fit for marginal prose-supported fit or whenever a labeled ICP disqualifier applies.
• Cap any disqualified matched result at weak-fit; quote the disqualifying words and still name established fit signals.
• Judge qualification and disqualification statements wherever they appear; freeform ICP prose has no fixed schema.
• Treat a thin ICP as a scoring caveat; do not invent criteria, points, weights, or thresholds to compensate.
• Derive Confidence and Needs review only from gaps or conflicts in the account's supplied facts or label.
• Keep confidence high when supplied facts completely establish what thin prose actually says; state that the ICP limits discrimination separately.
• Render Account, Label, Band, Reasoning, Confidence, Needs review, and Open questions.
• Restrict Confidence to high, medium, or low; set Needs review to true or false.
• Open bulk output with Counts by label, Band distribution, Low-confidence count, and Review-needed count.
• Proceed with zero questions or approval gates when inputs are complete; preserve evidence gaps in Open questions.
• Never enrich, open supplied links, or make a repo, filesystem, Git-history, CRM, or external-system change.

Calls

• Read references/context.md before scoring to resolve the repo and node, enforce node-local visibility, state context, and handle ambiguity or an empty visible set. If unavailable, report that the GTM context contract could not be loaded and stop without a band.
• Read references/output.md before banding and rendering to apply label validation, qualitative band boundaries, and the exact one-off or bulk shape. If unavailable, preserve the label and render the seven literal item fields from Details plus the required close.