[BUG] claude behavior problem / skill-creator issue

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

The skill authoring guidance (skill-creator) doesn't distinguish between soft and hard pointers to reference files. When a skill is split into a main SKILL.md and supporting reference files, the natural pattern is to write
pointers like "see preferences-event.md for details" — but Claude treats these as advisory and will skip reading the file if it believes it already knows the answer. This produces plausible-but-wrong output: for
example, a reference file specifying a fixed anchor date of 1989-09-15 was skipped, and Claude substituted today's date instead. The error is silent — the output looks correct and the user has no way to detect
it without knowing the expected value in advance.

What Should Happen?

The skill authoring guidance (skill-creator) should recommend explicit hard-read language for any reference file that contains values Claude cannot safely infer — values where a hallucinated answer would produce a silent,
plausible-but-wrong result. The recommended pattern should be something like:

▎ "You MUST read [file] before proceeding. Do not proceed from memory."

Soft pointers ("see X for details") are appropriate for procedural reference files where skipping just means less context. Hard-read language should be the recommended default whenever the reference file
contains a specific value (an ID, a date, a format) that the skill depends on being exact.

Error Messages/Logs

Steps to Reproduce

Create a skill split into SKILL.md + at least one reference file containing a non-obvious fixed value (e.g., a specific date, an event ID, a schema format).

1. Write a soft pointer in SKILL.md: "Refer to reference-file.md for the anchor date."
2. Invoke the skill in a new Claude Code session.
3. Observe that Claude skips reading the reference file and substitutes a value from training knowledge instead of the value defined in the file.

Claude Model

Sonnet (default)

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.123

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

_No response_