[MODEL] using heredoc consistently incorrectly

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude ignored my instructions or configuration

What You Asked Claude to Do

Not use heredoc

What Claude Actually Did

Heredoc quoting broke on that content. Writing the generator to a file instead.

Expected Behavior

Not use heredoc

Files Affected

Repository files

Permission Mode

Accept Edits was OFF (manual approval required)

Can You Reproduce This?

Sometimes (intermittent)

Steps to Reproduce

Give Claude, Opus 5 or Fable 5, a task that requires to retroactively fix prose that needs to be detailed instead of systematic. After creating the file with systematic prose, when it goes back to edit the file to change the prose to detailed and reasoned, it will waste effort to attempt to do the edit with heredoc even if instructions are given to not use heredoc. Behavior is intermittent, I suspect related to context window. Regardless, models need to be better trained on better methods to cleanly edit documents on complex projects.

Claude Model

Opus

Relevant Conversation

I need a new MD file next to LEGACY_STRINGS.md that lists each legacy code on its own line, below it I need you to write the path and filename to the related files, under each file fully explain the usage in a way I can understand, then line break a line gap, then do the next code and repeat. I will then manually go through it and next to each legacy code type a replacement leaf or word. Be very thorough. Do not use heredoc.

Impact

Low - Minor inconvenience

Claude Code Version

2.1.228

Platform

Anthropic API

Additional Context

Efficiency is important, although maybe a minor inconvenience, these add up quickly to create an environment that wastes time in situations where good codebase production is important. Claude typically writes code not to YAGNI or good practice principles (eg the public anguish over slop code) and balloons complex codebases out of control without harnessing capability that is NOT native to Claude Code. Endemic of the problem is these massive JSONL files, IMO you should be using SQLite3 to register EACH user reply in it's own row related to a conversation, and structure the data for precise retrieval. This is basic data management and Claude Code's design completely fails at doing intelligent data management for its own session data. Using a proper database per project with good structuring needs to happen like last year. This is why I have had to make my own CLI to further harness how Claude Code operates within my projects. Even with my own CLI based additional harnessing, there is so many edge cases to address and I'm only one human. You have a team of developers and models that can easily refactor the data management side of Claude Code to perform substantially better for both the customers and the models. Please stop dragging your feet and improve Claude Code.