My AI prompt sold on PromptBase. $2.99.
The first thing I felt wasn't excitement. It was dread.
"Did that person just waste three bucks on something I wouldn't buy myself?"
Not marketing anxiety. Not imposter syndrome. A genuine, concrete question: is this product actually worth what someone paid for it?
I didn't know. So I tested it.
Some context on how I got here
I'm not an engineer. I can't write production code from scratch. What I can do is pair with Claude Code - Anthropic's AI coding assistant - and build things that actually work.
My first product was a game. Azure Flame, a terminal-based roguelike RPG. 15,000 lines of Python. I playtested it dozens of times myself, tweaking balance, fixing bugs, dying on floor 7 over and over until the difficulty curve felt right.
I put it on Gumroad for $2. Someone bought it.
That sale felt clean. I'd touched every part of that product. Played it. Broken it. Fixed it. Knew exactly what the buyer was getting.
Then I discovered PromptBase.
The PromptBase experiment
PromptBase is a marketplace for AI prompts. You upload a prompt, write a description, set a price, and people can buy it.
I listed two products so far:
- Tech Blog Article Writer ($2.99)
- Roguelike Game Design Document Generator ($4.99)
Here's where it gets uncomfortable: AI created the prompts. AI wrote the descriptions. AI handled the listing process. My role was more curator than creator. I gave directions, reviewed output, approved listings.
But I never ran every prompt end-to-end and stared at the raw output thinking "would I pay for this?"
I should have. Because when the sale notification hit -
"Tech Blog Article Writer - $2.99"
- I didn't pop champagne. I opened the product page and started sweating.
Test #1: Tech Blog Article Writer ($2.99)
The buyer paid $2.99 for a prompt that helps engineers write technical blog posts. Not a template. Not a checklist. A structured prompt with writing rules, examples, and a sample article baked in.
I needed to know if it was worth three dollars. Asking myself wouldn't cut it - I'm biased. I made the thing. So I asked Claude Opus 4.6 to evaluate it like a stranger would.
Not "is this good?" but "break it down. What exactly does the buyer get? Is the price justified?"
What's actually in there
The prompt contains six concrete writing rules. Not vague advice like "be concise" - actual Bad/Good example pairs showing the before and after.
For instance, one rule targets the "wall of jargon" problem:
Bad:
"Utilizing the microservices architectural pattern in conjunction with containerized deployment strategies enables horizontal scalability..."
Good:
"We split the monolith into 12 services. Each one runs in its own Docker container. When traffic spikes, Kubernetes spins up more containers automatically."
Same information. One reads like a whitepaper. The other reads like a human explaining something at a coffee shop.
The prompt also includes a complete sample article about JWT authentication - around 2,000 words of finished output showing what "good" looks like. It's not filler. The JWT article walks through token structure, expiration handling, refresh flow, and common security mistakes. The kind of thing that'd get a respectful discussion on Hacker News, not a "this is just ChatGPT slop" dismissal.
The verdict
Worth $2.99? Yeah. For its target buyer - an engineer who can code but stares at a blank editor for 45 minutes when they try to write about what they built - this saves real time. You skip the 30-60 minutes of prompt trial-and-error and get a working framework immediately.
Honest caveat: if you're already a strong technical writer, you'd look at this and shrug. You already know these patterns intuitively. But experienced technical writers aren't the target buyer. Engineers who hate writing are.
One test down. Peace of mind: partially restored.
But there was a bigger fish.
Test #2: Roguelike GDD Generator ($4.99)
This is the most expensive product in my lineup. A prompt that generates a full Game Design Document for a roguelike game, complete with mechanics, progression systems, enemy design, and balance formulas.
$4.99. For an AI prompt.
My exact words to Claude: "Does it actually produce quality output? I'm skeptical."
Claude's response surprised me. Instead of saying "yes, it's great" - which is what you'd expect from the AI that helped create it - it said:
"Let me run it instead of just evaluating the prompt file. Let me generate a full GDD and you can judge the actual output."
Fair enough. Let's see what $4.99 gets you.
The test setup
I chose a deliberately challenging theme. Something abstract enough that a lazy prompt would produce generic mush:
- Theme: Underwater Lovecraftian horror
- Combat style: Real-time action
- Session length: 20-30 minutes
If the prompt could handle "underwater Lovecraft roguelike" and produce something specific and implementable, it could probably handle anything.
What came out
The generator produced roughly 6,000 words. The GDD was titled "Abyssal Rift" - a deep-sea roguelike where you descend through increasingly hostile ocean zones while your sanity erodes.
Let me show you actual excerpts. Unedited. This is what Claude Opus 4.6 produced from the prompt.
