Getting started

HyperHat measures how well you direct an AI coding agent through a real task. Each session runs in HyperHat Studio, a browser IDE with a live agent. HyperHat records your prompts, what the agent does, the checks you run and how you recover when something breaks, and turns that into a scored report. Your report stays yours. A team only ever sees group results.

For engineers

1. Sign up with Google, or with your email address. With email you confirm it by link or code, then set a password.
2. Pick a track and a level. They decide which sessions Dashboard offers, and you can change them later in Settings.
3. If a lead invited you to a team benchmark, join it before you start. The first session you start after joining is the one that counts for the team.
4. Start a session from Dashboard. Read the brief first; the clock starts only when Studio opens.
5. Direct the agent, check its work, and submit. Your report and replay are on Results.

Nothing is installed, and a session never touches your own repositories. Don't paste production code or secrets into it.

For team leads

1. Open Teams, name your team and create it. You become its lead, and its first benchmark opens straight away.
2. On the benchmark page, choose a plan and pay: Team, up to 20 engineers, or Starter, up to 5. Nobody can join, and no invite emails go out, until it is paid for.
3. Share the join link, or invite people by email. Invites added before paying are sent once it is paid.
4. Follow progress. Team results unlock once enough people finish.
5. Review the results, agree a few changes to try, and close the benchmark when the round is done.

Running a session

Each session is one timed task. You direct the agent the way you would at work: plan the change, give it instructions, check what it did, and recover when it goes wrong. It is a real brief worked with the agent, not a quiz.

Tracks and levels

There are eight tracks: Frontend, Full-Stack, AI Engineer, Rapid Prototyper, Data Analyst, Reliability, Technical Writer and Product Manager. Each has four levels: Entry, Mid, Senior and Expert. A level describes how demanding the task is, not how senior you are.

Time limits

• The brief page is untimed. The clock starts when Studio opens.
• There is no pause. The clock keeps running if you close the tab.
• You can resume. Refresh Studio, or choose Resume on Dashboard from any device, to return to the same session with the time that is left. It does not use another attempt.
• Time's up submits for you. At zero, the session is submitted and scored automatically.
• Abandoned sessions expire about an hour after their time limit if they never produced a report, and any credit they used is returned.

What is recorded

Your prompts, the agent's tool calls and edits, files you open and save, previews, terminal and lint runs, plan approvals and edits, copy actions, time away from the Studio tab, token use and the final state of the workspace, all with timestamps. These drive the score, the report and the replay. An AI model also reviews the transcript to write the report's review section, which does not change the score.

Your report

Every finished session is listed on Results. Open one for:

• Your score out of 1000, its band, your standing, and any cap with the reason for it.
• Six dimensions, each out of 100.
• Key moments, each linked to that point in the replay.
• Process integrity, activity and completion checks: how you directed the agent, what you ran, and which of the task's checks passed.
• A replay of the session and a prompt-by-prompt analysis with suggestions.
• An AI-written review of what worked and what to work on next. It does not change the score.

The replay and prompt analysis are included with every report and stay available, with or without a subscription.

Scores and evidence

Your overall score is out of 1000. It is criterion-referenced: it measures your session against the task's criteria and is not curved against other people. While a task is still being calibrated, the report labels the score Provisional.

Bands on the 0–1000 scale:

• Insufficient Evidence is below 430.
• Developing starts at 430.
• Proficient starts at 560.
• Strong starts at 700.
• Exceptional starts at 820.

A session handed to the agent wholesale is labelled Delegated instead of a band.

The six dimensions

Each dimension is shown out of 100, on your report and on team pages: Strong from 80, Solid from 65, Developing from 45, and Needs work below that. The names on your report are short; the research page uses longer names for the same six.

Task Decomposition · Task Decomposition and Planning

How you break the brief into steps: exploring before editing, opening with a plan, and reviewing the agent's plan instead of pasting the whole brief.

Prompt Quality · Steering Precision and Context

How specific your direction is: naming files and expected behaviour, saying what must not change, and giving the context the agent needs.

Verification · Verification and Code Review

Whether you check the agent's work yourself, with tests or commands, the preview, lint and reading what changed, and whether you do it between turns rather than once at the end.

Iteration Efficiency · Delegation Judgment and Scope Control

How directly you reach a working result: fewer wasted turns, no blind retries of the same request, and token use within budget.

Recovery & Debugging · Hallucination Recovery and Debugging

What you do when the agent's output is wrong or broken: diagnosing and changing course, rather than repeating the same instruction.

Output Quality · System Architecture and Code Quality

Whether the result meets the brief, judged by the task's automated checks and code hygiene such as lint.

How much each dimension counts varies by task. Output Quality and Verification carry the most weight on every task.

When the evidence is thin

A green outcome is not the whole score, and no session is left unscored. If the checks passed but you never verified the work, or you handed the brief to the agent unchanged, the score is capped and the report says why:

• Passed with flag: the task passed with little or no verification. The score cannot reach Proficient.
• Unverified: the checks passed, but the session shows no real direction or checking. The score stays inside Insufficient Evidence and has no standing.
• Criteria unmet: the task's required checks failed. The score stays below Developing.

Standing

Your standing ("top X%") compares your score with everyone who has a verified result on the same task and scoring version. It appears once at least 30 people qualify. Until then the report shows an estimate marked (est), based on a fixed curve rather than real results. Standing changes as more people take a task, and scores on different tasks are not directly comparable.

Teams

Teams is for a lead who wants a group picture. Open Teams from the rail. If you have no team yet, you name one and can invite people as you create it. If you already belong to one, you land on it. The team switcher, at the top of the side panel, moves between teams and starts another.

Leads and engineers

A team has two roles. Whoever creates it is a lead, and a team can have several leads. Any lead can make another member a lead, change a lead back to an engineer, or remove someone; leads cannot demote or remove themselves, so a team always keeps a lead. A lead can take sessions for their own report, but a lead's sessions do not count in the team's results.

Engineers see their own reports, never anyone else's score. Once a benchmark they took part in closes with results, they see the same team results as the lead. An engineer can leave a team from its page.

A benchmark is one round

Each benchmark is paid for once, before anyone can join. Places count the people who have joined, not invites or finishers. A re-benchmark adds everyone already on the team, which uses places straight away, and removing someone from the team does not free their place.

The first session someone starts after joining is their session for the benchmark. If they retake that same task before you close it, their best result counts. A session started before joining does not count.

Team results unlock once enough people finish, and then update in steps rather than after every finisher. A lead can make the minimum stricter in Team settings. It cannot be loosened again, and it applies to all of the team's benchmarks, past ones included. Each report an engineer has shared with the leads adds one to the number a lead's view waits for, because the leads already know that result.

• Close the benchmark when the round is done. That stops new joins and saves the baseline, and everyone on it can then read its team results. It cannot be reopened, and sessions finished afterwards do not count. If too few people had finished, it never gets team results.
• Delete permanently removes the benchmark with its team results, targets and adopted rules. Each engineer keeps their own report.

Re-benchmarking and targets

Re-benchmark later, usually about 90 days on, to see what moved. On a closed benchmark a lead can save a planned date, and the weekly lead email reminds you as it approaches. A re-benchmark is a new benchmark and is paid for like any other. It keeps the rules adopted on the benchmark before it, and its results show how often each habit came up before and after, marked where a rule went in.

The change card compares everyone, and also the same people on the same track and level where there are enough of them. Small moves are reported as no detectable change, and a change to the scoring between benchmarks is flagged. After two closed benchmarks, the team page charts the change per dimension.

On a closed benchmark's results a lead can set up to three targets for the next one, picked from suggestions: a dimension's typical engineer reaching a value, or a habit showing up in fewer sessions. Targets apply to team results only, and they stay fixed once the re-benchmark has results.

Team results

Results show the typical engineer on each dimension, the team's median, on the same 0–100 scale and bands as each engineer's own report. Each one is compared with typical on HyperHat: the middle result of everyone else who took the same session, one best verified result each and nobody from your team. With fewer than ten people, values are rounded to the nearest 5 and the spread is not shown, and small counts read "a few". Team pages never show an overall score or anyone's individual result.

On an unlocked benchmark the lead also sees:

• Where the team is strong and where the gap is, per dimension.
• How the typical engineer splits a session between directing the agent, working, verifying and time away, in the colours of the session bar on each engineer's report.
• Habits that showed up often, such as submitting without checking, and good habits worth keeping, such as checking the work as it goes.
• Suggested rules, which can be exported for CLAUDE.md, AGENTS.md, or .cursorrules. Mark a rule adopted once it is merged; adopted rules carry over to the re-benchmark.
• A delegation map: which kinds of change look ready for an agent with normal review, which need a plan and a verification gate, and which should stay human-led.
• A printable brief of the round.

The team page lists next steps: who to remind, invites to resend, rules to review, targets to set, and when to re-benchmark. Counts only, never names.

