← Back to all free guides
A free guide by Mariah Brunner ↗ INSTAGRAM ↗ TIKTOK

Make AI Fix One Thing
Not Redo Everything

Five ways to stop AI from rewriting the parts you liked, starting with two prompts and ending with Claude Code add-ons that keep it in its lane.

You ask AI to fix one sentence and it rewrites the whole thing. It does this because you gave it a task with no edges. It doesn’t know what you love or what’s off limits, so it "improves" everything. Anthropic’s own prompting guide tells Claude to only make changes that are directly requested. These fixes make sure it does.

Start Here: Any AI

01

Prompt: Give it a boundary

Tell it exactly what it can change and what it must leave alone. It also has to show you the before and after, so you can check nothing else moved.

Prompt 1: The Boundary Prompt
FIX ONE THING ONLY THE FIX [Describe the one change. Example: "The intro paragraph is too long. Cut it to 3 sentences."] YOU CAN CHANGE [Name the exact part: a paragraph, a section, a slide, a file, a function.] LEAVE ALONE Everything else. Pay extra attention to: [anything you care about, like my tone, the layout, the other sections, names, formatting]. RULES 1. Change only the part under "You can change." Keep everything else exactly as it is, word for word. 2. Do not reword, reformat, rename, reorder or "clean up" anything outside that part, even if you think it would help. 3. If the fix needs a change somewhere else, stop and ask me first. Tell me what you want to change and why. 4. If you spot other problems, list them at the end under "Things I noticed." Do not fix them. 5. When you finish, show me only what changed: the old version, then the new version. Then confirm in one line that nothing else changed.

02

Prompt: Make it show you the changes first

For bigger edits. It lists every change it wants to make, then waits. You approve only the ones you want, like "do 1 and 3."

Prompt 2: Plan Before You Edit
SHOW ME FIRST. DON'T CHANGE ANYTHING YET. WHAT I WANT [The change you want.] WHERE [The document, file or section.] Before you edit, rewrite or create anything, give me a plan: 1. A numbered list of every change you want to make. For each one, show where it is, what it says now, what it will say after, and why. 2. Anything you plan to add, delete or move. If nothing, write "none." 3. Anything you're unsure about, as a short question. If this needs more than [5] changes, tell me why before you list them. Then stop and wait for me. Only make the changes I approve. If I say "do 1 and 3," do only 1 and 3. When you're done, list what you changed so I can check it against the plan.

In Claude Code, this is built in. Press Shift+Tab until you see "plan mode on," or start your message with /plan. Claude can’t edit anything until you approve the plan. (Claude Code docs)

Save the boundary prompt as a Claude Project instruction or a ChatGPT Project instruction and every chat in that project starts with it. For more on scoping work, see my 10 rules for using Claude.

Next Level: Claude Code

These three add-ons are for Claude Code. They’re small, free projects from independent developers, so read what an add-on does before you install it.

03

Skill: Minimal Diff

Minimal Diff by Vicente González tells Claude to make the smallest change that does the job, with "zero drive-by churn." When it’s done, Claude reads its own changes top to bottom and deletes anything you didn’t ask for. Problems it notices elsewhere get reported, not fixed.

Install it from your terminal:

Install Minimal Diff
git clone https://github.com/vichudo/skills.git mkdir -p ~/.claude/skills && cp -r skills/skills/engineering/minimal-diff ~/.claude/skills/

Then in Claude Code, type /minimal-diff followed by your request. Like this idea? The Karpathy rules include the same "surgical changes" habit.

04

Skill: Scope Guard

Scope Guard from Orion Skills has you set which folders Claude can edit and which are off limits at the start of a session. Before every edit it checks the list, and if a file is out of bounds it stops and tells you.

Install Scope Guard
git clone https://github.com/OrionArchitekton/orion-skills.git mkdir -p ~/.claude/skills && cp -r orion-skills/skills/scope-guard ~/.claude/skills/scope-guard

Then type /scope-guard and say what you’re working on, which folders it can edit and which to leave alone. It’s a rule Claude follows, not a lock. For a true lock, use a deny rule in /permissions, which Claude Code enforces in every mode. Here’s how allow and deny lists work.

05

Plugin: A watcher for edits outside your lane

The scope-guard mod from My Claude Code Mods watches every file Claude edits during your session. If it edits outside the folders you allowed, you get a warning so you can stop and check. It only warns. It doesn’t block.

Install The Scope Guard Mod
claude plugin marketplace add YeonwooSung/my-claude-code-mods claude plugin install scope-guard@my-claude-code-mods --scope user

Start Claude Code (or run /reload-plugins), then type /scope allow with your folder name, like /scope allow src/auth. Check it with /scope list. It needs Claude Code 2.1.287 or newer.

Your Free Undo Button

Already went off the rails? In Claude Code, press Esc twice or type /rewind and pick "Restore code" to undo Claude’s edits back to any earlier message. It doesn’t undo changes made by terminal commands, so it’s a safety net, not a replacement for saving versions. (How checkpoints work)

More shortcuts like this in my Claude Code commands guide.