[carlrannaberg]

Visualize's entrypoint repeats trigger examples and procedural explanations, while several references repeat shared rules. Condense the entrypoint in place and have creation, polish, and review defer to the existing shared contracts where their guidance overlaps.

visualize/SKILL.md shrinks from 29,797 to 25,549 bytes (14.3%); its description shrinks from 917 to 381 characters. Entry headings, commands, permissions, absolute bans, model-specific rules, universal laws, and the creation shape gate are retained. Theme alternatives remain conditional on the user asking for alternatives.

Creation starts candidate-template selection from names, descriptions, and use/exclusion sections, with further sections available when structural fit is uncertain. The selected template's complete contract is still required before composing. Report, Comparison, and Research brief now expose distinct selection criteria: prose-led findings with ordinary evidence tables; a qualitative state-and-value matrix; and research-source synthesis or study documentation. Selection-critical conditions move up from deeper guidance without duplicating their full rules. Canonical changes and all three mirrors are committed together; scripts and package/catalog architecture are unchanged.

Validation at 8e6b020:

• Mirror, authoring-contract, template inventory, and whitespace checks pass. Inventory covers 35 contracts. The required matrix generated 34 fixture previews with zero detector errors (66 warnings, 6 info); authoring retains four existing easing warnings.
• A fresh independent native subagent completed the same fictional numerical decision brief without the diff, prior findings, expected template choice, loading advice, or truncation coaching. Supplied figures, caveats, brand, and approved structure were preserved. Four browser widths showed no horizontal overflow. A bounded refinement changed one CSS declaration (1/109 lines, 0.92%); repeating the request left the artifact byte-identical.
• A second fresh agent completed a nearby valid qualitative comparison using one matrix with the supplied cell states and supporting values, without inventing a winner or scores. Four browser widths showed no horizontal overflow. Desktop and phone screenshots for both cases were inspected. Both creation detectors reported zero errors and two favicon/unfurl warnings; these remain. Contrast checks were bounded arithmetic checks, not full accessibility certification.
• The actual provider tool transcripts show a limitation: the numerical brief agent saw all three revised descriptions but still loaded all three complete candidate contracts. Its initial query retrieved use/exclusion headings without their body text. The qualitative case loaded only Comparison's full contract, although its catalog output was truncated and exposed fewer alternative descriptions. Conditional reference routing worked, with polish loaded only for the separate refinement request. Truncation/recovery reads persist. These trials do not establish lean candidate loading, whole-task efficiency improvement, or automatic skill discovery.

CI passed on run 39434297241, attempt 2 at 8e6b020. Only the failed validate job was retried; the successful mermaid-windows result was retained. The first attempt failed in the unchanged image-route test CLI waits for SIGKILL and close before timeout cleanup (image-routes.test.mjs:93), whose 5 ms timeout/grace window expected bridge-timeout but received bridge-termination-unconfirmed. It passed unmodified on retry; the same complete 59-test image suite also passed locally. No image-route source or test changes were made.

Earlier independent Opus semantic-review findings were addressed. Previous behavioral trials covered shape approval, theme alternatives, motion-none handling, and bounded/no-op refinement; the fresh tests above exercise the current template-selection changes.

The generic skill-creator validator rejects the pre-existing version, user-invocable, and argument-hint fields on both the original and revised skill. Those host fields are preserved. Raw tool-event excerpts, transcript hashes, artifact comparisons, screenshots, and CI/local logs are retained locally under temp/slimming-evaluation/template-selection-content/.