Where the team is strong and weak
The typical engineer on each skill, against engineers on HyperHat who took the same session.
HyperHat · Team benchmark
See how your whole team works with AI
Each engineer spends 30 minutes directing an agent through one timed exercise. You get the team pattern: where the team is strong, where it stalls, which habits repeat, and which kinds of work are ready to hand to agents.
- Team-level by default — engineers see their own full result
- 30 minutes per engineer
- No install, nothing touches your repo
Finished12 /12Unlocked at 3
StrongestPrompt Quality+13 vs HyperHat
Biggest gapTask Decomposition−3 vs HyperHat
0–100 · same bands as each engineer's report
Engineers on HyperHatYour team
How it works
From one link to a team picture.
-
Create the team
Name it and open a benchmark. You lead it and choose who joins.
A few minutes -
Invite engineers
Send one link, or invite by email. Each person picks the track closest to their work.
One link -
Engineers finish
Each directs a live agent through a real brief in the browser, and gets their own full report.
30 minutes each -
Read team results
Team results unlock once three people finish, and fill in as more do. Share an executive brief.
Team-level only
The six skills
Six skills that show how your team works with AI
Every session is read on the same six skills. Five describe how the work was directed. Output Quality describes what the session produced.
The team page shows the typical engineer on each one, never a personal score. A skill with no logged check is not scored, rather than scored as zero. Levels are Developing, Solid and Strong.
The research behind the six-
Output Quality
Measures the resultDid the finished work do what was asked.
StrongThe work does what the brief asked.
WeakShips something a reviewer has to rebuild.
-
Verification
Measures behaviorHow often work was checked by running it before calling it done.
StrongRuns a test or opens the preview first.
WeakAccepts the first output.
-
Iteration Efficiency
Measures behaviorHow directly each round of agent work moved the task forward, without wasted turns.
StrongEach instruction changes the work toward the brief.
WeakRepeats turns that do not move the task.
-
Prompt Quality
Measures behaviorHow specific and contextual the instructions to the agent were.
StrongNames the file, the error, and the expected result.
WeakGeneric prompts like “fix it” or “make it better”.
-
Task Decomposition
Measures behaviorWhether the task was broken into steps the agent could carry out.
StrongBreaks the work into steps the agent can finish one at a time.
WeakPastes the ticket verbatim.
-
Recovery & Debugging
Measures behaviorWhether agent errors and dead ends were diagnosed before re-prompting.
StrongChanges course when an approach fails.
WeakRetries the same prompt.
The team report
What you read after people finish.
One picture of the team, built only from team-level numbers. Drawn here from an example team of twelve.
How the team spends a session
Where the typical engineer's time goes, in the same four parts as each person's own report.
- 19% Directing
- 55% Working
- 22% Verifying
- 5% Away
Habits worth keeping
Good habits, counted across sessions. One seen in more than half of them is a team strength.
- Edited the agent's plan before approving it8 of 12
- Fixed a failing check and confirmed it ran clean5 of 12
- Checked the work every other instruction3 of 12
Habits that repeat
The ones to fix. A count small enough to point at someone reads “a few.”
- Retried a failing approach without checking in between5 of 12
- Submitted without running any verificationa few
- Approved the agent's plan without changesa few
Rules to merge
Suggested from the habits above, as files for Claude, Codex and Cursor.
@@ Team rules @@
+ After a failed attempt, read the output and say what changed before retrying.
+ Nothing is done until it has been run.
Delegation map
Which kinds of change this team can hand to agents, and under what conditions.
- Bug fixesOne reviewer checks the diffReady
- New featuresEdited plan, then a verification gateGated
- RefactorsTests pass before and afterGated
- Data migrationsA plan, a rollback and a dry run firstGated
Change since the last benchmark
The typical engineer before and after, and whether each habit got more or less common. A move inside the measurement noise reads as no detectable change, and people on both rounds are shown as a paired view.
Verification50→66Improved +16
Recovery & Debugging68→74Improved +6
Submitted without running any verificationall but a few→a fewNo longer frequent
The team page
See the whole team page before you start.
Captured from the app with an example team of twelve on its second benchmark. Your team's results replace the example once three people finish.
Privacy
Team-level by default.
Leads read how the team works, never how one person scored. Engineers see their own full result, and decide who else does.
- The typical engineer, not a ranking
Each skill shows the team's median and spread. No names, no leaderboard, no individual scores.
- Small groups stay hidden
Team results unlock at three finishers. Under ten people, values round to the nearest 5, and any count small enough to point at someone reads “a few.”
- Engineers own their report
A lead sees one person's report only if that person shares it, and they can revoke it at any time.
Pricing
One price for the team.
$3,500per team
Up to 20 engineers.
- A full report for every engineer
- The team report and delegation map
- Rules files for Claude, Codex and Cursor
- An executive brief to share
$1,000to start
Up to 5 engineers.
- The same reports, for a smaller group
- The team report and delegation map
- Rules files for Claude, Codex and Cursor
- Credited toward the team of 20
FAQ
Questions leads ask.
Can I see each engineer's score?
No. You see team results: the typical engineer on each skill, the spread, and how common each habit is. An engineer can share their own report with the team's leads, and revoke it at any time.
How many people need to finish?
Team results unlock once three people finish and fill in as more do. Under ten people, values are rounded to the nearest 5 and the spread is left out, so no one result can be worked out.
How long does it take?
About 30 minutes per engineer, in the browser, whenever suits them. Team results appear as soon as three people have finished.
Does it touch our code?
No. Each session runs in an isolated browser sandbox on a prepared brief. Nothing connects to your repository.
How do we measure progress?
Run a second benchmark about 90 days later. It reports the change for the typical engineer and each habit, and calls a move within the noise no detectable change.
Get started
Open a benchmark for your team.
Name the team, invite engineers, and share one link. Each person takes one timed session in their own track.
