Manifesto
Spellbook is built on a simple claim: coding agents should not improvise inside fragile systems. Intent, domain truth, requirements, boundaries, verification, and human judgment must be part of the execution environment.
Ten rules for specification-driven development with coding agents inside real codebases.
Software should not emerge from vague prompts alone. Before code is generated, the system should know the problem, the audience, the expected behavior, the constraints, and what success means.
Important business rules should not live only in meetings, tickets, or someone's memory. They should be written, versioned, reviewed, and connected to implementation.
Requirements should not be decorative documents. They should shape plans, tests, quality gates, review criteria, and release decisions.
Coding agents should not operate as unconstrained text generators. They need rules that define what can happen, what must stop, and what requires review.
A completed task should not only produce a diff. It should produce evidence that lets engineers understand the change without reconstructing a chat session.
Every real repository has local knowledge. Naming, errors, logging, testing, structure, ownership, and deployment rules matter. Agents should learn these rules from the repo, not guess them from generic training.
If AI-generated code cannot be verified, it should not be trusted. Tests, gates, static checks, security checks, and runtime validation are part of the development loop.
Specification-driven development does not remove engineers. It gives engineers better leverage. Humans define intent, resolve ambiguity, own tradeoffs, review evidence, and improve the system.
Every task should improve the project's memory. When something fails, the lesson should update requirements, specs, invariants, forbidden shortcuts, tests, conventions, and architecture rules.
The future of software is not just more agents writing more code. The future is controlled collaboration between humans, agents, specs, tests, architecture, and runtime feedback.
More code for its own sake is not the goal. Correct, legible, enforceable, auditable, and constraint-aware software is the goal.
We do not need more code for its own sake. We need systems that are correct, legible, enforceable, auditable, and aligned with real-world constraints.
Spellbook exists for that future: a disciplined environment where coding agents do not merely generate code, but execute inside an engineering system.