All Briefs
Mission Brief · 003—·updated

by Łukasz Graliński

80/20 is over.
Meet 95/5.

In this illustrative model: in 2019, 20% of the effort got you 80% of an MVP; in 2026, 5% gets you 95% of what a demo shows. These are assumptions, not measurements by year. The remaining 5% of the product — the part under the waterline — still costs most of the work. Drag the sonar and see what it's made of.

5%
of the effort · the demo above water
46%
of PirxeyOS commits were fixes
1.6×
commit rate after go-live · half of them fixes
Lower the sonar
Iceberg · interactive

Lower the sonar. Price what the demo skipped.

Above the waterline: the demo — screens, the happy path, sample data. Below it: eight layers of work no demo shows and no prompt asks for, because nobody knows they exist yet. Pick your terrain, tell us how long the demo took, and drag the sonar down.

Terrain
hours

a weekend is ~16 h · a solo sprint ~80 h

WHAT THE DEMO SHOWSWHAT PRODUCTION REQUIRES25%50%75%100%SONAR 34% · 103 h uncovered
Visible above water16 h · 5% of the effort
Uncovered so far103 h of 304 h under water
20×the demo, to reach production on this terrain — a model, calibrated on our own repos

  1. Entries that cross midnight. Duplicates between pages. Time zones. Deletes that cascade.

    Nobody puts these in a prompt, because nobody knows they exist yet.

    PirxeyOS commitfix: handle midnight in tracker form

  2. Checking that the model did what it said. 140 test files in PirxeyOS alone.

    Generated code is cheap to produce and expensive to trust.

    PirxeyOS commitfix(e2e): fix failing smoke and fixture tests

  3. Who sees whose hours. Row-level rules. What a manager can and cannot open.

    The demo has exactly one user — you.

    PirxeyOS commitfix(reports): scope member filter by access

  4. Version skew between clients. Input validation. A vendor's undocumented quirk.

    The model cannot read the other system's mind — or its bugs.

    PirxeyOS commitfix: version skew error handling (#260)

  5. iOS zooming into inputs. Narrow viewports. Dark-mode placeholders. Empty states.

    The demo ran on one laptop, in one browser, with data that fit.

    PirxeyOS commitfix: prevent iOS mobile zoom also on time entry edits

  6. 81 database migrations. Error tracking. Deploys, rollbacks, importing the old tool's history.

    Nothing in a demo has a yesterday it must stay compatible with.

    PirxeyOS commitfix(clockify): resolve staging mappings safely

  7. Reports across long ranges. Four hundred entries in a day. Screens that stay fast.

    Ten rows of test data never get slow.

    PirxeyOS commitfix: fix slow entries render by not rendering all forms immediately

  8. Training users. Feedback loops. The second, third and fourth release.

    Shipping starts the work; it does not end it.

    PirxeyOS commitafter go-live, commits rose to 24 a week — half of them fixes

01 · Model, not measurement

How the numbers are built

Total effort = demo hours ÷ visible share. The terrain sets the share (12% · 5% · 3%). The tail is split across eight layers in illustrative proportions inspired by our fix categories, not measured shares of work or time.

02 · Calibration

587 commits, one product

PirxeyOS: a working system in ~20 days, then 30 weeks of keeping it alive. 46% of all commits are fixes; from the go-live week onward, the average weekly rate was about 60% higher. These are measured commit counts; the layer weights are model assumptions.

03 · What it doesn't say

AI is not the problem

22% of those commits were co-authored by Claude or Codex. The model wrote plenty of the tail. Somebody still had to know the tail existed, ask for it, check it and ship it.

Same feature · two products

What the demo shows. What production requires.

Four everyday features. On the left, what a weekend build delivers — and honestly delivers well. On the right, what the same feature has to do the day real users, real data and real money show up. None of the right column is exotic. It is just never in the demo.

01
Demo vs Production

Sign in.

DemoEmail and password. Works on the first try, because you typed it and you know it.
ProductionSSO for the one customer that requires it. Password reset. Rate limits. Session expiry. An audit trail of who signed in from where. Six features hiding inside one button.
The button was the 5%. The rest is what the button promises.
02
Demo vs Production

Save an entry.

DemoClick save, a row appears. The sample data fits on one screen and never overlaps.
ProductionEntries that cross midnight. Two people editing the same row. A delete that has to cascade to the running timer. Reports that must still add up afterwards. Every one of these is a real PirxeyOS fix.
Data has physics. Demos don't.
03
Demo vs Production

Show the report.

DemoOne chart, ten rows, instant. Everyone sees everything, because there is only you.
ProductionA manager sees only their own team. Six-month ranges stay fast. The export matches the screen to the cent. The old tool's three years of history are in there too. Permissions, performance, migration — three layers for one screen.
The chart is a screenshot. The report is a contract.
04
Demo vs Production

deploy — ship it.

