Build AI Team

English | 简体中文

Version: v1.1.0

Choose one Agent or the smallest reliable persistent AI team—with clear ownership, role-specific model routing, and source-aware Skill discovery.

Optimized for diverse real-world scenarios, including software, product, design, research, data, content, migrations, and releases.

Give Build AI Team a real request. It works backward from the final result, keeps one Agent when that is enough, and adds a persistent role only when ownership, separation, or collaboration value is concrete.

Your request → one Agent or the smallest reliable team
→ role ownership and model configuration → your confirmation

Highlights

  • Right-sizes the team. Starts with one Agent and adds a persistent role only when its ownership, separation, or collaboration value is concrete.
  • Keeps narrow questions narrow. Existing-team requests return only an add, merge, replace, or keep delta; confirmed replacements create and verify the new task, update affected responsibilities, re-check states, and archive the old task last. Model and Skill questions answer only the concrete configuration requested.
  • Makes every role accountable. Each current role gets one clear responsibility, owned outcome, dependencies, and boundaries.
  • Routes models by responsibility. Recommends a default model and reasoning effort, explicit upgrade conditions, and practical Token-saving options when safe.
  • Finds relevant Skills without blurring status. Distinguishes locally available Skills from public leads; on request, it searches public sources and checks fit, provenance, maintenance signals, and permission risks.
  • Keeps the planning task useful. After confirmation, the current planning conversation can remain the substantive project lead while only the other necessary persistent tasks are created. Persistent tasks are the live team state; no AI_TEAM.md roster is created.
  • Keeps new projects independent. A shared folder, similar title, previous task, memory, or prior team plan does not make a request the same project. Existing project scope and tasks are reused only when the user clearly continues the same work.
  • Separates every consequential action. Planning, team creation, public Skill search, installation, login, private-data access, writes, uploads, publication, and deletion remain distinct steps.

Planning ≠ team creation ≠ Skill search ≠ installation ≠ external permission.

Quick start

For a concrete project, role, or work package, Codex may select Build AI Team automatically when you ask for a new team setup, a change to an existing Agent team, a model configuration, or project-specific Skill advice. It selects the narrowest matching mode; you do not need to add an activation phrase.

Simple bounded tasks, ordinary discussion or mention, explicit negation, and TOOLKIT_MODE=member work stay outside the planning workflow. Automatic selection still depends on the host’s Skill matching, so $build-ai-team remains the reliable explicit fallback.

After installation, start a fresh Codex task and enter a real request:

I want to design, build, and launch a paid web app with accounts and subscriptions. Decide whether one Agent or a team is appropriate and give me the complete working setup.

The first response is a plan. No team, Skill installation, account connection, private-data access, or external write happens without confirmation for that specific action.

Support and validation

  • English is the single runtime semantic source. Clarification questions and complete recommendations follow the user’s latest substantive request language; technical identifiers remain accurate.
  • The v1.1.0 five-file runtime passed independent static review with 0 major / 0 minor / 0 blocker.
  • The frozen four-mode baseline completed 9/9 technical runs and 10/10 responses. Its independent review isolated one ambiguity in the user-visible public-Skill status contract; observations from non-invocation and member-priority paths were kept as host observations rather than attributed to a Skill that did not run.
  • After the public-Skill output contract changed only skill-discovery.md, a fresh targeted two-case / three-context / five-response batch passed 3/3 technical runs and independent content review with 5 pass / 0 minor / 0 major / 0 blocker.
  • A later narrow SKILL.md change required a replacement confirmation to state its complete safe task sequence. A fresh saved-project two-turn dynamic test independently passed with 1 pass / 0 minor / 0 major / 0 blocker and verified: create and confirm the replacement, update and confirm the affected current assignment, re-check both task states, and archive the old task last without deletion.
  • This is combined differential evidence, not a claim that this runtime reran the complete nine-case set. Earlier v1.0.1 project-scoped Codex Desktop checks remain evidence for unchanged planning, ownership, permission, and team-creation behavior.
  • These checks are limited real-host evidence, not a statistical guarantee or cross-host validation.
  • Codex Desktop project-based usage is the primary target.
  • Entry instructions for ChatGPT Work and Claude are included, but those hosts have not been validated in practice for this release.
  • This project is not an official OpenAI, Codex, ChatGPT, or Anthropic project.
  • Compatibility, model availability, and recommendation quality may vary by host and version; no vendor endorsement or compatibility guarantee is implied.
  • This version continues the MIT license introduced by the public v1.0.0-rc.9 commit.

