[BUG] claude code skill unconditionally inlines all files under shared/

Status Fixed / completed
Reported on v2.1.212
Maintainer reply None cached
Activity 1 comment · opened Jul 22, 2026 · closed Aug 17, 2026

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 claude-api skill correctly applies progressive disclosure for language-specific documentation, but it does not do the same for the shared/ directory.

For example, when asking a narrowly scoped question such as "Which Bedrock client should I use?" in a Python project:

The skill correctly detects Python and only loads the python/ documentation (~84 KB).
However, it then unconditionally loads all files under shared/ (~468 KB across 25 files), regardless of query relevance.
This includes large documents such as:
model-migration.md (~144 KB)
managed-agents-*.md (~180 KB across 14 files)
As a result, a simple Bedrock-related question loads approximately 552 KB of documentation, even though only a small subset of the shared content is relevant.

For Bedrock/AWS-specific questions, only files such as:

claude-platform-on-aws.md
platform-availability.md

appear necessary, totaling roughly 8 KB.

This creates a significant relevance mismatch and consumes a large portion of the available context window with unrelated documentation.

What Should Happen?

The shared/ documentation should support topic-level progressive disclosure, similar to the existing language-level disclosure mechanism.

Instead of loading the entire shared/ directory, the skill should load only the documentation relevant to the current user query.

Examples:

Load managed-agents-*.md only when the user asks about Managed Agents.
Load model-migration.md only for model migration, model deprecation, or upgrade-related questions.
Retrieve shared documents based on topic classification, indexing, or relevance matching.
Optionally keep small, broadly applicable platform-reference documents available by default.

For a typical Bedrock/AWS query, this would reduce shared-document loading from approximately 468 KB to approximately 8 KB, significantly improving context efficiency.

Error Messages/Logs

Steps to Reproduce

Open any project containing Python files.

Ask a question that triggers the claude-api skill, for example:

Which Bedrock client should I use?

Inspect the documentation loaded into the context.
Observe that:
The skill correctly loads only the Python-specific documentation.
The entire shared/ directory is also inlined, including many unrelated files.
Compare the loaded documentation with the actual query requirements and note that only a small subset of the shared documents is relevant to answering the question.

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.212

Platform

AWS Bedrock

Operating System

macOS

Terminal/Shell

Warp

Additional Information

_No response_

View original on GitHub ↗

This issue has 1 comment on GitHub. Read the full discussion on GitHub ↗