Define
In Define, your team agrees what a feature should do, what conditions it must meet, and a person approves it before any agent starts. Caliber keeps every version, points out gaps and gives each agent the approved spec to build against.
Agents can produce working code from almost any prompt. But when each engineer prompts in their own way, product decisions get lost. An approved spec gives engineers and agents the same up-to-date brief, so features follow what the team decided.
Code generation without Caliber
Each engineer briefs their agent in their own way.
When a prompt leaves a gap, the agent fills it with an assumption no one approved.
Decisions made while prompting stay in one engineer’s chat history.
Without a complete spec, reviewers check the code against their own understanding of the task.
Spec-driven development with Caliber
Every spec in Caliber follows the same structure.
A person approves the spec before any agent starts, so the team decides what the agent would otherwise guess.
The agent reads the approved spec through Caliber’s MCP connection and builds from it.
Caliber checks every change against each criterion defined in the approved spec.
Versioned specifications
Specs live in the Caliber database and in the project repository at the same time. Each is draft, in review or approved, and every version keeps who changed what. Keeping what a feature should do separate from how it is built stops the design from slowly going off course as agents add features.
Changing requirements
No team can define everything before work starts, and Caliber does not expect it. When the team learns something new, they update the spec and a person approves the new version. From then on, every change is checked against the new version, and the history shows what changed and who approved it.
Automated alignment checks
A model reviews each proposed spec for logical gaps and missing constraints. After the agent builds, it cross-references its implementation against the approved version, criterion by criterion. The finished code reflects what your team approved, with no guesswork from a loose prompt.
