KKenny0/card-skill / plugins/card-skill/skills/card-skill
card-skill
UnverifiedTurn supplied text or evidence into PNG cards, article covers, and explanatory illustrations. Use for content-to-image requests, including card series and reading-note visuals.
docs/astra-migration.md
View on GitHub# Astra instruction cleanup — 0.10.2
Based on [OpenAI's skill and prompt guidance](https://developers.openai.com/blog/rethinking-skills-and-prompts-for-gpt-6-astra), this change reduces entry-point context and resolves conflicting workflow instructions. Runtime validation, the review threshold, and local publication checks are preserved. Model validation also exposed a reading-notes renderer defect: ordinary poster inline sizing overrode its 64px book-title style. The renderer now respects that cap, with a regression check.
## Two rounds
| Round | Implementation | Acceptance |
|---|---|---|
| 1 — Production instructions | Shorten the description; make SKILL.md a conditional router; use one Visual Job v3 chain for Stable and Studio; scope artifact counts by source; require candidate selection only when requested; assign PNG review to the image-capable host Agent. | Preserve content, source/privacy boundaries, full composition requirements, one revision, and reviewed local delivery. |
| 2 — Evaluation | Remove reference lists and layout coaching from fresh-context prompts; add catalog/routing probes, interruption reports, and local self-checks; synchronize documentation and the generated package. | Compare old/new routing with the same probes, run deterministic checks and smoke, inspect representative PNGs, then run bounded model production cases. |
Each round leaves a usable skill. No dependencies, renderer modes, credentials, or services were added. The new command is `npm run eval:discovery`; its scope and limitations are in [the evaluation protocol](../references/eval-protocol.md).
## Observed results
- SKILL.md: 560 → 94 lines. The description value shrank from 1,112 to 176 characters.
- `npm test` and `npm run smoke` passed. The discovery self-test and a simulated missing-CLI interruption check passed.
- Packaging and a separate byte comparison verified all 141 packaged files against root sources. `git diff --check` passed.
- Re-rendered article-formula and editorial-wechat-cover fixtures were visually inspected and byte-identical to their existing gallery PNGs.
- Baseline discovery: 6/6 passed. Changed discovery subsequently completed 6/6 with zero errors or skipped cases after the earlier usage-limit interruption.
- The interruption exposed loss of in-memory probe results. The runner now saves completed cases plus the error and reports errors/skipped cases separately.
- Production validation exposed ambiguous evidence classification, artifact indexing, and source coverage. Ordinary claims and personal experiences are distinguished from explicit quotations; indices restart within each output (also documented at the public schema field); usable current evidence is tied to a visible artifact. Exact-quote checks, source coverage checks, and revision identity checks remain enforced.
- Long-mode guidance now matches its structured renderer: a short faithful title, complete body text, and no unsupported graphical metaphor. Formula reviews use the mode's actual visible fields.
- The too-text-heavy revision fixture now seeds five verbatim paragraphs, a mechanically checked bounded redundancy defect. Earlier unbounded seeds could overflow before visual revision; this fixture change limits comparisons with older runs.
The shorter entry point is measured; improved accuracy, total token cost, or speed is not established. The probes also do not prove end-to-end production behavior or performance among a large competing skill catalog.
## Release verification
On 2026-09-14, the final complete CardBench run passed all 29 cases with all output checkers passing. The committed [report](../evals/cardbench-results.json) has `scope.kind: full`, `complete: true`, and `prompt_profile: skill-routed-v1`; its measured visual score is 9.0/10. This includes 25 planning cases and four forced-revision cases. Earlier interrupted or failed runs were development feedback, not passing release evidence.
L2 Agent Critic is complete. The report leaves maintainer judgment and real-user reuse evidence false, and unmeasured content/agent/overall scores null. Representative PNGs were additionally inspected in the maintenance session, including split diagrams, long text, comic, qualified formula, and the reading-notes title fix. No claim of higher model accuracy or faster execution is made.
Reverting the changed sources and regenerating the mirror rolls back this change without a data migration.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis. | Stars: 472k | Updated | Claude CodeCodex | Verified | Install |
| diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness. | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependencies | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is available | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |