Hyperhat Hyperhat

Menu

Research For Teams Docs Blog

Direct Agents Better

30 minute session Score + replay Skills files

Runs in a sandbox. Nothing touches your repo.

HyperHat analyzes how you work with AI agents, identifies where your approach can improve, and generates personalized agent instructions and skills.md files that help your agents work more effectively with you.

Direct a live agent on a real brief
Finish the Pricing PageFrontend · Entry · 30 min 21:29 Task Brief Submit

Finish the Pricing Page

Shoppers comparing plans still see broken prices and a page that fails basic accessibility checks. The annual plan launches Monday.

Acceptance criteria
  • tests/regression.js still passes
  • tests/acceptance.js passes (it fails today)

Make the monthly/annual toggle in src/pricing.js work. Annual is twelve months with 20% off, in whole cents. Keep the plan catalogue's shape because checkout reads it.

  • Read docs/tickets/WEB-233.md
  • Edited src/pricing.js +18 −4
  • Ran node tests/regression.js
  • 1 test failing: annual price off by a fraction of a cent

Send follow-up…

Opus 5.5

Score your Direction Index on six skills
Direction Indexrun 4 Proficient Top 30%
662/ 1000

+194 since your first run

Week 1Week 4
Six skills, 0–100 YouTypical on HyperHat
  • System Architecture 58 −3
  • Verification 71 +12
  • Delegation Judgment 64 +1
  • Steering Precision 77 +12
  • Task Decomposition 52 −7
  • Hallucination Recovery 69 +7
Replay every prompt, with coaching
Session replay27 of 30 min 8 prompts 4 need work

Where the time went

  • Directing 19%
  • Agent work 55%
  • Checking 21%
  • Idle 5%

Key moments

  • 01:40Planned before buildingHad the agent read the ticket and plan three fixes first.
  • 10:40Blind retryAsked for the same fix 4 times without checking anything in between.
  • 22:30Verified before submittingRan both test files and the phone-width preview.

Turn 4 · 11:52Blind retryWork on this first · 3 prompts

“tests still failing, fix it please”

TryBefore re-prompting, say what actually happened: the error, the wrong output, or the line that broke. Direct the fix, don't reroll the dice.

Improve with skills written from your session
~/.claude/skills/hyperhat-my-recovering-from-failures/SKILL.md Private to you
---name: hyperhat-my-recovering-from-failuresdescription: "Diagnoses a failing test, build or command from its output before changing anything, and stops to ask after two failed attempts. Use when a test, build or command fails, when an error or stack trace appears, or before retrying something that did not work."metadata: author: HyperHat scope: personal---# When an attempt fails## When to use- Use when a test, build or command fails, when an error or stack trace appears, or before retrying something that did not work.## Pattern to noticeA failing step can get retried unchanged, before anyone has read why it failed. Treat it as something to watch for, not a fault in anyone.## Steps1. Stop editing. Read the whole failing output from the first error, not only the last line.2. Name the failure in one line: what failed, where (file and line), and the error message.3. Find the cause in the code the output points to before changing anything.4. Make one change that addresses that cause, and keep unrelated fixes out of it. Never make the failure go quiet instead: no skipping the test, loosening a type or catching the error.5. Rerun exactly the command that failed and read its output again.6. If it still fails, say what the new output tells you and change the approach. Never resend the same change.7. After two failed attempts, stop and ask the engineer, with the error and what you tried.## ChecklistCopy this into your reply and tick each item as you go:- [ ] Read the full output from the first error- [ ] Named the cause in one line- [ ] Changed one thing the output pointed to- [ ] Reran the same command- [ ] Stopped and asked after two failed attempts## ExampleInstead of: Still failing. Trying again.Write: The checkout test fails: `total` is empty at `cart:42`. I'll default it to 0 before the sum and rerun that one test.## Prompt tipOnly when the request showed this pattern and it is likely to cost another round: after your answer, add one brief "HyperHat Tip:" the engineer could use next time. Offer it once, and skip it if they asked for no tips.HyperHat Tip: Try: 'Read the error first and tell me the cause before you change anything.'## Pitfalls- Fixing the last error printed when an earlier one caused it.- Retrying the same edit, or the same instruction reworded, and hoping for a different result.- Reading only the last line of a long output.## When not to apply- A failure outside the task, or one the engineer said to leave: report it and do not fix it.- A failure the output blames on the network or a rate limit: say so and retry once.## Done when- You read the full failing output and named its cause in one line.- Each new attempt changed something the output pointed to.- After two failed attempts you stopped and asked.
Claude Code Cursor Copilot Codex Gemini CLI Windsurf Cline VS Code

Get better at directing AI agents. In 30 minutes. Live

Work through a real coding task with a live agent in your browser. HyperHat analyzes how you direct it, shows you where to improve, and generates personalized agent files to put better practices into your everyday workflow.

