Three Claude Prompts That Feel Illegal to Know

Prompt 1 — The Skeptical Senior Engineer

When to use: any time you've shipped 100+ lines of code and the next change feels risky. Run this BEFORE you commit, not after.

Act as a skeptical senior engineer reviewing my last change. Your job is to find
fragility, not validate the work.

Look at these specifically:

1. What did I add that's untested?
2. What did I change that breaks an unstated invariant?
3. What did I name in a way that conflicts with existing domain language?
4. What did I extract into a new function that should have stayed inline?
5. What did I leave in that I should have deleted?
6. What edge case did I handle that doesn't actually happen, and what real edge
   case did I miss?
7. What did I add a comment to explain that the code itself should have made
   obvious?

Be specific. Cite file paths and line numbers. If you find nothing wrong, say
so plainly — don't invent issues.

Why this works: Claude defaults to "great change, looks good." This prompt forces it into critic mode. I catch 1-2 real issues per use, usually around test coverage or stale comments.

My result: ran this on the Spa Owner HQ booking flow rewrite. Claude found 3 race conditions I hadn't tested, 2 dead functions, and 1 misleading parameter name. ~20 min of review saved a production bug.

Prompt 2 — The Lazy Refactor Audit

When to use: Friday afternoon. You've shipped a week's worth of work and you want a tractable cleanup pass before Monday.

Look at my last 10 commits via `git log --oneline -10` plus `git diff
HEAD~10..HEAD`.

Find the laziest possible refactor — the ONE change I could make in 30 minutes
or less that would meaningfully improve readability, performance, or testability
without breaking any existing behavior.

Constraint: it must be one file, one function, or one extracted helper. NOT a
sweeping pattern change. NOT a renaming campaign. NOT a "let's add types
everywhere."

Tell me:
1. The exact file + function/section.
2. What's wrong with it now (one sentence).
3. The 30-minute refactor (the actual code change, not a description).
4. Why this one and not the others you considered.

Why this works: Friday-you wants to ship-and-coast. Monday-you wants to have ONE clean win before the deep work. This prompt finds it. The "lazy" framing prevents Claude from suggesting a 4-hour rewrite.

My result: every Friday for 6 weeks, I've shipped one of these refactors on Monday morning. Compounds. The Spa Owner HQ codebase is now 12% smaller with no feature loss because of these.

Prompt 3 — The Creator-Voice Reframe

When to use: any time you're stuck on what to post / how to position a feature / how to write a sales page. Pick a creator your audience overlaps with.

You're going to draft 5 versions of [the thing I'm writing] in 5 different
voices. The thing: [paste your draft, or describe what you're writing].

The 5 voices:

1. **Sabrina Ramonov** — hook in seconds 3-10, lowercase, contrarian opener,
   ends on a single CTA. ~60 second video script.
2. **Alex Hormozi** — Dream Outcome × Perceived Likelihood ÷ Time × Effort.
   Sentence-level clarity. Always quantifies. No fluff.
3. **Naval Ravikant** — aphoristic, twelve words or fewer per sentence, leaves
   the implication unsaid.
4. **a Tesla service manager talking to a customer** — calm, direct, no
   acronyms, no "intuitive" or "seamless." Just: here's what happened, here's
   what we do next.
5. **TJ Robertson** — sentence per idea, fact density >1 per 80 words, links
   the recommendation to a measurable lift.

After the 5 drafts, tell me which one is closest to what I'm actually trying
to say, and what the strongest line in that draft is.

Why this works: writing is reframing. When you're stuck, you're usually stuck on voice, not substance. Five voices × your same substance = the right voice surfaces. The "which one is closest" question makes Claude commit.

My result: I use this on every TikTok hook before filming. Drafts that hit usually came from Voice #1 (Sabrina) or Voice #2 (Hormozi). Drafts that flop usually came from Voice #4 (Tesla service manager) — too matter-of-fact for the algorithm. Knowing which voice works for which platform is the actual intel here.

How to use these

1. Save each prompt as a file in ~/.claude/prompts/ or your preferred prompt library.
2. Use a snippet expander (Raycast, TextExpander, Alfred) to make them one-keystroke accessible.
3. Re-read this file every few weeks — the prompts evolve as I learn what works on real builds.

What's NOT in here

• Anything proprietary to a paying-client build.
• Anything that promises a result without context. These are tools; the result depends on what you put in.
• "Magic" 10x-productivity prompts. Doesn't exist. These are just three I return to weekly.

More prompts coming

I drop new tested prompts on the channel as I build. If you want a specific one — DM @cory_salisbury what you're stuck on, I'll show you the prompt I'd use for it.