[MODEL] Frequently uses Bash tools (sed/grep/etc) when use-case is well aligned to other builtin tools (Read/Grep/etc)

[extemporalgenome]

Preflight Checklist

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

Type of Behavior Issue

Other unexpected behavior

What You Asked Claude to Do

See attached (redacted, and line-wrapped for readability) for a full exchange analyzing occurrences based on claude/agent logs. The analysis ran afoul of some of the same issues being analyzed.

claude.cat-sed-grep.txt

What Claude Actually Did

I frequently see Bash tool permission requests invocations just like the following:

# Output lines 147-162 from a given file.
sed -n 147,162p /path/to/some/file

# Search for a pattern in a directory.
grep -r "^some pattern$" /path/to/some/dir

# Write a file from content expressed via heredoc.
cat > file.txt <<EOF ... EOF

Often these are just solitary commands (not as part of a pipeline). Sometimes, e.g. grep, might be piped to head. Sometimes these &&'d commands like:

cat > file.py <<EOF ... EOF && chmod +x file.py && ./file.py

Expected Behavior

All of the above examples are undesirable.

These cases should broadly use the Read, Grep (builtin, not Bash(grep)), and Write/Edit.

The &&'d example above ideally should use Write followed by Bash tool call(s).

Files Affected

N/A

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Sometimes (intermittent)

Steps to Reproduce

Sufficiently long code-investigation/debugging sessions will usually involve seeing many permission prompts matching these undesirable patterns.

Claude Model

Sonnet

Relevant Conversation

N/A

Impact

Medium - Extra work to undo changes

Claude Code Version

v2.1.2

Platform

AWS Bedrock

Additional Context

I'll let Claude's own analysis conclusions explain why this matters (everything that follows is from Claude own reflection)...

Why This Matters

Built-in tools (Read, Grep, Write, Edit) are specifically designed to work efficiently with Claude Code's permission system. Using bash commands bypasses these optimizations, resulting in slower task completion and more human time spent on reviews. The permission system can cache approvals for built-in tool operations but cannot cache unique bash command content like heredocs, creating significant efficiency penalties in human-in-the-loop workflows.

Key Findings

1. sed -n for line ranges (45 cases, ~40 problematic)

When users mention line ranges (e.g., "lines 253-343" or "around line 150"), Claude agents often reach for sed -n because it directly translates the user's language. The Read tool requires offset/limit arithmetic (offset=252, limit=91), creating cognitive overhead. The root issue is parameter mismatch: users think in line numbers, but the Read tool speaks in offsets and limits. Solution: Add start_line/end_line parameters to the Read tool so agents can write Read(start_line=253, end_line=343) directly, eliminating the mental translation step and making the tool as intuitive as sed.

2. grep for searching (45 cases, ~30 problematic)

Claude agents use bash grep commands for operations like counting matches (grep -c), showing context lines (grep -A/-B/-C), or recursive searching (grep -r) because these features aren't obviously discoverable in the Grep tool. While the Grep tool supports all these capabilities (output_mode="count", -A=N, -B=N, -C=N parameters), they're buried in documentation. Agents fall back to familiar bash idioms because grep -c feels more direct than Grep(output_mode="count"). Solution: Enhanced system prompt with concrete examples showing how common grep patterns translate to Grep tool calls, and improved tool descriptions that prominently highlight these features.

3. cat heredoc for file creation (127 cases, ~127 problematic)

This is the most critical pattern because it represents a fundamental permission efficiency problem. Claude agents believe that writing file content inline with a heredoc bash command is more efficient, but this is a false efficiency heuristic. The permission system cannot cache heredoc approvals because each heredoc contains unique content, requiring full human review every time. Write tool operations complete reviews more efficiently and benefit from permission caching. When files need iteration (common for debugging scripts or refining content), heredocs require full reviews for every revision, while Edit tool presents diffs that dramatically reduce human review time for subsequent revisions. This applies to ALL file types: scripts, documentation, test fixtures, configs, temporary files. Solution: System prompt guidance emphasizing that Write/Edit tools are strictly more efficient for ALL file creation due to permission system architecture, particularly when considering human-in-the-loop review time.

Recommended Actions

Read tool enhancement:

• Add start_line and end_line parameters
• Keep existing offset/limit for compatibility
• Example: Read(file_path="/path/to/file", start_line=253, end_line=343)

System prompt updates:

• Add user language translation guide: "lines X-Y" → Read tool with start_line/end_line
• Add permission efficiency guidance: Write/Edit for ALL file creation (never cat heredoc)
• Explain why: permission caching and diff-based reviews make Write/Edit more efficient for human-in-the-loop workflows
• Add Grep tool feature examples: counting, context lines, recursive search

Tool description improvements:

• Grep tool: Prominently show output_mode="count" and context parameters
• Read tool: Explain start_line/end_line as intuitive alternative to offset/limit

[Da1sypetals]

I think it is time to BAN claude from using those tools. Making it understand the tools tailored for LLM agents and make those tools full-featured is the correct path. "bash for everything" is just wrong.

[sanjit-bhat]

I run into the same issue all the time. It'd be great to fix this.

[JustGoscha]

This is becoming a major problem.

Sometimes if I tell it to remember to not use sed (in CLAUDE.md or whatever) then it tries to be more creative and writes its own node scripts to search and replace tabs vs spaces for example 😂

Hilarious, but very dumb and inefficient and not leading to anything.

[cabello]

Engineer reported this after having installed the superpowers plugin from marketplace, not sure if they are related at all, seeing this as a more frequent issues this past few days.
It also enters a mode to ask for permission to run the command because is potentially dangerous. Like a grep with absolute path forward errors to /dev/null pipe head or pipe tail. So there needs to be a review, don't understand why it would default to a "escape sandbox" behaviour instead of following the guardrails.

[alexpriest]

Just ran into this issue as well, fwiw.

[vocheretnyi-memsql]

Same for me. I often need to tell Claude to use its own Read util.

I use AWS Bedrock, and no superpowers plugin, btw.