A guide written so you can follow along even if you don't know a single line of code. Source: https://github.com/tobyilee/team-bigfive

0. One-Sentence Summary

A configuration bundle that makes multiple AIs divide up roles and collaborate like "one team," so you get better output from the same AI.The finding from Team Science that "a group of ordinary people with good teamwork produces better results" has been applied directly to AI agents. The point isn't to use a smarter AI — it's to makethe AIs collaborate with each other better.

1. Three Concepts to Understand First

Let's start with the parts non-developers get confused about most often.

(1) This is a "configuration bundle," not a "program"

It's not an app you double-click to run. It's a collection of rule and role files that sit on top of a tool called Claude Code. Think of it as a "recipe + kitchen duty roster" handed to a chef (the AI).

(2) It only works if you have "Claude Code"

Claude Code is a tool made by Anthropic that lets you give Claude tasks from a terminal (the black command-line window). Team Big Five Harness is an add-on configuration that switches how Claude Code operates into "team mode."

(3) What "Harness" Means

A harness is originally the tack put on a horse (reins, saddle, and the like). Here, it's used to mean a framework that keeps the AI from acting however it wants, and instead keeps it following a defined collaboration procedure.

2. Why Use It — What Gets Better

Giving one AI a big task all at once tends to cause these problems:

  • It forgets earlier decisions by the time it gets to later parts.
  • Two outputs that need to fit together (e.g., a web app's frontend and backend) end up mismatched.
  • Working alone, it can't catch its own mistakes.

Team Big Five Harness solves this the way a human team works.

In a human team In this harness
A lead sets the goal and divides the work team-lead (the leader AI)
Team members each do their assigned part contributor (multiple worker AIs)
One person cross-checks colleagues' work performance-monitor (the reviewer AI)
Shared goals get written on a whiteboard SMM.md (a shared notes file)
Repeating back an instruction: "OK, so you want me to...?" A message confirmation / restatement step

Key effect: the more a task requires multiple outputs to fit together, the bigger the payoff.

3. When to Use It — and When Not To

This is the most important judgment call. Using it unconditionally isn't automatically better.

Good fit (interdependent tasks)

  • "Design this API and wire up the frontend to it, all in one go"
  • "Gather multiple sources on this topic and synthesize/verify them into one"
  • "Build the worldbuilding and the plot for episodes 1-3 of a fantasy novel so they don't contradict each other"

Better to skip (simple or independent tasks)

  • Simple questions like "What's the English word for this?"
  • Short, one-off tasks
  • Multiple independent tasks that don't affect each other (just assign these separately — it's faster)

Analogy: Having 5 people roll a single piece of kimbap actually slows things down. A team shines when there are multiple stages that need to interlock, like a multi-course meal.

4. Installation (Getting Started)

This is the part non-developers get stuck on most, so we'll break it down step by step. Order: ① Set up Claude Code → ② Get the harness files → ③ Put them in your working folder → ④ Verify it works

Step ① Set Up Claude Code

Team Big Five Harness runs on top of Claude Code. If you don't have it yet, install Claude Code first.

  • Claude Code is a terminal-based tool and requires an Anthropic account (a Claude subscription).
  • The most accurate way to install it is to follow Anthropic's official documentation (search "install Claude Code").
  • Once installed, typing claude in your working folder will launch Claude Code.

If this is your first time using a terminal: on Mac it's called "Terminal"; on Windows it's "PowerShell" or "Command Prompt" — either one is the black window.

Step ② Get the Harness Files

Get the files from the GitHub page (https://github.com/tobyilee/team-bigfive). Pick whichever of the two methods is more comfortable.

Method A — Download via button (easiest, recommended)

  1. Go to the GitHub address above.
  2. Click the green Code button.
  3. Click Download ZIP.
  4. Unzip the downloaded file.

Method B — Download via command Paste this into your terminal as-is:

git clone https://github.com/tobyilee/team-bigfive.git

Step ③ Put Them Into Your Working Folder

This is the key step. The folder and file that act as the harness's "brain" are .claudeandCLAUDE.md.

  1. Inside the downloaded files, find the .claudefolder and theCLAUDE.md file.
    • (Folders starting with . are hidden by default and may not be visible. Turn on "Show hidden files" in your file explorer.)
  2. Copy both of these entirely into the actual project folder you'll be working in.

What your working folder looks like after copying:

my-project-folder/
├── .claude/          ← the harness's role/procedure definitions (copied in)
├── CLAUDE.md         ← the harness's instructions file (copied in)
└── (your working files...)

