One prompt. Seven days. Eighteen thousand dollars. Here’s the exact playbook — the prompt, the workflow, the mistakes, and why it actually works.
The Email That Started It
It was a Tuesday. I was halfway through my third coffee when a founder I’d met once at a meetup messaged me on LinkedIn:
“Hey — our AI ‘copilot’ feature is a mess. Users hate it. Can you audit it and tell us what to fix? Budget is open.”
I almost replied with my usual “sure, let’s scope a two-week engagement.” Instead, I opened Claude Code, pasted one prompt I’d been refining for months, and shipped a full technical audit + rewrite plan in 48 hours.
By Friday I’d invoiced $6,000. By the following Tuesday, two more founders had hired me off that same deliverable. Total for the week: $18,000.
This post is the prompt. And — more importantly — the reason it works.
TL;DR
- One prompt + Claude Opus 4.7 + a real repo = a deliverable clients will pay four-figures for.
- The prompt forces Claude to act like a senior consultant, not a chatbot.
- The output is a decision document, not code. That’s what gets you paid.
- The full prompt is at the bottom. Steal it. Adapt it. Send the invoice.
Why “Just Use ChatGPT” Doesn’t Work
I’ve watched a lot of devs try to monetize AI and bounce off. The pattern is always the same:
- They paste a vague request: “review my codebase.”
- Claude returns a polite, generic checklist.
- The client reads it and thinks: I could have Googled this.
- No second invoice.
The problem isn’t the model. It’s that most people prompt Claude like a search engine instead of a teammate. A senior engineer wouldn’t hand a client a list of “consider adding tests.” They’d say: “Your retry logic in chat_service.py is why your p99 latency is 11s. Here’s the fix, here’s the risk, here’s the rollout plan.”
That’s the gap the prompt closes.
The Setup (5 minutes)
You need three things:
- Claude Code (terminal, IDE, or desktop — pick your poison).
- Opus 4.7 selected for the heavy reasoning passes. Sonnet 4.6 for the cleanup.
- A real codebase — either the client’s repo (with permission) or a representative slice they’ve shared.
Install Claude Code if you haven’t:
irm https://claude.ai/install.ps1 | iex # Windows
curl -fsSL https://claude.ai/install.sh | sh # macOS / Linux
Drop into the repo:
cd ./client-repo
claude
Then paste the prompt. That’s the whole setup.
The Prompt That Did the Work
Here it is. No fluff, no “you are a helpful assistant.” Just the thing.
You are acting as a Principal Engineer doing a paid technical audit of this
codebase for a client who will read your output as a decision document.
Your job is NOT to be polite. Your job is to be specific, opinionated, and useful.
Do this in four passes, in order. Do not skip ahead.
PASS 1 — MAP
- Walk the repo. Build a one-page mental model of the system: entry points,
data flow, external integrations, deploy target.
- Output: a 10-line architecture summary a non-technical founder can understand.
PASS 2 — RISK
- Identify the top 5 things that will hurt this company in the next 90 days.
Examples: security holes, scaling cliffs, data loss vectors, vendor lock-in,
compliance gaps, on-call nightmares.
- For each: severity (Sev1–Sev3), the exact file/line evidence, blast radius,
and the cheapest credible fix.
- No generic advice. If you can't cite a file, don't list it.
PASS 3 — LEVERAGE
- Identify the top 3 changes that would 10x the team's shipping velocity.
Think: missing CI, missing types, missing observability, the one refactor
that unblocks four future features.
- For each: estimated effort (S/M/L), expected payoff, who on the team owns it.
PASS 4 — DELIVERABLE
- Produce a single Markdown document titled "Technical Audit — <repo name>"
with these sections:
1. Executive Summary (5 bullets, founder-readable)
2. Architecture at a Glance
3. Top Risks (from Pass 2, sorted by severity)
4. Top Leverage Moves (from Pass 3)
5. 30 / 60 / 90 day roadmap
6. What I would do first if this were my company
- Tone: senior, calm, direct. No hedging. No "it depends." Pick a side.
