CONCEPT

AI-Assisted Development

Properties1
tagsai-coding, claude-code, spec-driven-development, subagents, verification, web-dev

2 min read

AI-Assisted Development

Building software by directing coding agents (e.g. Claude Code) — the workflow layer of how to develop with an LLM, distinct from building AI features into a product.

Definition

AI-assisted development is the practice and discipline of driving an AI coding agent effectively. Austin Marchese, citing Claude Code creator Boris Cherny and Andrej Karpathy, frames a repeatable workflow: spec-driven development (co-design a detailed spec/“docs” before building — a spec collapses many possible build paths to one), sub-agents (parallel agent sessions, each with isolated context/prompt/permissions, for independent perspectives and scale), verification (give the agent a feedback loop and a way to see its own output — claimed 2–3× quality — plus human-validation zones sized by cost of error), reusable skills (built from validated real work, with a “gotchas” section), and judgment about augmentation vs. automation (prefer augmentation; most automations add operational debt). Global behavior is steered via a CLAUDE.md file injected at the start of every prompt.

Consensus is emerging: this is new, fast-moving, authority-based practice (leaning on Cherny/Karpathy/Anthropic guidance), not settled or independently benchmarked. It sits above the wiki’s artifact-focused web-dev concepts, relating to Clean Code & Maintainability (discipline, verification) and CI/CD & Automation (DORA Elite) (quality gates).

Key Properties

  • Spec-first: design detailed specs/docs before building; plan mode assists but doesn’t replace them.
  • Sub-agents: parallel, context-isolated sessions for perspectives, scale, and speed.
  • Verify before build: feedback loops + tools to inspect output; human-validation zones by error cost.
  • Skills from real work (with “gotchas”); steer globally via CLAUDE.md.
  • Augmentation > automation — automate only after a “taste test” / 80-20 analysis.

Examples from Sources

ExampleSource
Launching 10 sub-agents to check 10,000+ domains (a task infeasible in one session)Paste This Into Claude, It'll Make You Build 10x Faster (6 Claude Code power phrases)
”Interview me” → agent asks the questions you didn’t know to ask, then writes a specPaste This Into Claude, It'll Make You Build 10x Faster (6 Claude Code power phrases)

In the Sources

SourceContext
Paste This Into Claude, It'll Make You Build 10x Faster (6 Claude Code power phrases)Six “power phrases” for Claude Code (spec, sub-agents, verify, skills, augment).

Concepts: Clean Code & Maintainability, CI/CD & Automation (DORA Elite), 8-Pass AI Writing Pipeline Entities: Claude Code, Claude, Anthropic, Austin Marchese