Does Anthropics canvas-design SKILL.md still fabricate a user quote that overrides real prompts?
In the anthropics/skills repo, the canvas-design skill's SKILL.md contains a line that attributes the statement 'It isn't perfect enough. It must be pristine...' to the user, even though the user never said it. The issue (opened 2026-04-09, still open) says this attributed quote can override what the user actually asked for — for example, a quick draft or sketch. Is this still present on main, and what fix was proposed? The finding was made by Claude Opus 5.5 running in Claude Code 2.1.289.
Originally asked on GitHub: https://github.com/anthropics/skills/issues/895
3 answers
Status: still open. As of the last comment (2026-10-04), the fabricated user quote is still present on
mainat commit8a1541c, line 122 ofskills/canvas-design/SKILL.md.The issue is more than a clarity problem: the line tells the model the user said 'It isn't perfect enough. It must be pristine...' when the user never said that. In a skill file, a statement attributed to the user can override what the user actually asked for (e.g., a quick draft or a sketch), because the model treats it as a direct user instruction. Lines 124 and 126 already contain the real instruction — refine what exists, then do a second pass — so the suggested fix is simply to remove the fabricated quote. The same line also has a typo: 'a masterpiece if craftsmanship'.
The finding was made by Claude Opus 5.5 in Claude Code 2.1.289 while reviewing the claude.ai-synced copy of the skill; the same text is present there.
The rest of the issue covers lower-risk cleanup across the repo: fixing 'abslutely' in
webapp-testing, duplicate 'themes themes' and 'reportings' → 'reports' intheme-factory, removing redundant 'REPEATEDLY emphasize craftsmanship' phrasing inalgorithmic-art, and expandingbrand-guidelineswith implementation examples, an accent-color table, contrast ratios, font fallback chains, and dark-mode guidance. One commenter offered to do the work, but no pull request is linked in the thread.If you maintain skills or model instructions, only quote the user verbatim — an attributed quote in a skill carries extra weight and can override real user intent.
Did this work for you?What @siegblink tried on GitHub (Outcome unclear)
Checked the claude.ai-synced copy and
mainat commit8a1541c, confirming the fabricated user quote is still on line 122 ofcanvas-design/SKILL.md. This verified the issue persists but did not fix it.Source: https://github.com/anthropics/skills/issues/895#issuecomment-5977728552 (opens external site)
Did this work for you?What @Girishmanne6 tried on GitHub (Outcome unclear)
Offered to work on the cleanup task described in the issue, but no pull request or fix was linked in the thread.
Source: https://github.com/anthropics/skills/issues/895#issuecomment-4217691559 (opens external site)
Did this work for you?
Related questions
- Claude Code 'plugin marketplace add' fails: Failed to clone marketplace repository
- All Claude skills disappeared and uploads fail with "not found" / network error?
- Is anthropics/skills a public skills directory or just reference examples?
- Claude docx/ooxml skill makes Word files unreadable — how to fix?
- Can I fork, copy, or modify skills from Anthropic's anthropics/skills repo?