The creative spark for a new feature, analysis module, or custom calculator. Instead of diving into code immediately, the idea is funneled directly into deep domain research to validate feasibility and gather empirical context.

Deep fact-finding across primary databases (GTEx, ChEMBL, PubMed, Europe PMC, OpenTargets) or market datasets. This forms the absolute baseline parameters, requirements, and regulatory constraints of the project.

Maintaining structured files (`PRD.md`, `README.md`, `SCHEMA.md`, `GEMINI.md`). The Product Requirements Document (PRD) acts as a strict, ironclad contract between the human engineer and the AI agent, defining the absolute scope of the project.

Formulating context-rich instructions referencing active file descriptors (using `@` symbols) and strict PRD parameters. This provides the agent with deep context and focuses its attention exactly where needed.

The agent proposes a granular, step-by-step implementation plan, task list, and walkthrough. **Plan review is the primary gatekeeper** preventing vibration-prone, direct "vibe coding."

Auditing the planned changes and logic changes. Jonah acts as the chief architect, examining the proposed diffs against security standards, database schemas, and performance constraints.

Jonah signs off on the implementation plan. This gives the autonomous agent authorization to edit the workspace files, ensuring the human remains in ultimate control of the repository.

The agent executes the plan, applying targeted edits to the workspace files. Google Antigravity coordinates the file changes inside clean code buffers, ensuring non-disruptive workspace modifications.

Automated verification suites (PHPUnit, PyTest, local syntax check compilers) are executed across all modified files. If the code breaks or syntax errors emerge, compiler logs are dumped back to the chat.

The quality assurance boundary. If tests pass successfully, the branch is directed to document synchronization. If tests fail or syntax errors persist, the system loops back to adjust the prompt or models.

If the agent falls into cognitive loops or fails to debug after 2 attempts, Jonah resets the chat thread or swaps LLM models (e.g. from Qwen to DeepSeek). This prevents token waste and cognitive deadlocks.

Following a fully passing build, the agent is prompted to update the PRD, README, and schemas to align documentation with the live codebase, ensuring the source of truth never falls stale.