DemoOne command. You are the only user, so nothing can break for anyone else.
ProductionMigrations that run without downtime. Error tracking. A rollback plan. Version skew between someone's old tab and your new API. Support when it breaks on Monday at 8:03. Shipping starts the work; it does not end it.
A prototype is finished when it runs. A product is finished when it survives.
The rule · what changed

Pareto didn't break. It got steeper.

The 80/20 rule was never about effort being evenly spread — it was about the last stretch being disproportionately expensive. AI didn't remove that stretch. It made the first stretch nearly free, which makes the last one look even bigger by comparison.

01
5% of the effort

The visible part got cheap

A working screen, a data model, an API, a login — an evening with Claude or Codex. This is real. We use it every day. It is also exactly the part that a demo, a pitch and a screenshot show.

— Pirxey field model · calibrated on PirxeyOS, 2026
02
46% of commits are fixes

The invisible part didn't move

Entries that cross midnight. Duplicates between pages. Who is allowed to see whose report. iOS zooming into a field. None of it is hard — it is just not visible until a real user hits it. In our own product, nearly half of all commits are fixes.

— pirxey-os repo · 587 commits · Feb–Sep 2026
03
15 → 24 commits / week

Go-live starts the clock

Before rollout the repo saw ~15 commits a week. From the go-live week on: 24 — and 52% of them fixes. The demo ended in week 4. The product started in week 22.

— pirxey-os repo · weekly commit history
04
22% co-authored by AI

AI wrote a lot of the tail, too

This is not an anti-AI argument. One in five commits carries a Claude or Codex co-author line — including fixes. The point is who knew the fix was needed, asked for it, checked it and shipped it. That person is the 5%.

— pirxey-os repo · Co-Authored-By trailers
05
30–40% → 15–20%

Speed is felt before it is real

Stanford measured nearly 100,000 developers: AI raised delivered code volume 30–40%, and rework ate half of it — net gains about 15–20%, close to zero on complex existing codebases. Developers' own estimates of their productivity were off by about 30 points. The visible part is faster; the delivered part waits on the tail.

— Stanford software engineering productivity study, 2025
06
permissions · data · money

The tail is where the risk lives

The layers under water are also the ones that leak data, double-charge a customer or lose a day of entries. A prototype that skips them isn't 95% of a product — it is 100% of a demo.

— Pirxey delivery playbook · 2024–2026
Evidence · receipts

The numbers — ours and other people's.

The iceberg's weights are assumptions. The commit counts and research below are the evidence you can open and check.

587
commits, 46% of them fixes.
PirxeyOS — the operating system we are building for our own company: time tracking, HR, a skills matrix, employee records, onboarding, still growing. Feb 7–Sep 7, 2026: 104 feature commits, 269 fix commits, 81 database migrations, 140 test files. A working system in ~20 days — then 30 weeks of keeping it alive.
pirxey-os repository · Pirxey case study
19%
AI made experienced devs slower on mature code.
METR's 2025 randomized trial: 16 senior open-source developers, 246 real tasks on codebases they knew. With AI allowed, tasks took 19% longer — and the developers still believed they had been faster.
METR (2025)
+91%
More PR activity, without an organization-level delivery gain.
Faros telemetry: developers in high-AI teams interact with 47% more PRs per day; PRs are 154% larger, review time 91% longer, bugs per developer 9% higher. AI adoption was not correlated with organization-level delivery gains. These are associations, not proof of what caused them.
Faros AI — Lab vs Reality
80% / ↓
feel more productive with AI; delivery stability still drops.
DORA 2025, nearly 5,000 respondents: 90% use AI at work and more than 80% believe it made them more productive. Adoption now correlates with higher throughput — and still with lower delivery stability. DORA calls AI an amplifier. We call the gap the waterline.
Google Cloud — 2025 DORA report
~33%
of a developer's week goes to technical debt.
Stripe's 2018 Developer Coefficient survey put technical debt at roughly a third of a developer's week — 13.5 of 41.1 hours. Its broader maintenance measure was 17.3 hours. The tail was always there. AI just made the head cheaper.
Stripe — The Developer Coefficient (2018)
8
layers. Fixed proportions. Open to argument.
The layer weights (edge cases 24%, tests 16%, permissions 14%, integrations 12%, devices 12%, operations 10%, performance 6%, people 6%) are illustrative assumptions inspired by our fix categories, not measured shares of work or time. Your product will differ. That's what the call is for.
Pirxey field model
Mission control standing by

Send us the demo. We'll map the iceberg under it.

We write custom software, some pieces are ready-made, and we join your team and work alongside it. First step: we size the invisible 5% of your product — layer by layer, in hours — before anyone commits a budget. Free. No slide deck.

Price the invisible 5% with usBook a call
Pirxey · Aleja Grunwaldzka 472, 80-309 Gdańsk, Poland·130+ engineers · 100+ missions delivered