Constraints:
- Cite file paths and line numbers for every claim.
- If you don't know something, say "Unknown — need to ask: <question>."
- Do not write code in this document. Code goes in follow-up tickets.
- Length target: 1,500–2,500 words. Anything longer, cut it.
Begin Pass 1.
Run it against a real repo. Walk away for ten minutes. Come back to a document that — with light editing — is what a $250/hr consultant would have delivered after three days of meetings.
Why This Prompt Actually Works
Four design choices, each load-bearing:
1. It assigns a role with stakes. “Principal Engineer doing a paid audit” isn’t flavor text. It changes the posture of the response. Claude stops hedging.
2. It forces sequential passes. Most prompts let the model jump straight to recommendations. This one makes it understand before it judges. The Map → Risk → Leverage → Deliverable pipeline mirrors how real consultants think.
3. It demands evidence. “If you can’t cite a file, don’t list it.” That single line is what separates a deliverable from a horoscope.
4. It defines the artifact. Claude isn’t asked to “help.” It’s asked to produce a specific document with specific sections in a specific tone. Constraints aren’t a cage — they’re what turns output into something a human will pay for.
The Real Multiplier: The Follow-Up Conversation
The $6k audit is the door-opener. The $12k that came after wasn’t from new clients — it was from the same client asking the obvious next question:
“Okay, this is great. Can you actually fix #1 and #2 from the risk list?”
This is where Claude Code earns its keep. The audit document becomes the spec. Each Sev1 risk becomes a ticket. Each ticket becomes a branch. Opus 4.7 drafts the fix, you review, Sonnet 4.6 writes the tests, you ship the PR.
A workflow that used to take a sprint now takes an afternoon. The client sees a Sev1 close before they’ve finished reading your audit. That’s when the follow-on invoice gets approved without negotiation.
Mistakes I Made So You Don’t Have To
- I undercharged the first one. $6k was cheap for what landed. Senior eyes on a codebase, with a written deliverable and a roadmap, is $10–15k of value minimum. Price the outcome, not the hours.
- I tried to automate the client conversation. Don’t. The prompt produces the document; you present it. The 30-minute walkthrough is where trust (and the next contract) is built.
- I let Claude write the executive summary first. It came out generic. Now I make Claude do it last, after Passes 1–3, so the summary is grounded in actual findings.
- I forgot to ask permission. If you’re running this on a client’s private repo, get written consent. Not optional.
What This Is Not
It’s not a get-rich scheme. It’s not “AI replaces consultants.” It’s not even particularly clever.
It’s one well-crafted prompt that turns Claude into the version of itself most people never see — the one that behaves like a teammate who actually read the code, formed an opinion, and is willing to defend it.
The reason it makes money is the same reason senior engineers make money: someone with taste, evidence, and the guts to say “do this, not that.” Claude can do the reading. You still have to bring the taste.
Steal the Prompt. Send the Invoice.
Three asks before you go:
- Run the prompt on your own codebase first. You’ll be uncomfortable with how accurate it is. That discomfort is your proof it’s ready for a client.
- Edit the output by hand before sending. Strip Claude-isms. Add one observation only a human would make. That’s the difference between a deliverable and a leak.
- Charge what it’s worth. A technical audit that ships in 48 hours and accurately predicts the next outage is not a $500 deliverable.
The age of the lone senior engineer with leverage isn’t coming. It’s here. The prompt above is one of the doors.
Now go open it.
About the Author
Suraj Khaitan — Gen AI Architect | Building scalable platforms and secure cloud-native systems
Connect on LinkedIn | Follow for more engineering and architecture write-ups
Tried the prompt? Made your own version? Drop the result (or your invoice number, anonymized) in the comments — I read every one.
Top comments (0)