← HUBCOMMUNITY/PROMPTS
05 · PROMPT PACK · COPY + USE

Battle-tested prompts.
Tested across Claude, GPT, Gemini.

Each prompt was used at least 10 times in real work before it landed here. They're not "you are a helpful assistant" filler — they're structured, named, and tuned to the outcome you actually want.

2 OF 10 CATEGORIES LIVE · 100+ PROMPTS COMING20% COMPLETE
live prompts · ready to copy

Two flagships — the prompts I use daily.

Full prompt text + variables + tested-on + gotchas. Drop into Claude Code with /prompt-name or paste directly. Both work across Claude 4.7, GPT-5, Gemini 3 with model-specific tuning notes.

PROMPT 01 · DEBUG · v1.3

Find the bug — not just any bug, the right one.

For when something's broken and "fix this" isn't enough context. Forces the AI to hypothesize before patching, isolate the root cause, and give you the minimal fix — not a rewrite.

USES · 84×
HIT RATE · ~92%
STATUS · v1.3
CATEGORY
DEBUG
BEST MODEL
Claude 4.7
TOKENS
~480 in
COST AVG
$0.008

When to use

  • A test is failing and the message isn't obvious
  • Production code worked yesterday, doesn't today, and you can't pinpoint the diff
  • You're 80% sure what's broken but want a sanity check before shipping the fix

The prompt

CONTEXT: I'm working on {project description}. Something is broken. WHAT I OBSERVE: {paste the error, the failing test output, OR the wrong behavior} WHAT I EXPECT: {1 sentence on the correct behavior} RECENT CHANGES: {paste the last git diff OR "no recent changes — was working last week"} TASK: 1. Hypothesize the 3 most likely root causes. Rank by probability. 2. For the top hypothesis: what's the MINIMAL test or check that would confirm or refute it? 3. Once confirmed, what's the MINIMAL fix? Patch — not a rewrite. Do not propose a fix until I've confirmed the hypothesis. Refuse refactoring requests in this thread.

What it does differently

  • Forces hypothesis-first thinking — no jumping to "here's the fix" before isolating the cause
  • Refuses scope creep — the "refuse refactoring" line stops the AI from rewriting your file unprompted
  • Asks for the minimal test — usually a 1-line check, not a full test suite
  • Ranked probabilities — you learn which gotcha to check next time
TUNED FOR Claude 4.7 (best at hypothesis chains) · GPT-5 (good but tends to skip step 2 — add "ALWAYS run step 2 before step 3") · Gemini 3 (rambles in step 1 — limit to "3 hypotheses, 1 line each").
PROMPT 02 · CODE-REVIEW · v1.1

Code review — like a senior engineer who's been on-call.

Reviews like someone who's actually shipped this kind of code at 3am. Catches security gotchas, perf footguns, missing edge cases. Won't say "looks good" — explicitly lists what was checked, and if it's clean, says so with the list.

USES · 52×
CATCHES · ~3 / PR
STATUS · v1.1
CATEGORY
CODE-REVIEW
BEST MODEL
Claude 4.7
TOKENS
~620 + diff
COST AVG
$0.014

When to use

  • Before every PR merge — second opinion on the diff
  • Before deploying to production — even on solo projects
  • When you wrote it tired — catch the obvious stuff you'd otherwise miss

The prompt

ROLE: Senior engineer who's been on-call. Reviews like the next 3am page might come from this code. CONTEXT: {1 sentence — what this change does and why} DIFF: {paste git diff OR the changed files} CHECK FOR (and report on EACH explicitly): 1. Security — auth bypass, injection, exposed secrets, missing input validation 2. Performance — N+1 queries, unbounded loops, blocking I/O on hot paths 3. Edge cases — null/empty/timeout/concurrency 4. Observability — does it log enough to debug at 3am? 5. Naming + readability — would a teammate understand without context? 6. Scope — anything that should be a separate PR? OUTPUT: - For each of the 6 areas: either flag a specific issue with line ref, or say "clean — checked X, Y." - Then a prioritized fix list: P0 (blocks merge), P1 (fix this PR), P2 (follow-up acceptable). - If everything is clean, end with "Approve. Checked: [the 6 areas]." Do not propose stylistic preferences. Do not rewrite. Flag and explain only.

What it does differently

  • Forces explicit per-area reporting — no "looks good" hand-waving
  • Severity-ranked — you know what to fix now vs. later
  • "Refuses rewrites" — prevents the AI from sending you a 200-line "improved version"
  • Frames as on-call engineer — shifts the model toward defensive, observable code
PRO TIP Add this as a Claude Code /code-review skill — frontmatter user-invocable: true so it only runs when you call it.
10 prompt categories · the full library

What's coming.

Each category will ship with 8–12 prompts. Two are live (above). Eight more in queue. Total target: 100+ prompts across the library when complete.

01
Debug
Hypothesis-first, minimal-fix debugging across stacks
12 PROMPTS
LIVE
02
Code-review
Senior-on-call style reviews · security, perf, edge-cases
8 PROMPTS
LIVE
03
Research
Multi-source synthesis · citation-required · uncertainty-aware
10 PLANNED
NEXT
04
Content
Blog drafts · X threads · newsletter · LinkedIn carousels
12 PLANNED
NEXT
05
Ops
Daily brief · weekly metrics · investor update drafts
8 PLANNED
SOON
06
Sales
Outreach · follow-up · objection-handling · qualification
10 PLANNED
SOON
07
Email
Cold · warm · re-engagement · cancel/save · onboarding sequences
12 PLANNED
SOON
08
PRD & spec
Product requirements · scope cuts · acceptance criteria
8 PLANNED
SOON
09
Customer support
Empathy-first · de-escalation · forward-looking · doc-pointers
10 PLANNED
SOON
10
Technical writing
Docs · changelogs · post-mortems · API references
10 PLANNED
SOON