DEV Community

Alair Joao Tavares
Alair Joao Tavares

Posted on • Originally published at activi.dev

AI-Augmented Developer: How to Turn Your AI Agent into a Disciplined Engineer

The Problem Nobody Wants to Admit

If you build software with AI, you know the loop: you describe a feature, and the agent jumps straight into writing code before understanding the problem. It skips the tests, invents context, and three hours later, you're reviewing an 800-line diff only to find half of it doesn't do what you asked.

This isn't just the model's fault. It's the absence of process. Senior engineers don't open the editor first—they specify, plan, test, and review. The AI agent is missing that exact same methodology.

That's exactly what the AI-Augmented Developer framework delivers.

What is AI-Augmented Developer?

AI-Augmented Developer (aiadev) is a complete workflow framework for coding agents. It installs a set of composable skills and bootstrap instructions that ensure the agent uses them automatically—so you don't have to remember to prompt it correctly.

The philosophy is blunt:

  • Spec-first: No code is written without an approved specification.
  • Test-first: RED-GREEN-REFACTOR is a contract, not a suggestion.
  • Evidence over claims: The agent must verify its work before declaring success.
  • Simplicity as a primary goal: YAGNI (You Aren't Gonna Need It) and DRY (Don't Repeat Yourself) are laws, not tips.

The standard pipeline enforces an eight-stage engineering lifecycle:

specify → clarify → plan → tasks → implement
                                       │
                          test-driven-development (per task)
                          systematic-debugging (on failures)
                          checklist (security, perf, a11y, i18n…)
                                       ↓
                               analyze → requesting-code-review → finishing-a-branch
Enter fullscreen mode Exit fullscreen mode

Each stage is a "skill" that fires autonomously at the right moment. The agent doesn't skip steps or hallucinate context; it explicitly shows you the spec before writing its first test.

The Constitution: Seven Non-Negotiable Rules

The heart of the framework is constitution.md, containing seven principles every technical decision must honor:

  1. Spec-first — No approved spec, no code.
  2. Test-first — A failing test must exist before implementation.
  3. Simplicity — Build the simplest thing that works.
  4. Evidence over claims — Run it, prove it, show it.
  5. Provider pattern — Keep external dependencies behind interfaces.
  6. Privacy by design — Sensitive data never leaks to LLMs.
  7. Attribution — Credit every derivative work.

Every plan produced by the plan skill carries a Constitution Check table. If the agent breaks an article, the violation goes into a Complexity Tracking section with a required justification. Without this discipline, the framework refuses to move forward.

Rapid Evolution: 9 Releases in 48 Hours

Between April 14 and 15, 2026, the project rapidly shipped from v0.3 to v0.11. Each release tackled a real friction point for daily users.

v0.3: Interactive aiadev install

A Python CLI replaced ad-hoc scripts. A single command now renders a preset (substituting variables and placing files) into your project, complete with --dry-run, --uninstall, and drift detection against hand edits.

v0.4: Cursor Support

The first platform handler beyond Claude Code. It introduced full end-to-end round-trip support, fully documented.

v0.5: Codex, OpenCode, and Gemini Integration

Three more platforms arrived in a single release. The five major AI development tools are now officially covered: Claude Code, Cursor, Codex, OpenCode, and Gemini CLI. Each handler is a self-contained ~30-line module with 100% test coverage.

v0.6: User-Level Scope

Running --scope user installs skills once per machine under ~/.<platform>/skills/. Every project on your workstation inherits the same catalog with no repeated setup, while files containing project-specific variables (CLAUDE.md, constitution.md) stay local to the project.

v0.7: PyPI Distribution

You can now simply run pip install aiadev. The wheel bundles constitution.md, templates/, schemas/, skills/, presets/, and agents/—no repo clone required. It is published via OIDC trusted publishing, with no tokens stored in the repository.

v0.8: Extension System

aiadev extension add <git-url> allows anyone to ship third-party preset catalogs. Community catalogs, private corporate presets, and experimental builds are now supported. Built-ins win on name collisions, notifying users with a yellow warning when an extension is shadowed.

v0.9: Full Pipeline Installation and aiadev sync

The biggest paradigm shift. The install command now equips a project with the entire pipeline at once: 14 slash commands, 3 agents, 5 coding rules, and the full catalog of generic skills. The new aiadev sync command pulls framework updates into existing projects and regenerates an <!-- aiadev:auto-stack --> block inside CLAUDE.md based on project introspection (package.json, pyproject.toml, Cargo.toml, go.mod, docker-compose, etc.).

v0.10: Namespacing and Sequential Specs

Slash commands gained a structured namespace (e.g., /aiadev:specify, /aiadev:plan). Specs transitioned from a feature-<slug>/ scheme to zero-padded sequential IDs (specs/0001-<slug>/). Additionally, aiadev init --language pt-BR configures the entire pipeline to operate in the user's chosen language.

v0.11: Universal Model Context Protocol (MCP)

The Model Context Protocol is now a first-class citizen. You declare servers once in mcps.yaml, and aiadev install translates them to each platform's native format:

  • Claude Code: .mcp.json
  • Cursor: .cursor/mcp.json
  • Gemini CLI: .gemini/settings.json
  • Codex: .codex/config.toml
  • OpenCode: opencode.json

MCP stops being repetitive boilerplate and becomes a simple configuration detail.

Why This Framework Matters

Nine releases in 48 hours, each solving a concrete pain point without regressions or breaking changes. That is the framework successfully applying itself to itself: specs live under specs/, plans were generated by the plan skill, and commits follow the feat(<area>): T<N> <title> pattern enforced by the tasks skill.

For anyone building with AI, aiadev solves four critical problems at once:

Developer Pain Point Framework Solution
Agent codes without understanding the problem The specify skill forces specification first.
Code lacks tests, or tests are an afterthought The test-driven-development skill enforces RED-GREEN-REFACTOR.
Forgotten decisions and scope drift The analyze skill reports divergence between the spec, plan, tasks, and code.
Tedious manual setup per project aiadev install + --scope user + extensions automate everything.

Best of all: you don't need to manually invoke anything. Skills fire autonomously at the right moment across all five supported platforms, backed by a single MCP server declaration, ultimately delivering a clean PR.

Getting Started

# 1. Install the CLI
pip install aiadev

# 2. Enter a project and install the preset that fits
cd your-project
aiadev install --preset lean              # Generic pipeline
aiadev install --preset django-drf-react  # Full-stack web
aiadev install --preset mobile-ops        # Cloud Run + Expo

# 3. Pick your preferred platform (default is claude-code)
aiadev install --preset lean --platform cursor

# 4. Working in another language? Initialize with a language flag
aiadev init --language en

# 5. Verify your installation
aiadev doctor
Enter fullscreen mode Exit fullscreen mode

Start a fresh session, ask for a feature in natural language, and watch the agent instinctively reach for specify before writing a single line of code.

Who is it For?

  • Solo developers who want maximum productivity without sacrificing code quality.
  • Teams that need a consistent, unified process across multiple human contributors and AI agents.
  • Enterprises looking to standardize AI usage without vendor lock-in to a single platform.
  • Platform engineering teams maintaining internal tooling—the extensions system perfectly handles corporate distribution.

What's Next?

With the core pipeline complete, five major platforms wired, and MCP fully integrated, the foundation is solid. The natural next steps include themed presets (data engineering, machine learning, infrastructure), opt-in telemetry to determine which skills generate the most value, and specialized agents to automatically validate specs.

But the core promise is already here: the framework is complete enough for daily use, disciplined enough for serious production projects, and open enough for the community to evolve.


Top comments (0)