AI-SDLC

// GitHub

// Problem

Developers were confused about how to actually work with AI agents. Having done every job on a product team—UX design, UI design, QA, product management, product leadership, development—I understood the full loop. But most people were skipping critical planning phases, jumping straight to code, and wondering why their AI-assisted development felt chaotic. The tooling at the time either locked you into specific AI providers or tried to do too much magic behind the scenes. What developers needed wasn't another AI wrapper—they needed a structured workflow that made the invisible visible.

// Solution

A CLI tool that enforces a disciplined 8-step development process: idea → PRD → PRD-plus → architecture → system patterns → tasks → tasks-plus → tests. Each step generates a markdown file that becomes context for the next. The key insight: everything is just a prompt. The tool doesn't call AI APIs directly—it generates prompts you can paste into any AI tool (Claude, ChatGPT, Cursor, whatever you prefer). This makes the workflow transparent, version-controlled, and completely AI-agnostic.

// What I Built

A Python CLI distributed via PyPI that creates a structured directory for feature development. Running 'aisdlc init' sets up prompts/, doing/, and done/ directories with 8 prompt templates. 'aisdlc new "Feature idea"' creates a feature folder and your first markdown file. 'aisdlc next' reads your previous output, merges it with the next step's prompt template (replacing <prev_step> placeholders), and generates a ready-to-paste prompt. You copy that to your AI of choice, save the response, and repeat. The workflow has two modes: Chat mode (steps 1-5) where you iterate with AI to refine ideas and architecture, and Agent mode (steps 7-8) where the prompts are designed for more autonomous task execution. Everything stays in git—your planning artifacts, prompts, and outputs are all diffable markdown.

// Technologies

// Lessons Learned

• 01Everything is just a prompt. Write that on your mirror. The entire AI development workflow—requirements gathering, architecture design, task breakdown, test generation—is just prompts with different system instructions. Once you internalize this, AI tooling becomes demystified. You're not using magic; you're orchestrating context.
• 02Harnesses (what we now call them in 2025) are just prompt orchestration with managed context. AI-SDLC is essentially a linear harness that passes output from one step as input to the next. The <prev_step> placeholder is primitive but effective. Modern tools like Claude Code do this more elegantly, but the core pattern is identical.
• 03Forcing structure prevents the 'vibe coding' trap. When you can jump straight to code, you will. The 8-step gating mechanism makes you articulate the problem, define requirements, and plan architecture before writing a single line. This friction is a feature, not a bug.
• 04Tool-agnostic beats tool-specific. By generating prompts instead of calling APIs, AI-SDLC works with whatever AI is best at the moment. When Claude 3 dropped, users could switch immediately. When GPT-4 Turbo improved, same thing. No vendor lock-in, no API key management, no rate limiting concerns.
• 05Early inspiration from AIDER's architect/code mode pattern shaped this project. The idea that you could separate 'thinking about code' from 'writing code' into distinct phases with different prompts was revelatory. AI-SDLC extends that to the full product development lifecycle.