In short
The GitHub Spec Kit involves a whole pipeline of steps and a CLI, but for most projects, three skills in Cursor and the rule “don’t write the result until it’s been approved” are enough. I’m figuring out what to cut and what to definitely keep.
GitHub Spec Kit is a set of tools for spec-driven development with AI agents: the steps “constitution,” “specify,” “plan,” “tasks,” and “implement,” a separate CLI, and the .specify/ directory. It sounds impressive, but in practice, the full pipeline often creates more noise than it’s worth. Separate plan and tasks steps, along with batch recompilation, multiply the number of files, increase the risk of overwriting agreed-upon changes, and bloat the model’s context without any real benefit—especially if tasks run one at a time rather than as a whole pipeline.
The idea behind this Habr article is simple: don’t abandon the methodology, but streamline it to the bare minimum and integrate it directly into Cursor. The formula is: draft (optional) → specification → human approval → implementation. The agent doesn’t improvise the product: it relies on sources (code, tickets, wiki) and the agreed-upon specification, or explicitly marks items as [UNKNOWN] / TODO. reasoning_temperature: low on the final implementation prevents default assumptions.
Three components instead of five steps:
.cursor/rules/*.mdc: project constitution, restrictions, pipeline. Always in context./project-init (one-time: constitution, vision, roadmap), /project-specify (draft → specification, focus update), /project-implement (only after human approval: code or publishable artifacts).constitution.md, specs/, feature.json (pointer to the active specification).A separate tasks.md is not needed. The plan resides in the vision (major packages), in the checklist within each specification (files, sections, readiness criteria), and in feature.json (what’s currently in focus).
The first is meaning: the constitution, specifications, drafts, and rules/skills. The second is the result: code, tests, and published documents. Until the agreed status is assigned, the output is not written. The status is assigned by a person—an agent cannot grant itself approval. This is, perhaps, the main rule of the entire approach.
If tasks proceed sequentially, coordination and protection against fabricated facts are important, but separate plan / tasks / analyze steps increase noise without adding value—a simplified workflow is sufficient. A full Spec Kit is appropriate when you need a ready-made CLI, a standard multi-step pipeline, rigid artifact templates, or if the team is already working within the specify ecosystem. The approaches are compatible: the constitution and specifications can be reused when transitioning to the full Spec Kit.
The main takeaway is that pipeline discipline is more important than directory structure. The three skills and the rule “don’t invent facts, don’t write the result until it’s agreed upon” provide 80% of Spec Kit’s value without the heavy infrastructure. But there’s a limitation: this approach is designed for sequential work on a single specification. If you have a parallel workflow or a large team, the simplified workflow might not be enough, and in that case, a full Spec Kit with a CLI is justified.
Source: All Articles in a Row / Artificial Intelligence / Habr