Repository navigation
feat(6.6.0): October 2026 refresh — R53–R55, craft styles 27–29, prompt linter - #12
Conversation
…pt linter - R53 decayed signals, R54 design for machine remix, R55 generated images never pass as records (EU AI Act Art. 50) - Style atlas §1a October 2026 refresh and styles 27–29 (proof of process, cut-paper collage, contact sheet) - Platform guidance: static-to-video, Muse Image variants, remix checklist, AI Act disclosure - scripts/prompt_lint.py with self-test; CI lints the repo's own example prompts - Example 09 with three linted prompts; quick-start headline now passes R52 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Bpqe9yNt9Q9EJ5pAYnruFs
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 469b55de02
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| clause_start = max(prompt.rfind(c, 0, start) for c in ".;:\n,(") + 1 | ||
| return bool(NEGATION.search(prompt[clause_start:start])) |
There was a problem hiding this comment.
Preserve negation across comma-separated exclusion lists
For a valid constraint such as No play button, close icon or notification badge., the comma resets the negation scope, so close icon and notification badge produce a hard R48-fake-ui failure. The CLI consequently exits 1 even though the prompt explicitly prohibits those elements. Since the new workflow requires fixing every FAIL before generation, this blocks compliant prompts. Preserve negation across coordinated exclusion lists rather than treating every comma as a new clause.
Useful? React with 👍 / 👎.
| holes = sorted(set(PLACEHOLDER.findall(prompt))) + [f'"{s}"' for s in strings if SINGLE_LETTER.match(s)] | ||
| if holes: | ||
| out.append(Finding("FAIL", "R25", "unresolved placeholder: " + ", ".join(holes))) |
There was a problem hiding this comment.
Exclude prohibited filler text from placeholder findings
A complete prompt ending with No other text, no lorem ipsum, no icons. fails R25 because the placeholder search unconditionally matches lorem ipsum, even when it is explicitly forbidden. This exact exclusion is already used in the style-24 prompt fragment in references/style-atlas-2026.md, so following the repository's prompt guidance can trigger a hard failure after all actual placeholders have been filled. Distinguish a negative instruction about filler text from unresolved filler in the intended output.
Useful? React with 👍 / 👎.
| from dataclasses import asdict, dataclass | ||
| from pathlib import Path | ||
|
|
||
| QUOTED = re.compile(r'"([^"\n]+)"|“([^”\n]+)”|„([^”“"\n]+)[”“"]|«([^»\n]+)»') |
There was a problem hiding this comment.
Parse quoted copy that spans multiple lines
When a prompt specifies a headline with an actual line break inside its quotation marks, every alternative here excludes the newline, so quoted() silently discards the entire headline. For example, a 4:5 prompt whose quoted headline is Unlock followed by a newline and flaky mornings produces only the warning for absent quoted copy: it misses the hard R52 failure for Unlock and, if omitted, the R50 failure for missing no other text. The default CLI therefore accepts copy it would reject on one line. Recognize multiline quoted strings so explicit layout line breaks do not bypass the copy checks.
Useful? React with 👍 / 👎.
| ] | ||
| FAKE_UI = [ | ||
| r"play button", r"play icon", r"close (icon|button|x)", r"notification (badge|dot|bubble)", | ||
| r"unread badge", r"red badge", r"system alert", r"tappable", r"progress bar", r"mute icon", |
There was a problem hiding this comment.
Limit badge failures to functional notification UI
For a legitimate performance-sticker ad, describing a red badge carrying a verified discount triggers a hard R48-fake-ui failure solely because of its colour. For example, a complete 4:5 prompt requesting a circular red paper-price badge printed with "20% off" fails, while changing only red to yellow produces no findings. Style 8 in references/style-atlas-2026.md explicitly permits saturated sticker badges containing verified offers; these are not notification widgets. Remove the unconditional colour-based match or require notification context so compliant retail prompts are not blocked.
Useful? React with 👍 / 👎.
Why
This brings the skill up to date with October 2026: what the feed rewards now, what Meta's own systems do to a static after upload, and the EU AI Act disclosure duty that applies from 2 August 2026. It also adds a linter that catches a failed prompt before any render credits are spent.
What changed
New rules (additive, no renumbering)
Style atlas
Platform guidance §6
Tooling
scripts/prompt_lint.pyfails on placeholders, quoted copy without "no other text", copy slop (R52) and fake functional UI. It warns on long headlines, questions, em dashes, density adjectives, decayed cues, visual slop, evidence words and a missing aspect ratio. Negated mentions pass.scripts/test_prompt_lint.pyadds 20 cases. CI also lints the repo's ownexamples/prompts/*.txt.Examples and docs
examples/09-october-2026-craft-styles.mdandexamples/prompts/craft-collage-styles.txtwith three prompts that pass the linter.SKILL.md,core.md§20, both READMEs, research, QA gate, variation matrix and CONTRIBUTING.Verification
python scripts/check_docs.py→ PASS (41 files)test_qa.py16/16,test_diagnostics.py7/7,test_prompt_lint.py20/20prompt_lint.py examples/prompts/*.txt→ 16/16 pass. One expected WARN: the deliberate "before" ad in the before/after promo.Caveats
Most facts about Meta's October rollouts come from trade blogs, not Meta. The docs label them as secondhand and point to Ads Manager or the Ads Guide for verification. The AI Act passage is a production rule, not legal advice.
Generated by Claude Code