Why Combos Beat Single Codes
A single Claude prompt code does one thing. It changes tone, or forces commitment, or strips AI tells, or surfaces assumptions. Useful.
Stacks of 2-4 codes do something different: they compound. One code sets up the context Claude attends to, another shapes the reasoning, a third controls output format, a fourth strips the tells. Each solves a different dimension of the response.
Most founders who ask me "what's your favorite Claude prompt" are asking the wrong question. The answer isn't one code. It's a stack for the specific task.
Below are 10 tested combo patterns. Each is a paste-ready template for a common task.
1. Cold Email That Gets Replies
Stack: /punch /trim /ghost
What each code does:
- /punch sharpens every sentence so it lands
- /trim cuts the polite filler that signals AI
- /ghost strips em-dashes, transitional phrases, "I hope this finds you well"
Template:
/punch /trim /ghost Write a cold email to [role] at [company]. Context: [why you're reaching out — be specific]. The ask: [the one concrete next step you want]. Keep under 90 words. No greeting fluff. No closing sign-off niceties.
Why it works: Each code removes a different category of AI-reveal. /ghost alone leaves filler. /trim alone leaves AI phrasing. Together they produce email that reads like a real person typed it on their phone.
2. Decisive Recommendation on "Should I X or Y"
Stack: L99 /skeptic
What each code does:
- L99 forces commitment instead of hedging
- /skeptic rejects the question's framing if it's broken before answering
Template:
L99 /skeptic Should I [option A] or [option B]? Context: [what you're optimizing for, constraints, what's been tried]. Pick one. Give the reason. If neither is right, say so and propose option C.
Why it works: /skeptic catches the case where you're asking the wrong question. If the question is well-formed, L99 forces a specific recommendation instead of the usual "here are the tradeoffs" enumeration.
3. Pre-Mortem On A Plan
Stack: /blindspots /skeptic OODA
What each code does:
- /blindspots surfaces assumptions you didn't state
- /skeptic rejects the plan's framing if it's broken
- OODA structures the response (Observe, Orient, Decide, Act)
Template:
/blindspots /skeptic OODA Here's my plan for [what you're doing]: [paste plan]. Audit it. What assumptions am I making? What am I not seeing? If the plan is directionally wrong, propose the actual plan.
Why it works: Most plans fail on the unstated assumptions. /blindspots finds them. /skeptic tests whether your overall approach is even correct. OODA structures the feedback so you can act on it.
4. Reframing A Stuck Problem
Stack: /skeptic /blindspots
Template:
/skeptic /blindspots Here's what I think the problem is: [your current framing]. Here's what I'm planning to do about it: [your current plan]. Tell me what I might be wrong about — both about the problem and about the solution.
Why it works: You've been looking at this problem for weeks. You're in a cognitive rut. Two codes that both attack assumptions + framing break you out of it.
5. Summarize A Long Document Without Fluff
Stack: /trim /raw L99
What each code does:
- /trim strips meta-commentary and filler
- /raw forces plain text without "Here's a summary:" preambles
- L99 forces Claude to commit to which points actually matter
Template:
/trim /raw L99 Summarize this document. Only include points that change a decision or change my understanding. Skip background. No introduction, no conclusion. Output as a numbered list ranked by importance. Document: [paste]
Why it works: Raw-trim combo removes the "here's what I'm about to say" framing Claude normally adds. L99 forces ranking — you get the 3 points that matter, not 10 equally-weighted bullets.
6. Debug A Flaky Test
Stack: /skeptic L99 OODA
Template:
/skeptic L99 OODA This test fails about [N%] of the time. Code: [paste test + SUT]. Stack trace when it fails: [paste]. Don't propose a fix yet. First: rank the 3 most likely root causes. Pick the most likely. Propose the diagnostic step that would confirm or rule it out.
Why it works: /skeptic stops Claude from accepting your guess at the cause. L99 forces commitment to a most-likely root cause. OODA structures the response so you get a diagnostic plan, not a random fix.
