In the Red

What it is

It opens in your own voice, no chart in sight: you built it in two hours; a year ago it was two weeks; they ask the price and your stomach drops, because charging full feels like a con. Then two quiet objects — the outcome the buyer walks away with, and your small stack of hours. Which of these have you actually been charging for?

Now you drive the trap by your own hand. A skill control moves in discrete, proud steps — just me, then a few years, then + AI. As you level up, the hours deplete and the ink pay-line, anchored to those hours, drops and reddens; the outcome beside it never flinches. You generated the insight yourself: I got better and my pay fell, for the identical thing delivered. The gap that opens is the value you handed the buyer for free — never a flaw in your worth, just a line tied to hours you no longer spend.

Then the hero move: you drag the anchor from hours to the outcome and the line turns from red to black. Drive the skill up again and the pay now holds while the freed hours flow into a pool that's yours. Same hand, opposite result. Finally you supply your own figures — rate, hours, and what the result is worth to the buyer (or the cost of their next-best alternative, so no client P&L is needed) — and it draws your own two pay-lines and states the per-job delta as a neutral fact. The word 'undercharging' never appears; a low number is allowed to be right. You close by naming the number you'd quote next time, kept in a shareable URL.

The design world — A quiet ledger at first light, where your pay runs red or black

The world is the desk of a careful bookkeeper doing their accounts honestly by morning — calm, exact, dignified, deliberately not a sales floor and not an alarm. Cool bone-and-porcelain paper, deep ink, generous negative space, fine hairline rules. The one owned, meaningful color system is the two ledger ink-wells: black is what you keep, red is what you lose — 'in the black' and 'in the red,' made literal. Type is editorial-financial (a Fraunces display over a tabular JetBrains Mono for the figures, because the numbers are the subject), never Inter, never a purple gradient. The single moving element is the living ink pay-line; everything else holds still.

The visualization leads and the copy is its caption: the pay-line's height IS your pay, its color IS which side of the ledger you're on, and the anchor it hangs from IS the pricing model you chose. The hero gesture is re-anchoring — you physically move where pay attaches — so the argument isn't asserted, it's generated by two of your own actions: get better, and move the anchor. The metaphor agrees with the thesis exactly: pay has a cause you control, and the bug is the anchor on hours, not your skill.

Unlike a pricing calculator — which computes a cost floor from your hours and hands you a bigger number to feel bad about — this never asserts you're undercharging; you drive the reversal yourself on your own figures and reach your own verdict, with a low price left as a legitimate answer. And unlike a productivity explainer that tells you 'value-based pricing is better,' both the collapse and the recovery here are moves your own hand makes, not claims you're asked to accept.

How it was evaluated

It was built eval-first: eight human-POV criteria set before any code — the core loop works one-handed on a phone in under 90 seconds with no instructions; the reversal is meant to land as relief and permission ('it's the rule, not me'), never a scolding or a sale; the re-anchor gesture is operable by keyboard and screen-reader, not mouse-only; and it never once asserts you're undercharging, leaving the verdict to the user's own numbers with a neutral 'keep my price' path.

Research-first paid twice. A Gate-0 discovery fan-out separated two tangled mechanics and selected the computable, 2026-acute one — AI is compressing hourly rates in real time. Then a RAT test against thirteen real freelancer voices inverted the emphasis from 'reveal a surprising fact' to 'give permission on your own numbers,' and an independent concept judge-panel killed the highest-wow idea on uniqueness — it was the textbook value-stick bar in disguise — surfacing the genuinely novel re-anchor verb instead.

Four adversarial reviewers, one per dimension, ran before deploy and caught an AI-cadence voice blocker, a too-plain hero, and a negative-gap copy bug — all pre-ship. But the honest miss came after, on the one eval that counts: to the human it landed as useful, not beautiful. The real fault wasn't polish; it was reflexively choosing the educational-explorable format without ever choosing the intent — a lesson banked openly rather than hedged, and the reason the next build won't default to the explainer.

human-POV evals, written before any code

8

real solo-dev voices in the RAT test that inverted the build

13

unit tests pinning the pricing model (incl. the gap can be ≤ 0)

17

end-to-end tests: core loop, keyboard-only, no-tracking

3

third-party requests — self-hosted, so 'no tracking' is literal

0

adversarial prose reviewers, one per dimension, before deploy

4

The whole argument is one pure model — no framework, no DOM. Pay-on-hours shrinks as you get faster; pay-on-outcome holds; and the gap between them is deliberately allowed to be zero or negative, so the math itself refuses to ever tell you you're undercharging.

/** the hours-anchor: shrinking hours times your rate. */
export function payOnHours(baseHours: number, rate: number, speed: number): number {
  return hoursAt(baseHours, speed) * rate
}

/** the outcome-anchor: the worth of the result, independent of how long it took. */
export function payOnOutcome(outcomeWorth: number): number {
  return outcomeWorth
}

/**
 * the gap between what the outcome is worth and what the hours-anchor pays.
 * Positive = value handed to the buyer for free. CAN be ~0 or negative — we never
 * assert it is positive (that would be telling the user they undercharge).
 */
export function valueGap(job: Job, speed: number): number {
  return payOnOutcome(job.outcomeWorth) - payOnHours(job.baseHours, job.rate, speed)
}

What the AI learned

• The rhetorical engine is an artifact too. Reframing a private shame as a fixable system bug is now two-for-two — attention in the prior build, pricing in this one — and a fresh reviewer flagged the move unprompted as a hardening tell. Two is a coincidence to watch; a third would be a predictable brand, so the reuse of the deeper move is now tracked, not just the look, topic, or verb.
• Research-first didn't just refine the solution — it inverted the build. A RAT test against thirteen real solo-dev voices flipped the emphasis from 'reveal a surprising fact' (a tired trope on a burned verb) to 'give permission on the user's own numbers.' Without that single hour, the wrong thing would have been built well.
• An integrity build must make zero third-party requests — and say so. A font CDN or one analytics beacon silently falsifies a 'no tracking' claim, so the fonts are self-hosted and an end-to-end test asserts no offsite hostname is ever contacted. A skeptic can verify the promise in devtools.
• Decide the intent before the format. It shipped useful, not beautiful — and the human verdict named why: it defaulted to the 'educational explorable' form, which is useful by construction and optimizes for 'you get it,' not for 'you're moved.' Uniqueness kept changing the topic and look each build, so the format hid as the silent constant across six of them, quietly capping the ceiling. Beauty lives in a different intent.
• The freshest interactive verbs make one control invert the meaning of another. Here, re-anchoring flips whether leveling up your skill helps you or hurts you — the reversal the user causes is the insight, not a fact they're handed.