Af Project Management Expertise

Project Management Expertise

When to Use This Skill

Load this skill when you need to:

• Plan a weekly work cycle for a team
• Prioritize backlog issues for scheduling
• Manage team capacity budgets
• Assign issues to Linear cycles
• Review estimation quality across a backlog

Common triggers:

• /pm:plan-week command invoked
• PM asks "what should we work on this week?"
• Sprint/cycle planning sessions
• Backlog grooming and prioritization

Quick Reference

Capacity Constants

Note: Capacity varies by team size and composition. Define your team's weekly capacity in the project's CLAUDE.md (e.g., Team capacity: 50 pts/week).

Issue Status Types

Prioritization Order

Issues are scheduled in this order:

1. Approved issues (approved and estimated, ready to start)
2. Discovered issues with estimates (backlog items with Discovery Estimates)
3. Discovered issues without estimates (need estimation before scheduling)

Within each group, sort by:

1. Priority field (1=Urgent, 2=High, 3=Normal, 4=Low)
2. Estimate size (smaller issues first — complete more work)
3. Age (older issues first — prevent stagnation)

Core Concepts

Per-Team Planning

The PM agent plans for one team at a time. Each team has its own:

• Capacity budget (defined per project, see CLAUDE.md)
• Cycle (weekly time-box)
• Backlog of issues in Discovered/Approved states

The agent is scoped to the team it's installed in. If installed in the Agentview team, it only plans for Agentview issues.

Cycles (Not Milestones)

Linear cycles are the scheduling container:

• Team-scoped — each team manages its own cycles
• Time-boxed — weekly or biweekly duration
• Auto-rollover — incomplete work rolls to the next cycle automatically
• No auto-advance — adding an issue to a cycle does NOT change its status
• Independent of status — cycle assignment and issue status are separate concerns

Estimation Quality Tiers

The estimation skill (af-estimation-expertise) defines two tiers:

For planning purposes:

• Issues with Refined Estimates → schedule with high confidence
• Issues with Discovery Estimates → schedule with medium confidence (pad 20%)
• Issues with no estimate → estimate first using af-estimation-expertise, then schedule

Checking Estimation Status

To determine if an issue has been estimated and at what quality:

1. Has estimate? Check estimate field on the Linear issue (non-null = has story points)
2. What quality? Search the issue's comments for [Discovery Estimate] or [Refined Estimate] tags
3. No estimate? Load af-estimation-expertise and run a Discovery Estimate (quick heuristic)

Linear Hierarchy Model

The Hierarchy

AgentFlow uses Linear constructs as follows:

Key Insight: Sub-issues Are Full Issues

Linear sub-issues have their own:

• Assignee (QA for Behaviour, UX for Storybook)
• Cycle (can differ from parent)
• Estimate (part of parent total, not additional)
• Status (tracked independently)

This enables scheduling spec work into refinement cycles before the parent's delivery cycle.

Milestone Usage

Milestones serve the role that "Epics" previously filled — grouping related issues toward a delivery goal:

• "Auth MVP" — Login + Registration + Session Management
• "v1.0 Release" — Core feature set for first release
• "Q2 Delivery" — Time-boxed quarterly goal

When to create Milestones:

• During Discovery (for capability areas)
• During Planning (for release groupings)

Milestones vs Projects:

• Milestones — For most delivery goals (recommended)
• Projects — Only for "super-epics" spanning multiple milestones (rare)

Scheduling Spec Work (Refinement Cycles)

The Pattern

Specification sub-issues ([Behaviour], [UX]) should be scheduled into cycles before their parent's delivery cycle.

Example:

• Cycle 12 (refinement): [Behaviour] and [UX] sub-issues for Login Flow
• Cycle 13 (delivery): Login Flow parent issue

This ensures specs are complete and approved before implementation begins.

Capacity Planning with Mixed Work

When planning a cycle, include both:

• Delivery work — Parent issues in Approved status
• Refinement work — Spec sub-issues in Approved status

Both consume capacity. A typical split:

• 60-70% delivery work
• 20-30% refinement work
• 10% buffer for unplanned work

Checking Readiness for Delivery

Before scheduling a parent issue into a delivery cycle:

1. Check that its [Behaviour] sub-issue is Done
2. Check that its [UX] sub-issue is Done
3. If specs aren't complete, schedule the sub-issues first

Issues ready for delivery: Status = Approved AND all sub-issues have status = Done

Standing Housekeeping Issues

The Pattern

Each cycle should have a standing "Housekeeping" parent issue for batching trivial work.

