Playground pilot (user-run): document-critique over fleet SKILL.md files plus one cloud smoke

[kyle-sexton]

USER-RESERVED follow-up (deferred question Q9 in docs/topics/playground-integration/PLAN.md). The wrapper plugin (playgrounds) shipped without this evidence by design; the pilot gathers it.

Protocol (local desktop, upstream plugin as-is):

1. Install playground@claude-plugins-official, then run the repo-native recipe: "Use the playground skill to review plugins/<name>/skills/<skill>/SKILL.md and give me inline suggestions I can approve, reject or comment". Judge whether the approve/reject/comment round trip beats reading a review in the terminal. Known going in (pinned commit ed404106): the critique template's Markdown renderer misses H3s, links, and lists, and two of its three prompt-output groups are placeholders, so the question is the loop's value, not the template's polish.
2. Cloud smoke: publish one generated playground as an Artifact and press its copy-prompt button; record whether clipboard access works inside the artifact sandbox. The result may reorder the delivery tiers in plugins/playgrounds/skills/use/SKILL.md (Artifact vs file-send first).

Intake: fold learnings into the wrapper's recipes, delivery-tier ordering, and consumer notes (plugins/playgrounds/skills/use/context/consumer-notes.md); never into generation. Close this issue with the learnings recorded, even if the verdict is "no changes warranted".

[claude]

---

Generated by Claude Code

[claude]

This was generated by AI during triage.

Triage outcome: human-gated (user-reserved, and a capability blocker)

Two independent reasons this cannot be delegated, either of which is sufficient:

1. The body marks it user-reserved, as deferred question Q9 of its originating plan.
2. The protocol requires capabilities an agent lane does not have. Step 1 requires a local desktop running the interactive approve, reject and comment round trip and a human judgment about whether that loop beats reading a review in the terminal. Step 2 requires publishing an artifact and pressing a button in it to observe whether clipboard access works inside the sandbox. Neither is observable from a headless lane.

Verification (2026-09-06)

The wrapper plugin shipped and is present, including the consumer-notes surface named as the intake destination for learnings. So the pilot has somewhere to land its findings, and the item's framing that the wrapper shipped without this evidence by design is accurate.

What the runner should know before starting

The body already pins the known defects at the tested upstream commit, and they matter because they determine what the pilot is measuring. Going in, the critique template's renderer is known to miss several markdown constructs, and two of its three prompt-output groups are placeholders. The question under test is the value of the approve, reject and comment loop, not the polish of the template. A runner who judges the template rather than the loop will produce the wrong answer, so this distinction is the pilot's design and should not drift.

Step 2's outcome has a concrete consequence rather than being informational: if clipboard access does not work inside the artifact sandbox, the delivery-tier ordering in the wrapper's skill may need to change, putting file delivery ahead of artifact delivery.

Closing condition

The body sets it explicitly and it should be honored: close this with the learnings recorded even if the verdict is that no changes are warranted. A recorded negative result is the deliverable, not a reason to leave the item open. Learnings fold into the wrapper's recipes, its delivery-tier ordering and its consumer notes, and never into generation.

Not blocked on information. Nothing further is needed from the filer; this needs the reserved arbiter, a local desktop, and someone's judgment.

Cluster: playground-integration deferred verticals

This item is one of five deferred verticals from the playground-article corpus effort:
#3612 (V3), #3614 (V5), #3615 (V6), #3616 (V7), #3617 (Q9). Each measures some part of this
fleet against upstream Agent Skills or plugin-distribution guidance and records a verdict.

They are independent enough to run separately, so no carrier is designated and no dependency edges
were created. Two shared notes apply to all five:

• Provenance paths are not on the default branch. The effort's plan lives on its own branch,
  so treat those references as historical rather than as files to read.
• Recorded counts are stale. Verified 2026-09-06: the fleet is now about 262 skills, not
  the "70+" and "237" figures these bodies carry. Re-inventory before trusting any recorded count.

This triage pass created, merged and closed nothing.

---

Generated by Claude Code

[kyle-sexton]

This was generated by AI during audit remediation. Source: the 2026-09-29 audit of the Cursor agent run (PRs opened 2026-09-27 to 2026-09-29).

Closed in error. PR #4963 closed this with Closes #3617 and "Pilot not executed". Its park ledger was reverted by #5044, which left a closed issue and no artifact. I reopened it.

The pilot is user-reserved (Q9) and needs a local desktop and human judgment, so it stays on your queue. No agent lane will run it.

Steps (from the reverted checklist, git show 36c3c9bae:docs/out-of-scope/playground-pilot-user-run.md; posted here, the file is not restored). On a local desktop with playground@claude-plugins-official installed, judge the loop, not the template polish. Known at pinned upstream commit ed404106: the critique template's Markdown renderer misses H3s, links and lists, and two of three prompt-output groups are placeholders.

1. Document-critique round trip. Run: Use the playground skill to review plugins/<name>/skills/<skill>/SKILL.md and give me inline suggestions I can approve, reject or comment. Record whether approve/reject/comment beats reading a review in the terminal.
2. Cloud smoke. Publish one generated playground as an Artifact and press its copy-prompt button. Record whether clipboard access works inside the artifact sandbox. If not, consider reordering the delivery tiers in plugins/playgrounds/skills/use/SKILL.md (file-send ahead of Artifact).
3. Intake. Fold the learnings into the wrapper's recipes, the delivery-tier order and plugins/playgrounds/skills/use/context/consumer-notes.md, never into generation.
4. Close with the result recorded, even if the verdict is "no changes warranted" (this issue's own closing condition).

[kyle-sexton]

Owner decision (2026-09-29): Run both steps of the pilot (document-critique round trip, then the cloud clipboard smoke) on a local desktop and record the result, even if the verdict is no change.

Next: the owner runs the pilot; no agent lane can. The issue stays open with needs-human. After the result is recorded here, an agent folds the learnings into plugins/playgrounds/skills/use/SKILL.md (delivery-tier order), context/consumer-notes.md and the recipes, then closes the issue.