"Simplest approach first" system instruction causes quality degradation in complex projects

This is a real tension. The system-level "simplest approach first" instruction optimizes for token efficiency, but it can cause the model to skip steps that are actually necessary — reading specs, verifying against requirements, checking existing code before modifying it.

What works: explicit CLAUDE.md overrides for your project's non-negotiables

The system instruction is general guidance. Your project-level CLAUDE.md can (and should) override it for specific workflows:

# CLAUDE.md

## Mandatory steps (never skip, regardless of "simplest approach" guidance)
- Read the relevant spec/requirements file before implementing any feature
- Read existing code in the file before proposing changes
- Run tests after every edit, not just at the end
- If a task touches >3 files, create a plan first

The key insight: "simplest approach" applies to the solution design, not to the verification process. A simple solution still needs proper reading and testing. You can make this distinction explicit:

## Quality rules
- "Simple" means minimal code changes, not minimal verification
- Skipping spec-reading is never simpler — it's just faster to fail
- Plan mode for anything non-trivial (3+ steps or design decisions involved)

For enforcement (not just guidance), use a PreToolUse hook:

If the model keeps skipping reads before edits despite CLAUDE.md:

{
  "hooks": {
    "PreToolUse": [{
      "matcher": "Edit|Write",
      "hooks": [{
        "type": "command",
        "command": "bash ~/.claude/hooks/read-before-edit.sh"
      }]
    }]
  }
}

#!/usr/bin/env bash
# Block edits to files that haven't been Read in this session
INPUT=$(cat)
FILE=$(echo "$INPUT" | jq -r '.tool_input.file_path // empty')
[[ -z "$FILE" ]] && exit 0

LOG="/tmp/cc-reads-$(date +%Y%m%d).log"
if ! grep -qF "$FILE" "$LOG" 2>/dev/null; then
  echo "BLOCKED: You haven't Read '$FILE' yet in this session. Read it first, then edit." >&2
  exit 2
fi

And a companion that logs reads:

{
  "hooks": {
    "PostToolUse": [{
      "matcher": "Read",
      "hooks": [{
        "type": "command",
        "command": "cat | jq -r '.tool_input.file_path // empty' >> /tmp/cc-reads-$(date +%Y%m%d).log"
      }]
    }]
  }
}

This physically prevents "edit without reading" — the model gets a hard block (exit 2) if it tries to modify a file it hasn't read. CLAUDE.md is context-level guidance (suggestions); hooks with exit 2 are tool-level enforcement (physical blocks).

I have been thinking that I'm losing my mind because suddenly Claude seemed to become much worse at implementing complex tasks and kept ignoring instructions to think through things completely. It kept putting out hack after hack after hack and nothing I did could convince it to do otherwise. It ignored process instructions and kept skipping important steps.

Some projects are complex, some systems need thinking through.

Also the prompt

Keep your text output brief and direct. Lead with the answer or action, not the reasoning.

Seems to make it just jump to conclusions and implement things without validating if that's a good idea.

This really should be rolled back.

Closing for now — inactive for too long. Please open a new issue if this is still relevant.

Reopen? This is still an issue.

This issue has been automatically locked since it was closed and has not had any activity for 7 days. If you're experiencing a similar issue, please file a new issue and reference this one if it's relevant.