Tokens Are Dev Hours. Waste Is a Feature, Not a Bug.

Last week I built a production-ready app from scratch using AI.
One week. One developer. Real users. Real infrastructure.

The catch? I didn’t do it alone — I had AI teammates.

And they weren’t free.

Not just in credits.
In attention. In retries. In debugging mistakes they made.

I’ve been thinking a lot about what that cost actually is. Here’s the pattern I’m seeing:

Tokens are dev hours.
Prompt tax is real.
And waste is no longer invisible.

The Traditional Model: Burn Time, Not Tokens

In a normal dev cycle, bad code and bad assumptions still cost time — but that time is baked into salaries. Hidden in Jira boards. Spread across "velocity." You don’t get a line-item receipt for every wrong guess.

If your senior dev spends 4 hours debugging an integration that was supposed to take one, you probably won't even know. At best, it shows up in a retro.

With humans, waste is absorbed.
With AI, waste is metered.

And once you see it — you can’t unsee it.

The AI Engineer Has a Price Tag

Every prompt has a cost. Every mistake has a token count. Every time the model forgets what you said three minutes ago, it burns another 1,200 tokens — and your patience.

Here’s what I noticed while building the LPO Bracket Manager:

• Initial schema prompt → ✅ Valid result (~$0.14)
• Rewrite with fix → ✅ Clean tweak (~$0.11)
• Regeneration attempt → ❌ Wrong syntax (~$0.21)
• Re-prompt with extra context → ✅ Good output (~$0.13)

That’s $0.59 just to get one chunk of usable SQL.
Now multiply that by 500 interactions a week.

Prompt Tax: The Cost of "Almost There"

I call this the prompt tax — the cost of getting the AI to do something you thought it already knew how to do.

You pay this tax when:

• The AI hallucinates an import path
• It forgets the file structure you just gave it
• It reverts a function to a prior, incorrect state
• You burn tokens teaching it not to repeat a bad pattern

A strong junior dev might take 3 hours.
AI might take 4 minutes — but with 5 retries and $1.12 of token waste.

Speed and cost are now uncoupled.
Attention is quantized.
And you’re the one footing the bill.

Waste Is Now Visible — and That’s Good

For years, engineering waste has been invisible:

• Slack threads re-explaining things
• Code reviews correcting basic logic
• Duplicate work across sprints
• Context resets after PTO

We never really tracked that waste — we just accepted it.

But with AI, every retry is measurable.
Every hallucination has a price.
And every bad output forces you to ask:
Why did this happen? How do I prevent it next time?

That’s not frustrating — that’s an opportunity.

Designing for AI Waste

Over the last few weeks, I’ve shifted how I build systems for AI teammates:

• I create prompt scaffolds with tone, formatting, and constraints
• I write coding guidelines specifically for the model
• I use Kilo AI to format prompts like tasks in VSCode
• I maintain Docusaurus docs-as-code to reduce re-explaining logic
• I added a theme management layer to isolate design tokens — normally a full-team feature, done in two rounds with AI

This isn’t just development.
It’s training the system not to waste my time.

You don’t just ship code anymore.
You shape the environment that makes code shippable.

From Token Burn to Team Discipline

If you’re leading a team, here’s the shift:

• Stop measuring just output — start measuring retries
• Budget for prompt tax — literally and emotionally
• Create AI-ready docs, not just human-facing ones
• Recognize that context shaping is now a team skill

The dev team of the future won’t just be fast — it’ll be self-healing.
Because it will treat AI not as magic, but as a junior teammate who can learn — if you build the systems for it to remember.

Closing Thought

AI doesn’t eliminate waste.
It exposes it.
In tokens. In retries. In your Stripe receipt.

That’s a feature, not a bug.

It means we can finally see where our attention actually goes —
and we can design systems to reduce waste at the source.
Not just by writing better prompts…

…but by teaching the system to forget less.

✶

I’m building real-time prompt analytics tools, documenting waste patterns, and architecting AI-led apps in the open.
Follow along at blog.nino.photos, or DM me to collaborate.

Let’s build like waste matters.