Give Claude Code an engineering process.
An open-source harness that builds a full engineering process into Claude Code: researched plans, enforced standards, agent-reviewed and tested code, and documentation that stays true to the codebase. What ships is engineered, not generated.
Fast code is easy. A maintainable product is not.
Claude Code is astonishing. Features appear in hours. You ship faster than you ever have.
Decisions nobody wrote down. Inconsistent patterns. Architecture that exists only in old chat sessions. Every new feature fights the last one.
Not a better prompt — a process. One the AI actually follows: plan, build, review, test, document. Every time.
From idea to reviewed, tested, documented code
Four steps. Each is one command in Claude Code. You approve the decisions; the agents do the work.
Give your project a single source of truth
One command reads your codebase and builds an organized documentation library from it: architecture, coding standards, requirements, and research, each in its own place. This is much more than a notes file: it's the structured knowledge that every later step reads, builds against, and keeps up to date.
One command: /uc:migrate · Read more →
Plan like a tech lead
Describe what you want to build. Ultra Claude questions the scope, researches your existing code, and turns the idea into a step-by-step plan. You review and approve it before anything gets built.
One command: /uc:feature-mode · Read more →
Build with a team, not a single bot
Every task in the plan gets its own small team: one agent writes the code, a second reviews it against your standards, and a third tests that it works, opening the app in a real browser for UI work. Nothing counts as done until all three agree.
One command: /uc:plan-execution · Read more →
Keep docs and code aligned over time
Whenever you like, Ultra Claude compares what the documentation says to what the code actually does and flags anything that no longer matches. The project stays understandable, to the agents and to you.
One command: /uc:doc-code-verification-mode · Read more →
Not a spec format. Not a framework. A process.
vs. raw Claude Code
Same tool underneath. Ultra Claude adds governance on top: review gates before code lands, standards enforced with citations, and an organized documentation library in your repo instead of knowledge scattered across chat history.
vs. spec-format tools
Specs here aren't input templates you hope the model respects. They're enforced during execution and audited afterward. Define → enforce → verify. The loop closes.
vs. multi-agent frameworks
No new runtime, no orchestration server, no SDK. It's skills and agents inside the Claude Code you already run. Delete it and your repo still works.
The full lifecycle, covered
Agent teams per task
Every task is written, reviewed, and tested by separate agents before it lands. The reviewer cites your standards; the tester runs the real thing.
Docs govern code
Architecture and requirements exist before code is written, and every plan is checked against them. When code and spec disagree, you fix the spec first, deliberately.
Standards that actually hold
Define coding standards once. The reviewer enforces them on every task with specific citations, and verification audits that they still hold months later.
Research once, reuse forever
Library docs and patterns are fetched once and committed as project documentation. Repeat lookups hit the cache instantly, at a fraction of the tokens of raw web search.
Survives sessions
Checkpoints save execution progress automatically. A session can die mid-plan and resume where it left off. Completed work is never redone or paid for twice.
Roadmaps for whole products
Too big for one plan? Roadmapping decomposes a product into a sequenced series of plans with dependency ordering. Each one feeds straight into feature planning.
Backlog management
Bugs, open questions, ideas, and tech debt are tracked in the repo alongside the code. Agents log what they find; nothing gets lost between sessions.
Knows your Claude Code limits
Execution rides each 5-hour window to 100%, parks itself on the limit, and a machine-wide sentinel wakes every session the moment the window resets — plus per-task budget tracking.
Debugging mode
Hypothesis-driven bug investigation: evidence gathered in parallel, the bug reproduced before it's fixed, and the fix validated by a dedicated tester.
Ready to build like a team?
10 minutes to your first governed feature. Two commands to install, one to set up. Everything else is a conversation.
# In any Claude Code session:
/plugin marketplace add duniecdawid/ultra-claude-code
/plugin install uc@ultra-claude
# Once per machine:
/uc:setup
# Once per project:
/uc:migrateThe questions everyone asks
Is this a framework?
No. Everything runs as skills and agents inside Claude Code itself, and the documentation it writes stays useful on its own. Uninstall it and your repo still works.
Does it work on existing codebases?
Yes. Initialization detects existing projects, reads the code, and derives the architecture docs and coding standards from what's already there. You review and correct them, then everything builds on that.
What does it cost?
Ultra Claude is free and open source. You pay only your normal Claude usage: no account, no subscription, no separate billing.
Do I need tmux?
No, it's optional. With tmux, agent teams appear in split panes so you can watch them work; without it, everything still runs the same way.
How is this different from just writing a CLAUDE.md?
A CLAUDE.md is a page of advice the model may or may not follow. Ultra Claude is an enforced lifecycle: plans are challenged before building, code is reviewed and tested against your standards on every task, and the docs are audited against the code afterward.
Make Ultra a collaboration platform for your organization
The Ultra Claude Dashboard is the dark counterpart to your local setup: a live, shared view of every Ultra Claude project (documentation, plans, backlog, and execution history), synced from your machine in real time. Anyone on your team follows along from a browser; no dev environment needed.
Watch plans execute, live
Plans appear on the board while they're still being planned and update as agent teams work through tasks: Kanban boards, timelines, and execution status in real time.
Docs your whole team can read
Rendered documentation with comments on any heading, automatic snapshots on every commit, and side-by-side diffs between any two points in time.
Workspaces for the whole org
Invite anyone to the workspace. PMs and stakeholders finally see the docs, plans, and backlog that used to be buried in developers' repos. No CLI, no dev environment.
One package: npm i -g ultraclaude-agent, then ultraclaude-agent setup connects your projects.