AI-assisted development workflow is the set of practices a software team builds around AI coding agents (Claude Code, GitHub Copilot, Cursor, Codex, and similar tools) so the whole software development lifecycle benefits, not just the moment of typing code. It matters because DORA's 2025 research found that higher AI adoption raises both delivery throughput and delivery instability at the same time: the tools amplify whatever engineering discipline already exists, for better or worse, instead of fixing weak practices on their own. The insight that separates teams who compound gains from teams who compound debt is sequencing: agree on intent before generating code, keep that intent and your team's conventions in durable files an agent actually reads, and treat every agent-authored change as something a specific human still owns, reviews, and can trace back to a decision. Get that sequencing right and an agent becomes a fast, tireless collaborator; skip it and the same agent quietly produces duplicated, ungrounded, sometimes insecure code faster than any reviewer can keep up with.
What This Cheat Sheet Covers
This topic spans 15 focused tables and 102 indexed concepts. Below is a complete table-by-table outline of this topic, spanning foundational concepts through advanced details.
A jump-to index of every table row in this cheat sheet.
An interactive map of every table and concept in this topic.
Table 1: Workflow Philosophies and Mental Models
Before adopting any specific tool or ritual, it helps to name the mental models practitioners actually use to reason about working with AI agents — these frames explain why the more concrete practices in later tables exist.
| Approach | Example | Description |
|---|---|---|
Prompt-first workflow | Ask an agent to "add a login page" and iterate on whatever it returns | Fast for small, contained tasks • requirements and edge cases live only in the prompt, so context is lost between sessions and reviewers. |
Write a structured spec, then generate a plan, tasks, and implementation from it | Captures intent in a shared artifact first; the agent generates against that artifact instead of inferring meaning from scattered prompts. | |
"Make a tool that scrapes a page and emails me a daily summary," with no line-by-line review | Natural-language-only prompting where the output is trusted because it "just works"; fine for throwaway scripts, risky for production systems. | |
Deliberately curating which files, rules, and tool outputs an agent's context window contains before a task starts | Curating what the model sees so it gets the right information, not the most information; too much context degrades results as much as too little. |