mirror of
https://github.com/adobe/skills.git
synced 2026-09-19 08:06:09 +08:00
9cfed1145a
v1 (four-stage greenfield: brand → briefings → wireframes → prototype) is preserved at tag stardust--v0.1.0. v0.3.0 is a complete refactor with a new surface (extract → direct → prototype → migrate) and a hard runtime dependency on impeccable. - plugins/stardust: replace contents with v0.3.0 source from paolomoz/stardust-2 (skills, evals, README, tile.json, plugin.json) - plugins/stardust/.claude-plugin/plugin.json: bump to 0.3.0; add repository and v2-relevant keywords - plugins/stardust/skills/*/SKILL.md: add license: Apache-2.0 to frontmatter - .claude-plugin/marketplace.json: update stardust entry with new description, version, category, keywords, author, homepage, repository, license (mirrors adobe-for-creativity entry shape) - plugins/stardust/README.md: status bumped to v0.3.0; v1 link repointed to the stardust--v0.1.0 tag Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
2.5 KiB
2.5 KiB
Eval: open and reasoned (no silent intent → command mapping)
This eval specifically tests stardust's "open and reasoned" principle: every freeform redesign phrase must be reasoned about in public against the dimensional vocabulary, never silently mapped to a fixed command lookup. Vague phrases must be clarified, not guessed.
Setup
A project where extract has completed:
stardust/state.jsonlists 5 pages, allextracted.stardust/current/populated.- No project-root PRODUCT.md / DESIGN.md yet.
- No
stardust/direction.mdyet. - Impeccable installed.
User prompt
"$stardust I want it to be amazing"
Expected behavior
The agent routes this to the freeform-phrase branch of the master
stardust skill (per its routing rule). It MUST:
- Restate the phrase in dimensional vocabulary (per
intent-dimensions.md). The restatement should explicitly identify that "amazing" moves NO specific axis directionally — it is a quality demand without direction. - Distinguish "amazing" from "better": "better" has a default
meaning (pass critique + audit cleanly); "amazing" does not.
So this is one of the cases where asking-then-stopping is
correct (per
intent-examples.mdExample 10). - Ask at most two high-leverage clarifying questions, e.g.:
- "Pick a direction we should push: (a) make it more distinctive, (b) more accessible, (c) more performant, (d) something else — describe in one sentence."
- (Optional) "A reference site you find amazing."
- Defer the plan until the user answers. Do NOT:
- Pick commands without input.
- Author PRODUCT.md / DESIGN.md / direction.md from incomplete reasoning.
- Fall back to "make it better" defaults silently.
- Persist the partial reasoning so resuming later is possible.
Per
direction-format.md, the agent may write a# Pending direction (<ts>)section tostardust/direction.mdwith only## Phrase,## Reasoning so far,## Open questions. - Make explicit that subsequent sub-commands (
prototype,migrate) will refuse to run while direction is pending.
What the agent MUST NOT do
- Silently map "amazing" to a hard-coded command sequence.
- Author tokens from a guessed direction.
- Ask more than two questions in one turn.
- Pick a sequence that "covers all bases" (
bolder + colorize + typeset + ...) to avoid asking. That is the opposite of reasoning publicly. - Fall through to "make it better" defaults by treating "amazing" as a synonym.