Create at cycle start:

Title: Housekeeping — Week of {date}
Estimate: 2-3 points (collective)
Cycle: Current cycle

During the cycle: Add trivial items as sub-issues (titles only, no individual estimates):

• Typo fixes
• Config tweaks
• Unused import removal
• Small copy changes

Benefits:

• Trivial items don't inflate velocity individually
• Aggregate effort is tracked
• Nothing falls through the cracks

Workflow: Plan a Weekly Cycle

Step 1: Gather the Backlog

Query all schedulable issues for the team:

Using Linear MCP tools:
1. list_issues(team: "TeamName", state: "Approved")
2. list_issues(team: "TeamName", state: "Discovered")

For each issue, note:

• identifier (e.g., AGV-34)
• title
• estimate (story points, may be null)
• priority (1-4)
• createdAt (for age sorting)
• labels (for context)

Step 2: Assess Estimation Coverage

Group issues by estimation status:

For unestimated issues:

1. Load af-estimation-expertise
2. Use the Estimation Heuristics section (by change type) for quick estimates
3. Set story points on the issue
4. Post a [Discovery Estimate] comment

Step 3: Prioritize and Fill the Cycle

With the team's weekly capacity budget (defined in project CLAUDE.md):

1. Start with Approved issues — these are committed work
2. Add Discovered issues with estimates — by priority, then size, then age
3. Add Discovered issues with estimates — if capacity remains
4. Include refinement work — spec sub-issues ([Behaviour], [UX]) consume capacity too
5. Leave buffer — aim for 80-90% of capacity (uncertainty margin)

Capacity rules:

• Never exceed the team's weekly capacity in a single cycle
• Leave 10-20% buffer for unplanned work
• If a single issue exceeds remaining capacity, skip it (don't split)
• Flag issues over 21 points — suggest decomposition

Step 4: Present the Plan

Before making any changes, present the proposed cycle to the human:

## Weekly Cycle Plan: [Team] — Week of [Date]

**Capacity:** 50 points | **Planned:** X points | **Buffer:** Y points

### Scheduled Issues (X points)

| # | Issue | Title | Points | Priority | Est. Quality |
|---|-------|-------|--------|----------|-------------|
| 1 | AGV-XX | Title | 5 | High | Refined |
| 2 | AGV-YY | Title | 3 | Normal | Discovery |
| ... | | | | | |

### Newly Estimated (Z issues)

| Issue | Title | Points | Basis |
|-------|-------|--------|-------|
| AGV-ZZ | Title | 3 | Heuristic: new UI component |

### Not Scheduled (Remaining Backlog)

| Issue | Title | Points | Reason |
|-------|-------|--------|--------|
| AGV-AA | Title | 21 | Exceeds remaining capacity |
| AGV-BB | Title | null | Needs detailed estimation |

### Recommendations
- [Any issues that should be decomposed]
- [Any estimation quality concerns]
- [Suggested focus areas]

Step 5: Execute on Approval

After human approves:

1. Get or create the current cycle for the team:

   list_cycles(teamId: "team-uuid", type: "current")
   

   If no current cycle exists, note this and inform the human.

2. Assign issues to the cycle:

   update_issue(id: "AGV-XX", cycle: "cycle-name-or-id")
   

3. Set newly estimated story points:

   update_issue(id: "AGV-ZZ", estimate: 3)
   

4. Post estimation comments for any issues that were estimated during planning.

5. Post summary to Zulip with the final plan.

Workflow: Review Estimation Quality

When reviewing a backlog for estimation health:

Metrics to Report

Report Format

## Estimation Health: [Team]

**Total backlog:** X issues
**Estimated:** Y (Z%)
**Refined:** A (B% of estimated)
**Avg points:** C

### Issues Needing Attention
- [AGV-XX] No estimate — needs at least Discovery Estimate
- [AGV-YY] 34 points — should be decomposed into sub-issues
- [AGV-ZZ] Discovery Estimate only, in Approved status — needs Refined Estimate

Integration Points

Estimation skill (af-estimation-expertise):

• Load for any estimation work (Discovery or Refined)
• Use heuristics table for quick batch estimates
• Follow structured output format for Linear comments