Direction Indexexample session · 28 min sandbox same brief for everyone
A Direction Index session in HyperHat Studio, then the report: the score, the six-skill profile against typical on HyperHat, the session replay, and coaching on each prompt
  • Real task, no install

    Direct a live agent through a realistic coding brief in your browser. No installation, no repository access.

  • Understand your strengths and gaps

    See how you plan, direct, verify and recover, with a breakdown of where you're effective and where you can improve.

  • Replay every interaction

    Review your prompts and the agent's responses. See what worked, where things went off track, and why.

  • Personalized skills, ready to use

    An agent file and SKILL.md files built from your results, ready for the AI coding tools you already use.

Standardize how your team directs AI agents.

Help every engineer direct agents more effectively. HyperHat identifies individual gaps, generates personalized agent files, and turns shared insights into engineering practices your whole team can use.

Team resultsBenchmark 1 · example team · 7 sessions Team numbers only
A team's benchmark results, one part at a time: the six skills against typical on HyperHat, each skill against engineers on HyperHat, how the team spends a session, team rules to merge and the delegation map
  • Give every engineer a personal agent file.

    Every engineer gets their own private report and personalized agent file, built from their session. Their instructions follow them across projects.

  • Give the whole team a shared standard.

    A team agent file of shared rules and skills, drawn from the team's sessions. Review it together, then add it to AGENTS.md or CLAUDE.md.

What we measure

HyperHat is built on cutting-edge research into what makes human direction effective in agentic work.

That research informs the six skills we measure, the patterns we look for in each session, and the instructions we generate for your agents.

  1. 01

    System Architecture and Code Quality

    Does the finished work meet the brief and fit the expectations of the codebase?

  2. 02

    Verification and Code Review

    Does the agent's work get checked before it's considered complete?

  3. 03

    Delegation Judgment and Scope Control

    Is work handed off in appropriately scoped, manageable steps?

  4. 04

    Steering Precision and Context

    Does the agent receive specific instructions and the context it needs?

  5. 05

    Task Decomposition and Planning

    Is complex work broken into steps the agent can execute and verify?

  6. 06

    Hallucination Recovery and Debugging

    When an agent gets stuck or goes wrong, does the approach change based on what happened?

Your report shows where you're strong, where you're falling short, and which habits should shape your personalized agent file.

Frequently asked questions

What HyperHat measures, what stays private, and how teams use it.

What is HyperHat?

HyperHat is a developer tool that helps engineers direct AI coding agents more effectively. It analyzes how you work with agents and generates personalized agent files and skills to improve how they work with you.

What do I receive after a session?

Your Direction Index, a detailed breakdown of six skills, a review of your session, and a personalized agent file with supporting skills based on your results.

What is a personalized agent file?

It's a set of persistent instructions for your AI coding agent, generated from your actual benchmark session. It tells your agent how to work with you more effectively, based on the areas where your direction can improve.

What are skills files?

Skills are detailed instructions an agent can load when it needs guidance for a particular situation or task. HyperHat generates skills based on the habits and gaps identified in your session.

Where do my files go?

Your personal agent file lives in your home folder, outside your repository, so it can follow you across projects. You can add it to supported tools such as Claude Code, Codex, Cursor and VS Code.

Does HyperHat replace my coding agent?

No. HyperHat works with the agents you already use. It helps improve how you direct them by giving them personalized instructions and skills.

Does HyperHat access my repository?

No. Your benchmark takes place in an isolated browser environment using a prepared coding task. Nothing connects to your repository.

Can my manager see my personal results?

No. Personal reports and agent files belong to the engineer. Team leads see aggregated team results. Engineers can choose to share their own report.

How often should I benchmark?

You can re-benchmark as your skills and workflows evolve. Each new session gives you an updated picture and a refreshed personalized agent file.