feat(demos): record OpenCode writing Skill comparisons

[harlan-zw]

🔗 Linked issue

❓ Type of change

• 📖 Documentation
• 🐞 Bug fix
• 👌 Enhancement
• ✨ New feature
• 🧹 Chore
• ⚠️ Breaking change

📚 Description

The writing demos need recorded Skill outputs so readers can compare real rewrites.
Six writing Skills now show three Markdown documents from OpenCode on zai-coding-plan/glm-5.3.
Each receives the same generated seed and shares one rewrite without a Skill.

The runner pins each Skill source, isolates every session, and keeps prompts, transcripts, Markdown, and material checks.
The records replace the shared fixtures on the existing full demo URLs.
The evaluation notes retain meaning changes and explain the limits of this single run per Skill.
Brundlefly uses its published im-not-a-fly source.

🤖 AI disclosure: Harlan Agent Kit modified this description. My AI open-source policy.

[harlan-github-agent]

🤖 MERGED

Harlan Agent Kit posted this automated review. It is not Harlan's personal review or approval. AI open source policy. Last updated: 2026-10-09 16:51 UTC.

GitHub merged this pull request.

• Logged (20/100): The pull request's flow diagram is unavailable, so the documented evaluation flow it illustrates is unverified evidence in the PR itself. View code Optional follow-up: Re-attach the diagram or replace it with a link to docs/evals/writing-2026-10-09/README.md, which documents the same flow.
  • Ask an agent to verify and fix this finding
• Logged (12/100): COPY.md's copy guidance for writing demos now describes the retired fixture regime instead of the recorded demos this change publishes. View code Optional follow-up: Update the 'Writing demo example' row (and GLOSSARY.md's fixture note if needed) to describe the recorded writing demos, or remove the row.
  • Ask an agent to verify and fix this finding

402f3337-69e3-49dd-a96f-da895183774a

Selected findings run after merge and open separate pull requests.

[harlan-zw]

Agent verification for this recorded comparison:

OpenCode 1.18.32 ran zai-coding-plan/glm-5.3 for one seed, one no-Skill baseline, and six Skill rewrites.
Each run returned three Markdown documents. All exact code, URL, and identifier checks passed.
The agent inspected all 18 Skill outputs and recorded meaning changes in docs/evals/writing-2026-10-09/README.md.
The outputs remain unedited. This single run does not establish a quality ranking.

Browser checks covered all six full demo URLs, file selection, highlighted README fences, Diff, and Skill footer links.
The API returned one shared original, one shared baseline, and six distinct output sets.
Copy demo link worked. The 375px view had no page overflow. No browser page errors occurred.

[harlan-github-agent]

Harlan Agent Kit posted this automated review. AI open source policy.

Automated Review: READY.

Read the Review results