AI-native SDLC with Cursor

The Cursor SDLC track implements the shared six-stage lifecycle with Ask, Plan, and Agent modes, project rules, worktrees, Bugbot, Cloud Agents, and automations. Use this page as the Cursor adapter: it identifies the native surface for each gate and links to the single canonical procedure for the stage.

The lifecycle stays independent of the editor. Read the AI-native SDLC overview first, then use this map to choose the correct Cursor surface without duplicating the process.

Verified surface

CLI commands were checked on 4 September 2026 against Cursor Agent 2026.01.28-fd13201 and the current Cursor CLI parameters. Worktree, Cloud Agent, and review behavior was checked against Cursor’s official documentation. Re-verify before encoding commands in automation.

Stage map

Stage	Accepted artifact	Cursor surface	Canonical procedure
Plan	intent.md	Ask mode for discovery; Agent mode to save the reviewed artifact	Capture intent
Design	spec.md	.cursor/rules/*.mdc, AGENTS.md, skills, image context, and approved MCP tools	Write the spec
Build	plan.md, diff, tests	Plan mode, Agent mode, Agents Window worktrees, /worktree, /best-of-n	Build from an accepted plan
Test	Test evidence	Terminal commands, browser tools, checkpoints, and the active Agent feedback loop	Close the feedback loop
Deploy	Reviewed pull request	Bugbot, Cloud Agent Autofix, required CI, human merge and release gates	Layer review and approvals
Maintain	Incident record and next intent.md	Cloud Agents and Automations triggered on a schedule or supported event	Close production back into planning

Configure the build path

Prerequisites: a Git repository, Cursor authenticated, a clean working tree, and the repository’s real verification commands documented in project rules.

1. Explore without edits in the IDE’s Plan mode or from the CLI:

   cursor-agent --plan \
     "Read docs/intents/FEATURE.md and docs/specs/FEATURE.md. Inspect the repo and propose an ordered plan with risks, tests, and rollback."

2. Review the plan and save the accepted version as docs/plans/FEATURE.md. Treat .cursor/plans/ drafts as working state unless the team explicitly versions that directory.

3. Isolate implementation with the Agents Window or an IDE worktree skill:

   /worktree implement milestone 1 from docs/plans/FEATURE.md

4. Give the implementation run bounded proof:

   Follow docs/plans/FEATURE.md and the applicable .cursor/rules files.
   Implement only milestone 1. Run typecheck, lint, unit, integration, and visual checks.
   Diagnose failures in the implementation; do not weaken tests or skip required checks.
   Finish with changed files, command results, screenshots where relevant, and remaining risk.

5. Let Bugbot and CI review the pull request, then keep merge and production release behind named human approvals.

Cursor’s .cursor/worktrees.json can install dependencies and copy local configuration into worktrees. Those setup commands execute code, so review them like any other repository automation and never copy production credentials by default.

Put policy in the right layer

• Put scoped, version-controlled instructions in .cursor/rules/*.mdc.
• Use AGENTS.md for portable instructions shared with other compatible agents.
• Put reusable procedures in skills rather than pasting the same prompt into every chat.
• Use Ask or Plan mode for read-only discovery; use Agent mode only after the artifact gate.
• Use branch protection and deployment-environment approvals for merge and production authority.

Automate bounded checks

Cursor Agent supports non-interactive output through --print; --output-format accepts text, json, or stream-json:

cursor-agent --print \
  --sandbox enabled \
  --output-format json \
  "Review the current diff against docs/specs/FEATURE.md. Report blocking findings only."

--force auto-allows commands unless explicitly denied. Reserve it for externally isolated runners with scoped credentials; it is not required for ordinary review.

Prove the adapter works

• cursor-agent --help shows --plan, --mode, --print, --output-format, and --sandbox.
• Plan or Ask mode explores the repository without changing source files.
• /worktree creates a separate checkout and git worktree list shows it.
• The implementation session runs the repository’s actual checks and preserves test intent.
• Bugbot or another agent adds evidence, but a human and platform policy still control merge and production.

Common failure modes

A model or product name ages faster than the workflow. Keep the process model-agnostic. Choose models from the current picker and evaluate them against repository tasks instead of hard-coding a marketing name.

A worktree misses dependencies or configuration. Add reviewed setup commands to .cursor/worktrees.json, then confirm secrets, ports, and databases point at isolated local resources.

Headless mode has more authority than expected. --print has tool access. Keep sandboxing enabled, scope the workspace, and omit --force unless the runner is independently contained.

Continue the track

Artifact chain Create one traceable handoff from intent through production feedback.

Cursor quick start Install Cursor, configure the project, and complete a first feature.

Project rules Place durable context and constraints in the correct Cursor rule layer.

Cloud Agents and Automations Run isolated background tasks and event-driven maintenance workflows.