Bootstrap and maintain an Atlas workspace in your repository, then gate AI tool development with evidence-based reviews across intake, plan, review, gate, and postmortem modes.
Coding agents are only as good as the context your repository gives them.
Atlas makes that context structured, durable, and machine-checked. One command scaffolds the shared instructions, repo memory, artifact paths, and managed skills agents need — plus two kinds of gates: structural gates that keep the workspace sound and process gates that turn reviews into recorded verdicts.
The payoff compounds over time. You keep working normally, while Atlas turns useful project context into documentation that grows with the repo and makes each next agent run more effective.
Atlas creates the deterministic structure, then hands the rest to your local agent so it can inspect the repository and finish the setup with project-specific context.
Start Here
npx --yes @blazity-atlas/core@latest init
Atlas first asks where the workspace should live — default .ai, any repo-relative path works — then previews the files it wants to write, asks for confirmation in an interactive terminal, creates the workspace and agent entrypoints, and prints the next prompt to give your coding agent. If it detects an agent CLI on your machine (claude, codex, cursor-agent), it can offer to launch it with that prompt directly.
Non-interactive setups pass --root <dir> instead. Custom roots are discovered through a one-line .atlas pointer file at the repo root; the default .ai needs no pointer, and every example below uses it.
Claude Code users can start from the Atlas marketplace instead:
Both paths use the same published package, @blazity-atlas/core. The Claude Code plugin exposes the managed skills; the CLI still owns the deterministic file structure.
What Atlas Adds
Atlas keeps AI-facing documentation small, explicit, and owned by the repository:
.ai/config.json is the source of truth for artifact locations. If your repo already has useful docs, Atlas can map conventional paths into the .ai/ workspace instead of inventing another documentation system.
How Setup Continues
The first command only writes the shared structure. The printed prompt tells your agent to read .ai/skills/atlas-setup/SKILL.md and follow it. The atlas-setup skill then:
inspects the repository before asking questions;
lets the agent recommend a template after reading the project — the five templates differ only in path aliases (which conventional docs folders get migrated), so the choice is low-stakes, agent-proposed, and refinable later;
fills AGENTS.md, project vocabulary, and stable memory files;
keeps Claude, Codex, Cursor, and similar agents pointed at one shared workspace;
leaves plans, decisions, research, and results in predictable locations.
This keeps the human flow simple: install Atlas once, then let the local agent adapt it to the actual repository. As work continues, plans, decisions, research, and lessons accumulate where agents can find them instead of disappearing into chat history.
Reviews That Leave a Verdict
Atlas manages a second skill: atlas-review. It walks a change through five modes — Intake, Plan, Review, Gate, and Postmortem — and writes its verdict (pass, conditional pass, or fail) as an artifact into the results path, where the next agent run can find it.
Claude Code users run /atlas:atlas-review. Any other agent gets the same behavior from one instruction: "read .ai/skills/atlas-review/SKILL.md".
Why Atlas Exists
AI can generate code quickly. That is no longer the hard part.
The hard part is keeping the output understandable, reviewable, secure, and aligned with the system your team actually needs to run.
Atlas is built for teams that want AI acceleration without giving up ownership. It brings agents into the delivery process through explicit rules, traceable artifacts, structural and process gates, and human review at the points where judgment matters.
Make a codebase agent-ready by scaffolding AGENTS.md, ARCHITECTURE.md, and docs/ structure following progressive disclosure patterns. Creates CLAUDE.md as a symlink for Claude Code compatibility.
Core AIWG utilities for context regeneration, workspace management, development kit, and @-mention traceability. Essential foundation for other AIWG plugins.