Step ④ Verify It Works

  1. Open a terminal in that folder and type claude to launch Claude Code.
  2. Try typing one of the trigger phrases from section 4 ("Usage Examples") below.
  3. If the AI doesn't just answer on its own, but instead does something like "form a team / divide into roles," it's working correctly.

The key takeaway: installation = putting the .claude folder and CLAUDE.md file into your working folder. That's it.

5. How to Use It — the "Trigger Phrase" Is Everything

Once installed, you don't need any complicated commands. Ask as you normally would, in your own words, and just include one "magic word" below, and team mode turns on.

Trigger words (any one of these is enough)

  • "as a team"
  • "agent team"
  • "team big five"
  • "high-performing team"
  • "with multiple agents"
  • "handle this collaboratively"

The formula for building a sentence

[what you want done]+[trigger word]+please

That's the whole thing.

Note: since Claude interprets natural language, you can phrase this in whatever language you're working in — the examples above are just illustrative English equivalents of the concept.

6. Four Real Usage Examples

Examples any non-developer can type in directly.

Example 1 — Building a Website (Development Task)

Build a company landing page and a contact form, and make sure they're wired together correctly, as a team.

One AI handles the UI, another handles the form logic, and the reviewer AI checks that "the two actually fit together."

Example 2 — Research and Reporting (Non-Development Task)

Research 2025 EV market trends from multiple sources and synthesize them into one report, using an agent team.

Multiple worker AIs split up the research → the reviewer AI catches contradictions between sources → the lead AI merges everything into a single report.

Example 3 — Writing and Planning (Creative Task)

Build out the worldbuilding and the plot for episodes 1-3 of a fantasy novel with team big five, so nothing contradicts.

The person handling the world and the person handling the plot both look at the shared notes (SMM) together to avoid conflicts.

Example 4 — Analysis + Verification (Business Task)

Analyze this spreadsheet of sales data, and cross-verify the analysis is correct, using a high-performing team.

Analysis and verification are handled by separate roles, catching errors that a single person working alone would miss.

7. What Happens While It's Running (For Peace of Mind)

So you don't panic when unfamiliar terms scroll by on screen, here's the flow ahead of time.

  1. Setup — The lead AI writes the goal and "definition of done" into the shared notes file (SMM.md).
  2. Team formation — It creates several worker AIs plus one reviewer AI.
  3. Collaborative execution — Workers reference the shared notes as they work, and confirm with each other when handing off results ("Just to confirm, you meant it this way, right?").
  4. Review — The reviewer AI cross-checks whether the outputs actually fit together.
  5. Synthesis — The lead AI merges everything into a single final version.

Output is typically organized into a folder called _workspace/final/.

8. Frequently Asked Questions

Q. Does this work in the regular Claude app (web/mobile) too?No. This harness isexclusive to Claude Code (the terminal tool). It doesn't work in the web or mobile Claude app.

Q. I can't see the .claude folder.Folders starting with . are hidden by default. Turn on "Show hidden items" in your file explorer settings. (Mac Finder: Cmd + Shift + .)Q. I included the trigger word, but it's just answering normally.Check three things: ① is .claude and CLAUDE.md actually insideyour current working folder, ② did you launch Claude Code from that folder, ③ is the trigger word spelled correctly.

Q. Wouldn't it be better to always use team mode?No. Using a team for simple tasks is actually slower and less efficient. Use it only for "complex tasks whose outputs need to fit together" (see Section 3).Q. What if I don't like the result? Just continue the conversation and ask for revisions in natural language. The harness is designed to incorporate feedback and improve after each task.

9. One-Page Cheat Sheet

[Install]
1) Install Claude Code
2) Download the ZIP from GitHub (Code → Download ZIP)
3) Copy the .claude folder + CLAUDE.md into your working folder
4) Run 'claude' from that folder

[Use]
"(what you want done) + (trigger word) + please"

[Trigger words]
as a team / agent team / team big five / high-performing team / with multiple agents / handle this collaboratively

[Use it for] Complex tasks whose outputs need to interlock
[Skip it for] Simple questions, short one-off tasks, independent tasks

Theoretical Background (For Further Reading)

This harness is grounded in the classic Team Science paper Salas, Sims & Burke (2005), "Is there a 'Big Five' in Teamwork?" Its core claim: good team performance comes from five behaviors (team leadership, mutual performance monitoring, backup behavior, adaptability, and team orientation) resting on three underlying foundations (shared mental models, closed-loop communication, and mutual trust). A detailed explanation can be found in docs/TEAM-SCIENCE.md in the source repository.