Living system · versioned · verified against current tools
The Engineering Workflow Library
How modern AI-native engineers actually work. Not tutorials — working practice: each workflow prescribes one good way, with real commands and config, the artifacts it produces, the pitfalls, and a dated changelog.
A workflow is a documented way of working with AI tooling in real engineering — the practice, not the product feature. (Claude Code also ships a feature named "workflows"; the Claude Code workflow covers it inside the practice.) Guides explain topics; workflows prescribe how to work, and every one produces something you can adopt this week.
The workflows
Working with Claude Code
Run a repo where the agent does the typing and you keep the judgment — CLAUDE.md, skills, hooks, subagents, worktrees, and the new dynamic workflows feature.
Shipping an Agent SDK agent
Build and ship a Claude Agent SDK agent end-to-end — from first query() to permissions, evals, and production deployment.
MCP integration
Model capabilities as MCP servers, test them, and ship them — the how-we-work layer on top of the MCP server guide.
Context engineering in the repo
Treat context as a budget: structure repos, docs, and memory so agents act well — practice layer of the context engineering guide.
Reviewing AI-written code
Diff discipline and verification steps that keep AI-generated changes honest — trust the process, not the model.
Eval-first development
Evals as the new tests: build the harness before the feature — practice layer of the LLM evaluation guide.
The AI quality stack
How you know the LLM is delivering: nine layered gates from spec-first to red-team, one runnable check per layer.
Spec-first development
Brainstorm → spec → plan → implement — deterministic gates and human checkpoints before any code lands.
Red-teaming LLM systems
Attack your own agent before someone else does — adversarial suites, cadence, and the triage loop back into guardrails.
The security self-audit
Audit your own LLM app against a framework you didn't write, with an agent that never saw the code — checklist frozen first, findings to file:line.
Validating the validator
The first check you write for a trusted artifact is itself unchecked — fixture a known-bad and a known-good case before you act on a single one of its verdicts.
Parallel agents, one working tree
Disjoint file lists are not enough — a bulk sweep and a targeted edit both own the whole set. Partition by operation, and put the baseline method in the brief, not just the ban.
Deciding what to build
When agents implement faster than you decide, the backlog is the bottleneck — two classification axes instead of one, a weekly floor rather than a maintenance ceiling, and the rules we measured and threw away.
Auditing an inherited citation
A "same as above" citation is a pointer, not a value. Correcting the row it points to silently rewrites every row beneath it — the sweep to run before you fix one link in the chain.
Kept current, visibly
A stale workflow is worse than none. Every page carries a last verified date — the day its commands were run against current tool versions — and a dated changelog of what changed and why. When a tool release invalidates a step, the page gets updated or visibly flagged, never left to rot silently. That maintenance discipline is what the membership funds.
Adopt the workflow, then learn the system underneath it.
The workflows are free and public. The aiArch curriculum teaches the engineering underneath them — agents, evals, context, and production architecture — on a platform that is itself a production agentic system. The build is the curriculum.
See how aiArch helps senior engineers become AI-native, or compare Professional Membership pricing.
Free sample — no signup · every claim cited · full curriculum is waitlist-only