# The Job Sheet
The Job Sheet is a free one-page reference — a plain markdown file, no install, no account — that
lists which AI model to open for which job, and the single reason each one wins its row. It comes
with a short routing rule you can paste into your CLAUDE.md or your assistant's custom
instructions so the decision stops being something you re-argue every few weeks.
It exists because "which AI is best" is the wrong question, and the video that goes with it shows
why: the same four models, judged across three different jobs, physically trade places.
## The three rounds, and what they actually show
Writing code. Gemini is the weakest of the four here. ChatGPT is decent. Claude takes it — it is
where developers route real production work, which is a stronger signal than any single benchmark
row.
An hour-long recording you need the contents of. Claude is the skip, and not by a little: it
cannot accept a video file at all, in the API or the web app. ChatGPT does have real video handling
— it samples frames at roughly one per second and transcribes the audio — but it gets unreliable
past about five to ten minutes. Gemini takes this one because it tokenizes video natively, which has
been true since Gemini 1.5 and is architectural rather than a benchmark-of-the-week.
Ten thousand runs, as cheaply as possible. Claude is the skip. ChatGPT is decent. DeepSeek takes
it at roughly 15 to 100 times cheaper per token than the mainline paid tiers, for bulk work where you
do not need the frontier.
Now read those three rows together. **Claude is the pick in round one and the one to avoid by round
three.** Nothing about Claude got worse in the ninety seconds between them. The job changed.
The part nobody says out loud
ChatGPT lands "decent" in all three rounds.
That is not a knock, and it is the most useful line on the page. The model most people have open
right now is genuinely fine at everything — which is exactly why you never notice you are using the
wrong one. A tool that is never wrong is also never the reason a hard job goes well.
Why the sheet is organised by job, not by rank
A "best AI of 2026" list is wrong the week after it ships. The jobs you actually do are stable. So
the sheet is a routing table: when a model leapfrogs another, you edit one row and the structure
still holds. That is the whole compounding idea — organise around the half that does not move.
Install it in four minutes
1. Read the table once. You are looking for the two or three rows where you are currently using
the wrong tool. Not for a new default.
2. Do not switch your default. The tool you already have open is fine at everything. The entire
gain is in switching for the handful of jobs where the gap is large, and leaving the rest alone.
3. Paste the routing rule into your CLAUDE.md or custom instructions, so the decision is
written down instead of re-made.
4. Re-check one row when a model ships. Never the whole sheet.
What this deliberately does not do
It quotes no benchmark percentages. The public numbers for these models disagree across leaderboards
— the same model, the same month, different figures on different sites, a lot of them auto-generated
pages carrying invented scores. Every verdict here is a relative one between named alternatives on a
named job, because that is the part that survives the disagreement.
One round was rewritten outright before the video was recorded. The first version claimed Gemini wins
long-context work "and it isn't close." That turned out to be wrong twice over: context windows
converged to about a million tokens across all three labs in March and April 2026, so nothing chokes
on capacity any more, and on the stringent eight-needle retrieval test at that length Gemini places
last rather than first. The round became the video round instead.
Terms and tiers move fast. Re-check any pick against the provider's own page before you rely on it.
Get it
Comment TIERS on the reel and the sheet comes straight back to you, or take it here — every job,
the pick, and the one reason it wins.