7. Code Review With Actual Specificity
Stack: SENTINEL /blindspots /punch
Template:
SENTINEL /blindspots /punch Review this code. Focus on: failure modes I haven't handled, security/auth gaps, race conditions, and assumptions that aren't enforced. For each issue: name the line, name the risk, name the fix. Code: [paste]
Why it works: SENTINEL makes Claude scan for errors. /blindspots adds "what aren't you considering" scanning. /punch forces specific, actionable feedback instead of mealy-mouthed "consider edge cases."
8. LinkedIn Post That Doesn't Read Like LinkedIn
Stack: /ghost /punch /hook
Template:
/ghost /punch /hook Write a LinkedIn post about: [topic]. Make it feel like a real person typed it on their phone — short paragraphs, specific examples, one strong opinion. No "thrilled to share," no "so honored," no hashtag lists. Length: 4-6 short paragraphs.
Why it works: /ghost strips LinkedIn-isms. /punch forces concrete claims. /hook makes the first line earn the read in the feed.
9. Tone-Match An Email
Stack: /mirror /voice
Template:
/mirror /voice First read this email as a tone reference: [paste their original]. Now write my reply. Reply needs to: [your content points]. Match their tone exactly — same formality, same length, same rhythm.
Why it works: /mirror analyzes the reference writing. /voice locks Claude into the matched tone for the whole response instead of drifting back to default.
10. Synthesize Multiple Sources Into A Decision
Stack: L99 /punch PARETO
Template:
L99 /punch PARETO I'm trying to decide [actual decision]. Here are 5 sources: [paste or describe]. Don't summarize each one. Tell me: which 2-3 insights actually change the decision, what those insights are, and what I should do.
Why it works: L99 forces a recommendation. /punch keeps the synthesis specific. PARETO surfaces the 20% of insights that drive 80% of the decision.
The Meta-Rule For Stacking
Two codes: almost always better than one. Three: usually better. Four: sometimes better. Five or more: rarely better.
Beyond 4 codes, you're diluting the effect of each one. Claude's attention budget spreads thin, and the response becomes weirdly formatted (trying to satisfy all 5 codes at once).
The sweet spot for most tasks is 2-3 codes. Pick the stack based on what dimensions need to change:
- Reasoning dimension: L99, /skeptic, /blindspots, /deepthink, OODA
- Format dimension: /raw, /bullets, /json, /punch, /trim
- Style dimension: /ghost, /voice, /mirror, /casual
- Specificity dimension: PERSONA, /focus, SENTINEL
One code per dimension. 2-3 dimensions per task. That's the formula.
Bad Combos To Avoid
Stacking placebos: ULTRATHINK + GODMODE + ALPHA + EXPERT = 4 codes, zero effect. All placebo. See anti-patterns library.
Stacking conflicting codes: /punch + /deepthink is a contradiction — one wants terse, one wants thorough. Output will be confused.
Stacking the same dimension: /raw + /bullets + /json all control output format. Pick one. Using three gives Claude contradictory instructions.
FAQ
Is there a shortcut to finding the right stack?
Yes. The Combo Generator lets you type a task ("write a cold email," "debug a flaky test") and returns the exact stack + paste-ready template. 6 free recipes, 14 more in the Cheat Sheet.
Do these combos work on Claude Haiku 4.5?
Most do. Combos with /deepthink or long PERSONA setups work slightly worse on Haiku. The writing combos (/ghost /punch /trim) work equally well. See our Haiku 4.5 testing notes.
How long do combos take Claude to process?
About 30-50% longer than single codes. If you're doing interactive iteration, it feels slower. For one-shot important tasks, the quality gain justifies the wait.
Can I invent my own combo?
Yes. Pick one code per dimension (reasoning, format, style, specificity). Test pair-wise against single codes. Keep the stack if output is measurably better. See how to test Claude prompts.
Are there industry-specific combos?
Yes. The Pro Cheat Sheet has 20 bonus prompts that are essentially pre-built industry combos for legal review, sales coaching, content editing, and financial analysis.
Try One Combo Right Now
Pick the single biggest task you're procrastinating this week. Match it to a stack from above:
- Cold email →
/punch /trim /ghost - Stuck decision →
L99 /skeptic - Plan review →
/blindspots /skeptic OODA - Tone-match a reply →
/mirror /voice
Paste the template into Claude. Fill in your specifics. Run.
30 seconds in, you'll see why stacks beat single codes.