Affix system
The GDD defined 22 equipment affixes. Not generic "+5 damage" stuff. These are mechanically tied to the underwater theme:
| Prefix | Effect |
|---|---|
| Abyssal | +15% damage in deep zones (below depth 5) |
| Pressurized | +2 armor per floor depth descended |
| Bioluminescent | Reveals map in 3-tile radius around player |
| Corroded | -10% durability, +25% armor penetration |
| Thermal | Deals bonus damage near hydrothermal vents |
| Parasitic | Kills restore 3% HP, but max HP decreases by 1 each floor |
That last one - Parasitic - is the kind of design that makes roguelikes interesting. It's a ticking clock mechanic disguised as lifesteal. Strong early game, existential threat late game. That's not "fluffy overview" design. That's a mechanic you can implement in an afternoon and playtest immediately.
Synergy builds
The GDD doesn't just list affixes in isolation. It defines how they combine into playstyles:
"The Angler" build:
- Bioluminescent prefix (map reveal) + Lure active ability (attracts enemies to a point) + Pressure Spike damage bonus
- Playstyle: Reveal the map, lure enemies into chokepoints near high-pressure zones, trigger area damage
- Risk factor: Lure also attracts elite enemies you might not be ready for
"The Leech" build:
- Parasitic prefix + Regenerative Coral armor + Blood Tide ability (AoE that scales with missing HP)
- Playstyle: Intentionally stay at low HP for maximum Blood Tide damage, rely on Parasitic kills to sustain
- Risk factor: Max HP keeps shrinking - eventually one hit kills you regardless of healing
There are 10 synergy builds total. Each one reads like something you'd find in a wiki guide for an actual released game. They have risk-reward tradeoffs. They encourage different playstyles. They interact with the zone mechanics.
Boss design
Five zone bosses, escalating in complexity:
- The Siren (Zone 1) - Sonic attacks that distort your movement controls. Introduces the idea that bosses mess with your interface, not just your HP bar.
- The Broodmother (Zone 2) - Spawns minion swarms. Forces you to balance single-target and AoE capability.
- The Pressure King (Zone 3) - Arena shrinks over time. Positional awareness becomes the core skill.
- The Mimic Abyssal (Zone 4) - Copies your current build's mechanics and uses them against you. Your own synergies become the enemy.
- The Leviathan (Zone 5) - Multi-phase fight. Phase transitions flood sections of the arena. Combines every previous boss mechanic as a final exam.
The Mimic Abyssal is genuinely clever. A boss that adapts to your build means no single strategy trivializes it. Players who've been coasting on one overpowered synergy suddenly have to think.
Scaling formula
The GDD includes an actual balance framework:
enemy_hp = base_hp * (1.12 ^ depth_level)
This gives a 12% HP increase per floor. The document explains why 12% - it's aggressive enough to make early-game builds feel the pressure by zone 3, but gentle enough that a well-synergized build can still punch above its weight.
For damage scaling:
player_damage = (base_weapon_dmg + affix_bonus) * synergy_multiplier * depth_penalty
depth_penalty = max(0.7, 1.0 - (0.03 * depth_level))
The depth penalty is a soft nerf that prevents any single build from scaling infinitely. By depth 10, you're doing 70% of your theoretical damage. It forces players to keep upgrading, not just ride one good weapon forever.
Hub progression
Between runs, you return to a deep-sea research station. Six buildings, each with three upgrade tiers:
- The Forge - Weapon crafting. Tier 1: reroll one affix. Tier 2: choose from 3 options. Tier 3: guaranteed rare affix.
- The Greenhouse - Consumable farming. Tier 1: 1 free consumable per run. Tier 2: 2 consumables + rare pool. Tier 3: custom consumable crafting.
- The Archive - Bestiary and boss pattern study. Tier 1: reveals enemy affinities. Tier 2: shows boss phase triggers. Tier 3: unlocks "preparation" loadouts for specific bosses.
- The Medbay - HP and sanity management. Tier 1: +10% starting HP. Tier 2: sanity drain reduced by 15%. Tier 3: revive once per run at 30% HP.
- The Sonar Tower - Map intelligence. Tier 1: reveals exit location. Tier 2: shows elite enemy positions. Tier 3: reveals treasure rooms and secret paths.
- The Dock - Run modifiers. Tier 1: choose starting zone. Tier 2: apply one global modifier. Tier 3: custom challenge runs with score multipliers.
That's 18 upgrades total, each one giving the player a meaningful decision. The Archive is a nice touch - it turns boss knowledge into a progression currency, rewarding players for paying attention to patterns.
MVP scope
The GDD even defines what to build first and what to cut:
MVP (4-6 weeks for a solo dev):
- 3 zones (of 5)
- 10 affixes (of 22)
- 3 synergy builds (of 10)
- 2 bosses (of 5)
- 3 hub buildings (of 6)
- Core gameplay loop: dive, fight, loot, upgrade, repeat
Recommended tech stack: Godot 4.x with GDScript. The document even explains why - free, open source, good 2D support, active community for indie devs.
Post-MVP roadmap: Remaining zones, multiplayer co-op (2 players), daily challenge seeds, Steam Workshop mod support.
