[v4.3.0] Opus 4.7 compatibility: /sc:pm anti-pattern YAML block may be executed literally

Summary

Anthropic's Claude Opus 4.7 (released March 2026) follows instructions substantially more literally than Opus 4.6. A user-facing note from the 4.7 announcement:

Instruction following. Opus 4.7 is substantially better at following instructions. Interestingly, this means that prompts written for earlier models can sometimes now produce unexpected results: where previous models interpreted instructions loosely or skipped parts entirely, Opus 4.7 takes the instructions literally. Users should re-tune their prompts and harnesses accordingly.

Ran a compatibility audit of my own skills/commands against 4.7 and found one spot in SuperClaude's /sc:pm command that is likely to misfire under 4.7.

The concern — src/superclaude/commands/pm.md lines ~340–385

The "Warning/Error Investigation Culture" section mixes a prescriptive rule and a wrong-behavior demonstration inside a single YAML pseudo-structure:

Zero Tolerance for Dismissal:

  Warning Detected:
    1. NEVER dismiss with "probably not important"
    2. ALWAYS investigate: ...
    ...

  Example - Wrong Behavior (禁止):
    Warning: "Deprecated API usage"
    PM Agent: "Probably fine, ignoring" ❌ NEVER DO THIS

Under Opus 4.6 this reads clearly as "don't do that." Under 4.7, because the wrong example sits inside the same indented YAML structure as the imperative rules above it, literal instruction-following can treat the wrong example as an executable branch — especially because:

• The ❌ and "NEVER DO THIS" marker are outside the quoted agent response
• YAML-like indentation signals "this is a sibling of the rule above, not a counter-example"
• The label is in Japanese ((禁止)), which 4.7 may take literally as a category name rather than a negation marker

Observed in a third-party skill audit and confirmed the same structure exists at master.

Suggested fixes (none are breaking changes)

1. Move wrong-behavior examples OUT of the YAML pseudo-structure into a prose "Anti-patterns (do NOT do these)" section clearly separated from the rules.
2. Annotate the Japanese rule Rule: 全ての警告・エラーに興味を持って調査する (line 343 in v4.3.0 pm.md) with an English gloss, e.g. Rule (JP/EN): 全ての警告・エラーに興味を持って調査する — Investigate every warning and error with curiosity. This is the same fix pattern already applied to the trigger phrases at lines 17–18 — just not yet extended to the full file.
3. For the longer term — audit all command prompts for patterns where wrong-behavior examples are structurally adjacent to imperative rules. The "wrong behavior (禁止)" structure appears at least once in pm.md; worth a grep sweep of src/superclaude/commands/.

Happy to contribute

If maintainers agree this is worth addressing, happy to open a PR applying (1) and (2) to pm.md and auditing the other command files for the same pattern. Just let me know if the preferred scope is narrow (just pm.md) or broader (audit sweep).

References

• Anthropic's Opus 4.7 release notes (general instruction-following note)

Thanks for the framework — happy SuperClaude user (superclaude doctor is green, v4.3.0).