AI-Agents

Ready-made expertise and multi-agent runs for Claude Code and Codex.

Out of the box, Claude Code works as one assistant in one chat. This repository adds three things:

• Skills: instructions the assistant loads by itself when your task needs them, for example a code-review checklist.
• Subagents: specialists, such as a security reviewer, that work in a separate context and report back.
• Saved workflows: multi-step runs, such as "review my diff with 4 reviewers, then try to disprove each finding". The rules of each run are enforced in code, so an agent cannot skip a step.

Install it as a Claude Code plugin in three commands, or clone it and link it.

What you get	How many	Where
Skills: engineering, AI, data, QA, ops, docs, research, applied theory	140+	skills/universal/
Subagents with explicit boundaries	85	agents/claude/, agents/codex/
Team recipes	8	agents/teams/
Saved workflows with guarantees in code	8	agents/workflows/
Opt-in hooks: git safety, lint-config guard, notifications	3	hooks/
Language and domain rules	14 folders	rules/

The exact, current list is in the generated catalog.

Quick start

Pick one install path. Claude Code only: use the plugin. Codex, or you want to edit skills: clone and link.

Claude Code plugin

/plugin marketplace add vasilyu1983/AI-Agents-public
/plugin install ai-agents@ai-agents
/reload-plugins

Then, in a repository with uncommitted changes:

/ai-agents:adversarial-review

Four reviewers read your diff. Two more agents (refuters) then try to disprove each finding. A finding is dropped only if both disprove it. The run changes no files.

Cost: each reviewer and refuter is a full agent session on your own plan. One run starts 4 reviewers plus 2 refuters per finding, and takes several minutes. Run it on a small diff first.

The run ends with a verdict: CLEAN (no findings), FINDINGS (each with file:line, severity and title), or INCOMPLETE (an agent failed; re-run). Workflows shows the full result.

Clone and link (Claude Code and Codex)

git clone https://github.com/vasilyu1983/AI-Agents-public.git ~/AI-Agents-public
cd ~/AI-Agents-public
bash scripts/distribution/sync-skills.sh

Then start a new session and type /adversarial-review. In Codex, which has no workflow commands, invoke the run-workflow skill and name the workflow; the skill makes Codex follow the workflow step by step.

Getting started covers both paths, subagent installs, checks, updates and removal.

Documentation

Guide	Read it to
Getting started	Install, check, update and uninstall
Concepts	Understand skills, subagents, teams, workflows, hooks and rules, and how they connect
Workflows	Run each of the 8 workflows: arguments, stages, verdicts, cost
Skills	Find, use and read skills
Agents and teams	Use, brief and install subagents and teams
Hooks and safety	Install the hooks; what the git guard blocks; recovery
Codex	What differs in Codex
Customizing	Write your own skill, subagent, team, workflow or rule
Troubleshooting	Fix a symptom
Catalog	Every skill, subagent, team, workflow and board, generated from the files

The workflows

Workflow	What it does	Edits files?
adversarial-review	Four reviewers on a diff, then two refuters per finding	No
review-fix-loop	Review and fix in rounds, or build against frozen acceptance checks	Yes, in scope
source-check	Checks every quote, link, figure and attribution in a document against its source, then fixes confirmed defects	Yes, the target
feature-delivery	Plan, test first, implement, review. Stops for your approval after the plan.	Yes, the plan's files
build-mvp	Cuts a PRD into thin slices and builds each approved slice; a fresh judge checks every slice	Yes, each slice's files
epic-delivery	Delivers an epic as file-owning tasks in dependency waves, then one integration review	Yes, each task's files
expert-board	A board of specialists on one decision: blind memos, optional debate, weighted synthesis	No
marketing-campaign	Research, one positioning lock, channel drafts, review against the lock	No

No workflow commits, pushes, opens a pull request or approves anything. You review and commit.

Why it works this way

These are design choices. The workflow tests in agents/workflows/test-*.mjs check the ones enforced in code.

• Guarantees live in code, not in prompts. Fresh agents each round, a two-refuter quorum, frozen acceptance checks, round caps and plateau stops are enforced by the workflow runtime. A prompt that imitates a workflow has none of them.
• The tree is checked, not trusted. Writing workflows snapshot the working tree after every step. A write outside the plan, a staged file, a moved HEAD or a touched instruction file stops the run.
• Every subagent states what it does not do. Reviewers cannot edit, so they report instead of quietly patching.
• Skills hold judgment, not data. A skill tells the agent what it would get wrong without it. Versions, prices and limits change, so skills look them up instead of hard-coding them.
• Small context. The runtime reads only each skill's short description until a task needs the skill. References load only when a step asks for them.

Layout

skills/universal/<name>/   SKILL.md + references/, scripts/, data/, assets/
agents/claude/             subagent definitions (source)
agents/codex/