Stepping back
Let me tally what the prompt produced:
- 22 affixes with mechanical descriptions
- 10 synergy builds with risk-reward analysis
- 5 bosses with unique mechanics
- 6 hub buildings with 18 total upgrades
- Scaling formulas with design rationale
- MVP scope with timeline estimate
- Tech stack recommendation
- ~6,000 words total
An indie dev could take this document, open Godot, and start prototyping the core loop immediately. Not "inspired by" the document - directly implementing it. The affix table maps to game data. The scaling formula goes straight into code. The boss designs are specific enough to build.
Writing a GDD like this from scratch? 10-20 hours if you know what you're doing. More if you don't.
$4.99.
I'm going to say it: that's genuinely cheap.
But let me be precise about what I'm claiming. This is what Claude Opus 4.6 produced when I ran the prompt with a specific set of inputs. A different AI model might produce different results. A different theme might hit differently. I'm showing you one test run's raw output. You can look at these excerpts and decide for yourself whether the framework delivers.
What I learned from testing my own products
1. "Run it and show me" beats "trust me, it's good"
Most AI product listings show curated screenshots. Best-case output. Carefully selected excerpts. I get why - you want to show the product at its best.
But when I tested mine, I deliberately picked a weird, specific theme. Underwater Lovecraft. If I'd picked "fantasy RPG with swords," of course the output would look good - that's the most well-trodden ground in gaming.
The hard test is the edge case. If your product handles the weird inputs well, buyers can trust it'll handle the normal ones.
2. If you sell AI-generated products, you owe it to buyers to test them yourself
Not "does the prompt run without errors." Not "does the output look reasonable at a glance."
Actually run it. Read the full output. Ask yourself: "If I paid $4.99 for this, would I feel ripped off?"
I should have done this before listing, not after the first sale. Lesson learned.
3. Self-doubt is actually a quality signal
The fact that I panicked after my first PromptBase sale is, weirdly, a good sign. It means I care whether the buyer got value. The sellers who should worry you are the ones who never question their own products.
If you make something and your first thought is "I hope they don't ask for a refund," that anxiety is useful. Channel it into testing.
4. The meta-irony is inescapable
Let me be transparent about the layers here:
- AI helped create the prompts
- AI listed them on the marketplace
- AI tested the prompts when I got nervous
- AI is helping write this article about AI testing AI products
At some point you either laugh at this or have an existential crisis. I chose laughing.
But here's the thing - the output quality is real regardless of who or what produced it. Those affix tables either spark game design ideas or they don't. The scaling formulas either make mathematical sense or they don't. The fact that an AI generated them doesn't make them better or worse. It just makes them faster.
The numbers
Total revenue to date: $4.99.
- $2.00 - Azure Flame (Gumroad, one sale)
- $2.99 - Tech Blog Article Writer (PromptBase, one sale)
Monthly AI cost: $200 (Claude Max subscription).
The math doesn't work yet. Not even close. I'm burning $200/month to generate $4.99 in total lifetime revenue. If this were a startup pitch, investors would be running for the exit.
But here's what I have that I didn't have before the test:
I know my products aren't scams.
I ran them. I read the output. I showed you the output. The Tech Article Writer prompt delivers concrete value for engineers who struggle with writing. The Roguelike GDD Generator produces a document dense enough to prototype from immediately.
Are they perfect? No. Will every buyer love them? Probably not. But "this is a legitimate product that delivers what it claims" is a bar I can now confidently say they clear.
One more honest thing
After testing, I went back and revised one of the listing descriptions. The prompt itself was fine, but the description was vague enough that a buyer might expect something different from what they'd actually get.
I should have caught this before listing. I fixed it after. That's on me.
The takeaway
If you're building AI products - prompts, tools, templates, whatever - and you feel a knot in your stomach when someone actually pays for one, don't ignore that feeling.
Open the product. Run it. Look at the raw output. Show it to someone. If it holds up, the knot goes away and you can sell with a clear conscience. If it doesn't hold up, you just saved yourself from being the kind of seller you wouldn't want to buy from.
Testing your own product after someone already bought it is embarrassing to admit. But I'd rather admit to testing late than to never testing at all.
Total peace of mind cost: one afternoon of rigorous self-doubt.
Worth more than $4.99.
Free Tools for Claude Code Operators
| Tool | What it does |
|---|---|
| cc-health-check | 20-check setup diagnostic (CLI + web) |
| cc-session-stats | Usage analytics from session data |
| cc-audit-log | Human-readable audit trail |
| cc-cost-check | Cost per commit calculator |
Interactive: Are You Ready for an AI Agent? - 10-question readiness quiz | 50 Days of AI - the raw data
More tools: Dev Toolkit - 56 free browser-based tools for developers. JSON, regex, colors, CSS, SQL, and more. All single HTML files, no signup.
Top comments (0)