Conciseness

A telegraphic, table-driven runbook with zero explanation of concepts Claude already knows — e.g. "cutoffs → population → feature CTEs → label → final SELECT" and "Never ask for `td.apikey` — AI Signals provides it". Every line is instruction, routing, or a decision rule; no padding, no time-sensitive version details, and the phase summary is compact navigation rather than duplication.

Actionability

Provides copy-paste commands ("tdx tables \"<database>.*\"", the information_schema.columns classification SQL), exact config keys ("sink_database", "model_name", "target_column", "n_train_snapshots", "spacing_days"), exact file paths ("workflow/sql/input_features.sql", "config.yml"), naming rules, and quantified checks ("AUC above ~0.98, one dominant feature → drop the feature, re-run"). Remaining placeholders like <project_slug> are inherently per-project variables, not pseudocode, so this sits at the fully-executable anchor.

Workflow Clarity

Five phases with numbered steps, explicit user checkpoints ("Never skip", "ask once", "wait for approval", Phase 5 "only when the user says the model is ready"), and feedback loops around the batch push/run operations ("data_prep fails → fix the SQL, re-run", "Leakage… → drop the feature, re-run", "compare with the previous run in project_state.md"). Validation is present at every risky stage, so the batch-operation cap does not apply and the content matches the anchor with explicit validation, recovery loops, and confirm-checklists.

Progressive Disclosure

The "How to Start" routing table maps each question to one of eight real references/ files, all one level deep, and every cited section verifies against the bundle (workflow-setup.md §1/§2/§4, dataprep-guide.md §2 "Choosing the snapshots" and §4/§5/§6, agent-setup.md "Build checks"/"Finalize"). Bulk detail (SQL patterns, API parameters, output schemas) lives in the reference files while SKILL.md stays at orchestration level; the ../shared/*.md and project_templates/ pointers are package-level cross-skill files rather than nested bundle references, so navigation remains one level deep and clearly signaled.

Specificity

Lists multiple concrete actions spanning the full lifecycle — "Builds leakage-safe stacked training snapshots, trains XGBoost/CatBoost with tuning, threshold selection and SHAP, writes Parent Segment scores, and builds the project agent and dashboard" — with named technologies (XGBoost/CatBoost, SHAP, Parent Segment) and no vague filler. Coverage runs from data prep through deployment, matching the comprehensive anchor rather than the minor-gaps anchor below it.

Completeness

Explicitly answers both 'what' ("Sets up AI Signals binary classification (propensity scoring) on Treasure Data… writes Parent Segment scores, and builds the project agent and dashboard") and 'when' ("Use to start, configure, change, re-run or resume a project or rebuild its agent. Trigger on: …") with concrete trigger phrases — the exact anchor-5 pattern, not merely the implicit-'when' anchor 4.

Trigger Term Quality

Over twenty natural trigger phrases with synonym coverage across use cases — "propensity model", "predict which customers will churn", "score leads by likelihood to convert", "build a churn model", "build a lead scoring model", "classifier runbook" — plus operational variants ("resume classifier setup", "rebuild the classifier agent", "configure the classifier"). This exceeds the anchor-5 example's synonym breadth; absent literal phrases like "upsell model" are covered by the use-case list, so it does not fall to the 'a few natural terms missing' anchor.

Distinctiveness Conflict Risk

The niche is tightly pinned to "AI Signals binary classification (propensity scoring) on Treasure Data" and nearly all triggers are domain-qualified ("ai signals propensity model", "classifier output tables", "training snapshots"). A few broad phrases ("build a churn model", "classification workflow") carry only minimal overlap risk since they still resolve to this skill's niche, fitting anchor 5 rather than the minor-overlap anchor 4.