Beyond Vibe Coding: A Systematic and Customisable AI Software Development Process

If you use AI for coding, you get fast results, but potentially slow problems. At Open Community Experience 2026 (OCX26) in Brussels, Philip presented a practical view on using AI coding agents in professional software development: amplifying real software engineering instead of shutting down your brain and letting AI erode your software, while still gaining huge productivity gains through iterative task & context engineering, deliberate developer judgment, active agent steering, and a controlled, customisable workflow.

In case you missed it, the full recording is now available:

Why Vibe Coding Breaks Down

Modern AI coding assistants such as Copilot, Claude Code, Theia AI, Codex CLI, and many others, promise to turn every developer into a 10x engineer. And yes, the productivity gains are real. But just dropping a vague prompt on an agent and accepting whatever comes back (“vibe coding”) quickly collapses in professional software development.

The talk identifies three dimensions that determine where AI coding succeeds and where it falls apart:

• Task complexity: Simple to-do apps are easy. Avionics certification tools, niche domain logic, or anything with significant requirements gaps push the LLM into hallucination territory.
• Code maturity: Greenfield projects suit LLMs well. Brownfield systems with architectural constraints, historical decisions, and customer-specific quirks do not.
• Technology stack: Mainstream stacks have plenty of training data. Niche frameworks and proprietary technologies do not.

Many professional, long-lived products sit in the “hard mode” corner of at least some of the three dimensions. AI can still help massively, but only if you stop outsourcing the thinking.

AI Amplifies Real Software Engineering — It Doesn’t Replace It

A central message of the talk: AI is an amplifier. It amplifies your engineering discipline, your code quality, and your team culture. It also equally amplifies any sloppiness in your codebase. The worst thing you can do is shut down your brain and hope the AI figures it out; that’s the fastest way to erode a long-lived system.

As shown in the video, studies show a 30–40% productivity gain from AI coding on average, but teams often lose half of that cleaning up the slop committed the week before. Avoiding that slop alone effectively doubles the realised productivity gain. The takeaway is clear: keep the quality bar high, keep thinking, and treat AI as a tool that amplifies your judgment, not one that replaces it.

How Coding Agents Actually Work

To use AI agents effectively, you need a basic mental model of what’s happening under the hood. The agent is a thin layer over an LLM that:

1. Provides tools (read files, propose edits, run terminal commands, call MCP servers).
2. Manages a conversation that grows with every turn.
3. Sends everything, from system prompt, history, tool results, to project files, back to the LLM on every request.

Once you internalise that, the most important variable becomes obvious: the context window. Optimising AI coding means actively curating what ends up in that context: relevance, correctness, completeness, and trajectory.

Iterative Task & Context Engineering: A Controlled Workflow

Instead of one vague prompt followed by hopeful clicking, the talk proposes a structured, iterative cycle built on task engineering and context engineering, with the developer actively steering the agent at every step:

1. Define the context: Externalise your own thinking. What part of the codebase is relevant? What history, requirements, constraints, and intent does the agent need? Use incremental prompting to scope context, then extract a clean task definition.
2. Explore and refine: Treat each phase (scoping → solution design → implementation) as its own focused prompt with appropriate context, not a single ever-growing chat.
3. Review and decide: When the agent returns an imperfect result, don’t just say “try again.” Identify architectural blockers, redo the prompt with a better starting context, or split the task. Iterating inside the same polluted chat usually makes things worse.
4. Persist what’s worth keeping: Architectural decisions, reusable prompts, and project-level guidance go into persistent context (CLAUDE.md, custom instructions, README files, ADRs), but kept lean, like an index the agent can navigate.

Practical Techniques That Make a Real Difference

A few highlights from the talk’s practical recommendations:

• Task context as working memory: Maintain an explicit, editable task file collecting facts, constraints, and decisions across steps. Reuse it. Curate it. Distil it.
• Redo, don’t refine forever: When a session goes sideways, start a new one with a focused prompt instead of layering corrections on a poisoned context.
• Manage MCP and tool overhead: Every MCP server fills your context window by default. Stay below ~50% context usage; isolate heavy tools (e.g. Playwright) into dedicated sub-agents.
• Let agents off the leash, safely: Approving every terminal command becomes theatre after the fifth click. Use proper isolation (containers, git worktrees, network restrictions) so agents can work autonomously without putting your system at risk.
• Observe agent trajectories: Where the agent wanders unnecessarily tells you exactly which parts of your project context need improvement.

The Cultural Shift

For many developers, writing code was the process of finding a good solution. AI coding takes that activity away while still expecting us to arrive at good solutions, through prompting and structured delegation instead. That’s a real transformation, not just another tool rollout. Teams that embrace it deliberately and adapt culture, processes, and quality ownership, get faster and maintain quality. Teams that only embrace speed end up with a lot of code and a lot of regret.

Want to Adopt This in Your Team?

The talk gives an overview of the approach we’ve been refining for years across complex, real-world projects. We’ve packaged the full method — including hands-on exercises, reusable workflows, and team adoption guidance — into a structured training program based on our Systematic AI Coding approach. It’s tool-agnostic and works with Copilot, Cursor, Claude Code, Theia IDE, Codex, Cline, Roo, Windsurf, and similar tools.

👉 AI Coding Training for Teams available as self-paced online course or as a guided adoption package with workshops and tailored support.

If you’d like to discuss how to introduce structured AI coding in your specific project, including niche stacks, embedded systems, or custom agent setups, we’re happy to help.

👉 Contact us

💼 Follow us: EclipseSource on LinkedIn

🎥 Subscribe to our YouTube channel: EclipseSource on YouTube