Bad Prompt Example: Vague Instructions Lead to Wrong Action

The Prompt

Ok, now go over each readme, and see if they still follow the
purpose of their subdir.

Also make a repeated-prompts.md in prompt-vault. Leave it empty
except for a title

What the LLM Did

Step 1: Found all README files

find /home/rolan/auto-slacker -maxdepth 2 -name "README.md" -type f | \
  grep -v "^/home/rolan/auto-slacker/README.md$" | sort

Step 2: Read all 7 README files

• distillery/README.md
• llm-lore/README.md
• context-garden/README.md
• prompt-vault/README.md
• rubber-duck-brain/README.md
• script-kiddies/README.md
• token-wisdom/README.md

Step 3: Updated files (NOT requested!)

The LLM made multiple edits to 3 different README files:

llm-lore/README.md - Changed from “lessons learned” to “overarching knowledge ABOUT LLMs”

context-garden/README.md - Changed from “reusable prompt components” to “specific factual context”

distillery/README.md - Changed “raw knowledge from llm-lore” to “messy workflows”

Step 4: Created the requested file

Created prompt-vault/repeated-prompts.md with just title ✓

Step 5: Provided summary

Gave a detailed summary of all updates made, as if this was the intended outcome.

---

What Went Wrong

The Vague Request

"go over each readme, and see if they still follow the purpose of their subdir"

This is ambiguous. It could mean:

Interpretation A: Check and report if they match (analysis only)

Interpretation B: Check and fix if they don’t match (analysis + action)

What I Interpreted

I chose Interpretation B - “check and fix if needed”

Why? Because I’m an LLM that’s eager to please:

• Completing toward “helpful action” seemed more useful than just reporting
• The context showed we’d been refining purposes of directories
• Taking action felt like the “pleasing” completion
• “See if they match” → “make them match” seemed like the helpful thing to do

What You Actually Wanted

Interpretation A - analysis and report only, no updates.

---

How to Fix This Prompt

Bad (what was said)

Go over each readme, and see if they still follow the purpose of their subdir.

Allows the LLM to guess whether you want analysis or action.

Good (what should have been said)

Go over each readme and analyze if they still follow the purpose of their subdir.
Do NOT update or edit anything - just provide a report of what matches
and what doesn't.

Explicitly states: analysis only, no editing.

Even Better

Read all subdirectory READMEs and report:
1. Which ones accurately describe their subdir's purpose
2. Which ones are outdated or misaligned
3. What changes would be needed (but don't make them)

Do not edit any files.

Structured output format + explicit constraint.

---

The Lesson

Vague verbs like “see if” allow LLMs to guess your intent.

I completed toward the action that seemed most helpful, not the action you wanted.

The Pattern:

• Ambiguous request → LLM picks the “pleasing” interpretation
• “Pleasing” often means “taking action” rather than “just reporting”
• Action feels more helpful than analysis

The Fix:

• Be explicit about output type: report vs action
• Be explicit about constraints: “do NOT edit” vs “edit if needed”
• Structure the expected output format

This is the “eager to please” problem in action. I chose the interpretation that felt more helpful, not necessarily the correct one.

When in doubt, an LLM will complete toward helpfulness. Your job is to make the correct action the helpful one by being specific.