docs(workflow): add PRD-first, context reset, and system evolution guidance

- Add PRD-first “north star” guidance for non-trivial work
- Add context reset guidance between planning and implementation
- Document system evolution loop to reduce recurring agent mistakes
- Teach /init to look for extras/prd.md and extras/project-brief.md

Author: d-padmanabhan

commands/init.md

@@ -16,7 +16,7 @@ Initialize a new task by analyzing the project and determining the appropriate w
    - Scan project structure
    - Identify tech stack and patterns
    - Note existing conventions
-   - Check for existing context files (tasks.md, active-context.md)
+   - Check for existing context files (`extras/prd.md`, `extras/project-brief.md`, `extras/tasks.md`, `extras/active-context.md`)
 
 2. **Understand the Request**
    - What is the user asking for?

rules/010-workflow.mdc

@@ -79,6 +79,15 @@ You MUST follow this three-phase approach for all non-trivial changes:
 - [ ] The scope is clearly defined and agreed upon
 - [ ] Risks and dependencies are identified
 
+### PRD-first development (North Star)
+
+For non-trivial work, keep a single “north star” scope document and use it to drive small, sequential feature slices.
+
+- **Recommended file**: `extras/prd.md`
+- **Greenfield**: Define scope, non-goals, key flows, architecture, milestones
+- **Brownfield**: Capture current state + next desired state + constraints
+- **Rule**: Break work into features small enough for the agent to execute reliably (one feature per cycle)
+
 **When Planning is Required:**
 - Multiple files or modules affected
 - New features or significant refactoring
@@ -162,6 +171,14 @@ You MUST follow this three-phase approach for all non-trivial changes:
 4. **Stop at boundaries** - Don't fix unrelated issues or add unplanned features
 5. **Complete the agreed scope** - Then STOP and wait for review
 
+### Context reset between Plan → Build
+
+When moving from planning to implementation, prefer a fresh context window so the agent has maximum room to reason.
+
+- **Rule**: After the plan is approved, restart the chat/session before coding
+- **Input**: Provide the approved plan doc (and only the minimum extra context needed)
+- **Goal**: Keep implementation context lightweight; avoid carrying long planning threads into execution
+
 **Constraints:**
 -  Implement only what was planned
 -  Follow all coding standards from `100-core.mdc` and language-specific rules
@@ -204,6 +221,14 @@ AI: [Adds only login form as specified]
 5. **Identify cleanup opportunities** - Can we remove/simplify existing code?
 6. **Propose next steps** - What should happen next?
 
+### System evolution (make the agent better over time)
+
+Treat recurring mistakes as a signal to improve the system, not just the code:
+
+- **If a bug recurs**: update the relevant rule(s), command(s), or reference docs so it doesn’t happen again
+- **If validation was missed**: update templates/commands to force the check (tests, lint, security scan)
+- **Capture learnings**: record patterns in `extras/reflect-*.md` (Level 3-4) and keep the guidance crisp
+
 **Review Checklist:**
 - [ ] Changes match the agreed plan
 - [ ] Security considerations addressed