How to properly execute speckit stages?

[voipp]

There is optional hook called /speckit-agent-context-update which compressess context between stages specify, plan, tasks...

But i prefer to clear context and start new session during stages so safe my tokens)

So, what is the prefferred way ?

Off-topic question: good models set for speckit stages?

[identity-wael]

[Removed by author.]

[voipp]

what is the "agent context file"

[mnriem]

An agent context file is a repository-level instruction file that your coding agent may load automatically—such as AGENTS.md, CLAUDE.md, GEMINI.md, or .github/copilot-instructions.md. It is not the model’s conversation context window.

Importantly, the Spec Kit workflow commands do not require an agent context file or the agent-context extension. Each command contains the instructions needed for its stage and reads the relevant generated artifacts, such as spec.md, plan.md, and tasks.md.

The optional agent-context extension is only a convenience. It maintains a small marked section in the agent’s instruction file pointing to the current plan.md; it does not compress conversation history or preserve hidden session state.

Starting a fresh session between specify, plan, tasks, and implement is therefore perfectly valid. The generated artifacts provide the durable handoff between stages. If you do not install or enable the extension, the workflow commands still operate normally and Spec Kit will not modify an agent context file.

Drafted on behalf of @mnriem by GitHub Copilot (model: GPT-5.6 Sol); fully AI-drafted for human review.

[Dotoryman]

Starting a fresh chat between stages is fine. Spec Kit carries the work forward in the feature artifacts (spec.md, plan.md, tasks.md), rather than requiring one continuous conversation. In the new chat, keep the same repository/feature branch and point the agent at those artifacts before running the next stage. Review each artifact before moving on: specify → plan → tasks → implement → converge; add clarify and analyze when the requirements or cross-artifact consistency need checking.

agent-context.update is a separate, opt-in extension for maintaining a section of an agent instruction file (such as AGENTS.md or CLAUDE.md). It is not a required chat-history compression step. If you don't want Spec Kit to maintain that file, you can leave the extension uninstalled/disabled and still use the core workflow. If you do use it, commit or otherwise retain the managed context file so a new session sees it.

I wouldn't prescribe one model per stage as a Spec Kit rule. Pick a model that reliably follows the commands and can read the relevant artifacts; spend more attention on reviewing spec.md and plan.md before implementation than on keeping the same model or chat open.

References: quickstart/workflow, agent-context extension.