Work management (af-work-management-expertise):

• Linear status flow: Discovered → Refining → Approved → In Progress
• Issue updates via linearis CLI or Linear MCP tools
• Current-task.md for tracking planning sessions

Process skills:

• af-discovery-process — Features should get Discovery Estimates during Phase 4
• af-refinement-process — Features should get Refined Estimates before approval
• af-delivery-process — Story points must be set before starting implementation

Linear MCP tools:

• list_issues — Query backlog by team and state
• list_cycles — Get current/next cycles
• update_issue — Assign to cycle, set estimates
• create_comment — Post estimation breakdowns

Duplicate and Overlap Detection

Before scheduling, scan the backlog for duplicate or overlapping issues:

When to Check

• During cycle planning (Step 2.5 — after gathering backlog, before scheduling)
• After batch issue creation (new issues often overlap existing ones)
• When the backlog exceeds 50 issues (overlap probability increases with size)

Detection Method

1. Theme clustering. Group issues by keywords (e.g., "cloud", "migration", "health", "costs", "activity"). Any theme with 3+ issues warrants review.
2. Parent-child overlap. Check if Epic/parent issues carry their own estimates when children already carry the work — this causes double-counting.
3. Entity vs. deep link overlap. An entity issue (e.g., "Environment Entity") may subsume a smaller issue (e.g., "Deep Links: Environments"). Make the smaller one a sub-issue or mark as duplicate.
4. API + UI refactor pairs. Verify that paired refactor issues (API Refactor + UI Rethink) don't overlap with a broader parent issue covering the same scope.
5. Description cross-reference. If an issue's description says "likely overlaps with X" or "see also Y", investigate immediately.

Resolution Actions

Impact on Totals

After resolving duplicates, recalculate project totals. Report the delta:

• "Removed X pts of double-counting (was Y, now Z)"

Cycle Lifecycle Management

Verify Cycles Exist Before Assigning

Before assigning issues to cycles, always verify the target cycle exists in Linear:

list_cycles(teamId: "team-uuid")

If a cycle doesn't exist, create it first using the Linear GraphQL API:

mutation {
  cycleCreate(input: {
    teamId: "team-uuid",
    startsAt: "YYYY-MM-DD",
    endsAt: "YYYY-MM-DD"
  }) {
    cycle { id number }
  }
}

Never "plan" cycles conceptually without creating them in Linear. Cycle assignments will silently fail if the cycle doesn't exist, and issues will appear unscheduled despite being "planned."

Cycle Naming Convention

Cycles are numbered sequentially per team. When creating future cycles, continue the sequence from the last existing cycle.

Common Pitfalls

1. Over-scheduling. Don't fill 100% of capacity. Leave 5-10 points for unplanned work and interruptions.
2. Ignoring estimation quality. A Discovery Estimate for a Discovered issue is fine. A Discovery Estimate for an Approved issue means Refinement didn't complete properly.
3. Scheduling unestimated work. Never add an unestimated issue to a cycle — estimate it first, even if just a heuristic.
4. Ignoring dependencies. Check parent/child relationships. Don't schedule a sub-issue if its parent isn't ready.
5. Planning across teams. Each team has its own capacity. Don't mix team budgets.
6. Double-counting Epic parents. If an issue has the "Epic" label and its children carry estimates, the parent should have estimate = 0. Otherwise totals are inflated.
7. Forgetting to create cycles. Always verify cycles exist in Linear before assigning issues. Conceptual planning without cycle creation leads to silent failures.
8. Missing delivery by-products. Issues created during delivery (follow-on work, "what's left" breakdowns) bypass the estimation workflow. Run an estimation sweep after each delivery cycle.

Remember:

1. Capacity is project-specific — check project CLAUDE.md for the team's weekly budget
2. Plan for one team at a time — the team the agent is installed in
3. Prioritize: Approved > Discovered
4. Always present the plan before executing
5. Leave capacity buffer for unplanned work (10-20%)
6. Estimate unestimated issues before scheduling them
7. Schedule spec sub-issues ([Behaviour], [UX]) into refinement cycles before parent delivery cycles
8. Create a standing Housekeeping issue each cycle for trivial work
9. Scan for duplicates and overlaps before finalizing cycle plans
10. Verify cycles exist in Linear before assigning issues to them