What it does not do automatically

  • create, replace, reassign, or archive Agents or persistent tasks;
  • install or run third-party Skills;
  • connect to accounts or business systems;
  • read private data;
  • write, upload, publish, or delete content;
  • guarantee that a third-party Skill is safe, current, or suitable;
  • guarantee one universally optimal team size.

Install in a Codex project

  1. Download or clone this repository.
  2. Copy the complete build-ai-team/ folder into the target project’s .agents/skills/ directory.
  3. Verify that the final path is:
<project>/.agents/skills/build-ai-team/SKILL.md
  1. Keep all five runtime files from the same version together:
build-ai-team/
├── SKILL.md
├── agents/
│   └── openai.yaml
└── references/
    ├── curated-skills.md
    ├── model-routing.md
    └── skill-discovery.md
  1. Start a fresh Codex task and describe a substantial project that may need team planning:
I want to design, build, and launch a paid web app with accounts and subscriptions. Decide whether one Agent or a team is appropriate.

If the host does not select the Skill automatically, use the explicit fallback:

$build-ai-team I want to design, build, and launch a paid web app with accounts and subscriptions.

Do not copy only SKILL.md, and do not place the whole repository root inside .agents/skills/. When updating, replace all five runtime files together; do not mix files from different versions.

If the Skill does not appear, check the directory hierarchy, the exact build-ai-team folder name, and whether all five files are present before trying other fixes.

Update or remove

  • Update: replace the complete build-ai-team/ folder with all five runtime files from one newer version, verify SHA256SUMS.txt, and start a fresh Codex task. Do not mix files from different versions.
  • Remove: delete only <project>/.agents/skills/build-ai-team/, then start a fresh Codex task. Do not delete the parent .agents/skills/ directory because it may contain other Skills.

Removing the Skill stops it from being loaded in future fresh tasks. It does not undo teams, files, installations, account access, or external actions that were separately confirmed earlier.

Example requests

I only need a clickable prototype from this finalized PRD. Decide whether one Agent is enough; no development is needed.

Plan a CRM migration with independent read-only verification. Do not connect to the CRM yet.

我想做一个能上架 App Store 的记账 App。请判断需要一个 Agent 还是多个 Agent，并给我完整工作安排。

Our current team has a backend owner and a read-only verifier. Should I add another member for the migration, or keep the team unchanged?

Which model and reasoning effort should I use for a data-migration verifier?

Which locally available Skills fit an iOS release-readiness review?

How to read a recommendation

  • Required: belongs to the current team after confirmation.
  • Required if…: the work package appears only when the stated observable condition becomes true; it is not created now.
  • Optional: may improve quality, speed, or coverage but is not needed to complete the current goal and is not created by default.
  • Owner: the one role accountable for a particular work package or deliverable.
  • project lead: an existing substantive role that additionally coordinates plan, dependencies, and final handoff status; it does not gain ownership of other roles’ professional deliverables.
  • independent reviewer: stays separate from production and does not edit the work being reviewed.

Local Skills and public search

Initial recommendations use only the current host's exposed local-Skill metadata and do not show unchecked public leads. Live public search starts only when the user asks for it, searches external sources rather than repeating the local list, and keeps an existing plan's primary next step visible.

A recommendation is not installation. Installation is not permission to run scripts, log in, read private data, write, upload, publish, or delete. Review third-party Skills and their licenses before use.

Third-party project names and links are references only. Inclusion does not mean endorsement, affiliation, current compatibility, or security certification, and those projects remain subject to their own licenses.

Checksums

SHA256SUMS.txt covers the five runtime files. Run the platform-appropriate SHA-256 check after downloading, and compare the results before use.

Repository documents

Relationship to Agent Team v1.0

Build AI Team is a separate Skill product. It does not replace the earlier five-file Agent Team v1.0 toolkit or its immutable release.

License and official source

This version is provided under the MIT License. The license change began with the public main commit containing v1.0.0-rc.9; that version and later MIT-licensed releases may be used, copied, modified, merged, published, distributed, sublicensed, and sold subject to the license terms.

The earlier v1.0-beta7 public source snapshot remains governed by the custom non-open-source license included in that exact commit. The MIT change is not retroactive.

The only official repository is itsedizeng/build-ai-team.