Agentic software delivery,
without the chaos.
A reusable Claude Code-native runtime for orchestrating planning, implementation, UI/UX polish, QA, documentation, and self-maintaining project memory across client projects.
- Open source
- MIT license
- Zero deps
- Single file
AI coding falls apart at scale.
One-shot prompts produce demos. Real client work is the opposite: long-lived projects where context drifts between sessions, decisions get forgotten, tasks slip into ambient memory, and design intent quietly erodes across dozens of small edits.
Without explicit lanes, contracts, and audits, agentic coding collapses into a generic-looking mess — plausible code that nobody can hand off, UIs that all start to look the same, and QA that exists only as "looks good to me."
Seven pieces. One contract.
A small, opinionated runtime. Every piece has a single owner and a single responsibility.
-
Central
Orchestrator
Sole authority over task state, memory writes, and final completion. Reads required files on every request, classifies the work, delegates to specialized workers, and closes out the task once QA and UI audits pass.
-
Workers
Specialized lanes: planner, backend, frontend, ui-ux, qa, docs, devops. Each stays in its own.
-
Memory
Living project, progress, architecture, decisions, conventions. Self-maintaining across sessions.
-
Contracts
API, database, frontend, UI direction, integrations. Workers read before acting, write as they go.
-
Tasks
Atomic files moving through queued, active, review, blocked, completed. State is on disk, not in memory.
-
Reports
QA, UI audits, bug records. Every delivered feature leaves an evidence trail you can hand off.
-
Impeccable UI/UX Skill
Vendored design audit. Catches generic AI-template patterns before the UI ships.
Every request follows the same path.
Lightweight variants exist — typo fixes don't need a planner — but the canonical pipeline is fixed and visible.
-
01
Request
User intent enters. Orchestrator reads required files.
-
02
Plan
Planner decomposes into atomic task files. Each names an owner.
-
03
Contract
UI direction, API shape, schema. Worker reads before implementing.
-
04
Implement
Frontend, backend, or devops works strictly inside its lane.
-
05
Audit
UI/UX runs Impeccable on the result. Pass or revise.
-
06
QA
Behavior validated against the task. Recommendation, not authority.
-
07
Document
Docs worker updates README, changelog, public surfaces — if changed.
-
08
Update Memory
Orchestrator appends progress, updates architecture if needed, closes the task.
Built for client delivery.
Designed by a freelancer, for freelancers. Solo developers shipping serious work across many projects.
-
Consistent delivery
Every project runs the same pipeline. No bespoke setup per client.
-
Less context drift
Memory survives sessions. Decisions don't get re-litigated.
-
Better UI output
Impeccable audits every frontend change. No generic templates.
-
Safer handoffs
Tasks, contracts, and reports leave a trail the client can read.
-
Reusable across projects
Stack-agnostic. Clone, initialize, ship.
What's inside.
A small, opinionated tree. Bridge files at the root for tools that need them; the truth lives in .claude/.
CLAUDE.md
PRODUCT.md
DESIGN.md
index.html
.claude/
agents/ # orchestrator, planner, backend, frontend,
# ui-ux, qa, docs, devops
commands/ # /initialize-project, /plan-feature, ...
memory/ # project, progress, architecture, decisions
contracts/ # api, database, frontend, ui-direction
tasks/ # queued, active, review, blocked, completed
reports/ # qa, ui, bugs
runtime/ # workflow, execution-rules, execution-state
skills/impeccable/ # vendored design audit
Clone the runtime.
Start your next project.
Open in Claude Code, run /initialize-project, and ship.
$ git clone https://github.com/webDevYuri/claude-engineering-runtime.git
$ cd claude-engineering-runtime
# In Claude Code:
/initialize-project