The Members page, for leads, lists each person's name, role, track and level, and where they are on the current benchmark: finished, in progress, not started or not joined. It never shows scores. When an engineer shares one of their reports with the team's leads, it appears there under Shared with you until the share ends.

Once a benchmark you took part in closes with results, its page shows you the same team results your lead sees, with the suggested rules to download. A card only you see puts your own result on each dimension next to the typical engineer on the benchmark. While a benchmark is open, its progress and results are your lead's to follow.

Sharing

Your report is only on your account unless you share it. Decide what to show and who should see it before you send a link.

• Public report link. Anyone with the link can read that report without an account; it never shows your other sessions. It does not expire. You choose what it shows; by default that includes your name, the replay, your prompts and the AI review. Revoking it turns that address off for good, and sharing again makes a new one.
• Share with your team's leads. If you are an engineer on a team, you can let that team's leads open the report in HyperHat for 7, 30 or 90 days, 30 unless you change it. Your prompts, the replay and the AI judge review stay out unless you include them. You can revoke it at any time, and it ends if you leave the team.
• Team report and brief links. A lead can share a benchmark's report or brief with someone outside the team as a link that needs no sign-in. Each link expires after 30 days, and the lead can revoke it sooner. Engineers can see on the team page which links still work and the day each expires.

Plans and billing

Team plans

• Team: $3,500 for one benchmark of up to 20 engineers. A full report for every engineer, the team report and delegation map, rules files for Claude, Codex and Cursor, and an executive brief to share.
• Starter: $1,000 for one benchmark of up to 5 engineers, with the same reports for a smaller group.

One payment covers one benchmark, and a re-benchmark is paid for again. Engineers never pay for their session on a paid team benchmark, and it does not use up their free personal session. While a Starter benchmark is open, a lead can upgrade it to 20 engineers by paying the Team price less what was paid for Starter. A team coupon, if you have one, applies to the Starter plan. A benchmark holds at most 20 engineers; for a bigger group, email hello@hyperhat.com. Prices are in US dollars.

Individual sessions

Your first session is free. After that, HyperHat Pro is $29 a month for unlimited sessions; start it from the banner at the top of the app, and manage or cancel it from Dashboard. A promo code gives 30 days of Pro. It does not renew, and if you already have Pro it adds 30 days. Past reports stay on Results either way.

Paid purchases are generally non-refundable, except where the law requires a refund or we say otherwise when you buy. See the Terms.

Account and privacy

Settings holds your display name, profile picture, and session track. The track is what Dashboard uses to recommend the next session.

Deleting your account, from the danger zone in Settings, is immediate and cannot be undone. It removes your name, email, password and photo, turns off your public report links, ends reports shared with leads, takes you off your teams and cancels HyperHat Pro. Your session results stay without your name and still count in team results they were part of. If you are the only lead of a team with other people on it, make someone else a lead or delete the team first.

Sessions run in an isolated sandbox, but what happens in them is recorded (see Running a session). The Privacy Policy covers what is collected, how it is protected and how long it is kept. For privacy requests, email privacy@hyperhat.com.

Co-Director

Co-Director is a separate product, in alpha with a small number of engineering teams. It runs locally alongside the agent in your terminal or IDE: it adds context and your team's rules before the agent runs, flags drift and loops on the same failure, and turns repeated corrections into a pull request against your rule files. It never commits for you. It is not part of a benchmark purchase. To ask for access, use Co-Director Alpha Access in the site footer.

FAQ

How long does a session take?

30 to 90 minutes, depending on your track and level. The brief page shows the exact limit, and the clock starts when Studio opens.

Can I take it again?

Yes. Each track and level has several tasks, all scored on the same 0–1000 scale. The first session is free. After that you need HyperHat Pro or a promo code.

Can my lead see my score?

No. A lead sees team results once enough people finish, never one person's. Your score, grade, and replay stay on your account unless you share that report yourself, by link or with your team's leads.

Why does my report say Unverified or capped?

The work may have met the brief, but the session showed little or no checking of the agent's output. If you did verify and still see this, email us with the session details.

What if my team is small?

The benchmark can still run. Team results stay locked until enough people finish. You still get your own report.

I paid, but the benchmark still says it isn't paid for.

Wait a minute and reload; the page updates as soon as the payment is confirmed. If it does not, email us. Don't pay again.

How do I contact support?

Email hello@hyperhat.com. We typically reply within one business day. Include your account email, the team and benchmark if any, when the session ran, and what happened. Never send passwords, card details or secrets. The full privacy terms are in the Privacy Policy.