Soufiane Keli, a consultant at OCTO Technology, proposes a methodical approach that allows generative AI to produce nearly 100% of the code in a real software project. Far from magical, this strategy rigorously combines proven practices within a process structured into four steps.
Step 0: Daily LLM Onboarding
Unlike a human developer who retains context across sessions, the LLM must be "rehired" every day. This crucial step consists of systematically restating the project's global context, its business and architectural objectives. Treating the model as a "junior who starts over every morning" enforces a beneficial documentation discipline: explicit context becomes an asset shared by the whole team, not merely tacit knowledge held by a few.
Step 1: Exploration and Atomic Planning
Before any generation, each user story undergoes a meticulous breakdown into atomic tasks with precise descriptions. Paradoxically, this planning itself uses an LLM to identify the optimal level of granularity. This step transforms vague functional objectives into actionable technical specifications, drastically reducing the ambiguity that models handle poorly.
Step 2: Iterative Spec-Driven Development
Development proper relies on highly structured prompts comprising four elements: a detailed technical specification, illustrative code examples, project standards and conventions, and an explicit Definition of Done (DoD) checklist. If the result is unsatisfactory, the approach prescribes a methodical adjustment of either the prompt or the context, avoiding random iterations. This rigor turns interaction with the LLM from an improvised conversation into a reproducible engineering process.
Step 3: Continuous Capitalization
After each cycle, the lessons learned progressively enrich an organizational knowledge base: successful patterns, effective prompt formulations, identified pitfalls, documented examples. This continuous improvement loop turns individual experience into collective intellectual capital, accelerating future projects.
Fundamental Principle and Field Validation
The guiding principle explicitly rejects monolithic generation: "Don't ask the AI to do everything at once". Instead, atomic tasks + clear standards + rapid iteration simultaneously produce velocity AND quality—objectives traditionally seen as antagonistic.
Crucially, this approach was demonstrated in a brownfield environment by Loïc Lefloch and Simon Belbeoch at OCTO Technology, proving its applicability beyond idealized greenfield projects. The brownfield context—with legacy code, existing architectural constraints, technical debt—represents the real working ground for the majority of developers.
Pragmatic Positioning
Keli explicitly positions this method as a pragmatic combination of publicly available good practices, not as a radical innovation. This strategic modesty reinforces its credibility: the approach does not require a cultural revolution, only a disciplined orchestration of known techniques adapted to the generative-AI context.