| name | pm |
| description | Project manager for GitHub issues: brainstorm design approaches, create structured issues optimized for LLM agent teams, triage and recommend what to work on next, audit and clean up stale issues, or deep-validate a single issue against the codebase. Triggers: create issue, plan work, new task, project manager, write ticket, draft issue, plan feature, plan project, start project, create ticket, review issue, brainstorm, explore idea, think through, pm. |
| user-invocable | true |
| argument-hint | [brainstorm | next | update | review ISSUE_NUMBER | -quick <description>] |
| allowed-tools | ["Task","Skill","Read","Write","Edit","Bash(gh:*)","Grep","Glob","AskUserQuestion","WebSearch","WebFetch"] |
| metadata | {"author":"claude-pm","version":"2.0"} |
Project Manager
GitHub issue lifecycle: brainstorm, create, triage, audit, and review.
Sub-Skills
| Skill | Command | Description |
|---|
| Create | /pm or /pm -quick <desc> | Create structured issues optimized for agent execution |
| Brainstorm | /pm:brainstorm | Explore problem space and design approaches before creating issues |
| Next | /pm:next | Triage open issues — recommend what to work on next |
| Update | /pm:update | Audit open issues — find stale, orphaned, and drift |
| Review | /pm:review ISSUE_NUMBER | Deep-validate a single issue against the codebase |
Usage
/pm → Create a new issue (interactive flow)
/pm -quick fix the login → Create issue with smart defaults
/pm brainstorm caching → Explore approaches before creating issues
/pm:brainstorm → Same (alternate syntax)
/pm next → Triage: recommend next issue to work on
/pm:next → Same (alternate syntax)
/pm update → Audit: find stale/orphaned issues
/pm:update → Same (alternate syntax)
/pm review 42 → Review: validate issue #42 against codebase
/pm:review 42 → Same (alternate syntax)
Routing
Parse the first argument:
| Argument | Route to |
|---|
brainstorm | /pm:brainstorm sub-skill |
next | /pm:next sub-skill |
update | /pm:update sub-skill |
review | /pm:review sub-skill |
-quick [desc] | Create Issue Workflow (quick mode) below |
| anything else / empty | Create Issue Workflow below |
If routed to a sub-skill, invoke it with Skill and pass remaining arguments. If the first
argument is -quick, skip the Ambiguity Check and go directly to the Create Issue Workflow
(quick mode is an explicit signal that the user wants fast issue creation, not brainstorming).
Otherwise, continue with the Ambiguity Check below.
Ambiguity Check
Before starting the Create Issue Workflow, check for existing specs and assess readiness.
Check for existing specs. Use Glob to find specs in .dev/pm/specs/*.md, then Read the
most recent candidates to check topic relevance and status. If multiple specs match, prefer the
most recent by date prefix in the filename. If a spec exists that matches the
user's topic and its status is Approved, use it as primary context for the Create Issue
Workflow — the brainstorming was already completed. Skip the ambiguity check and proceed directly.
If the matching spec's status is Draft (or missing), do not bypass ambiguity handling — the
brainstorm hasn't been approved yet. Continue with the assessment below.
Assess ambiguity. If no matching Approved spec exists, check for these ambiguity signals:
| Signal | Examples |
|---|
| Multiple valid approaches | "we need better caching", "rethink the auth flow" |
| Unclear scope | "improve the plugin system", "make it faster" |
| No obvious issue type | request doesn't map clearly to bug/feature/refactor/chore |
| Architecture decision needed | "should we use X or Y", "how should we structure this" |
| Exploration language | "I'm thinking about...", "what if we...", "I'm not sure how to..." |
If two or more signals are present, suggest brainstorming before issue creation:
Question: "This sounds like it could benefit from brainstorming before creating issues —
there are multiple approaches and the scope isn't fully defined yet. Want to explore first?"
Options:
- Yes, brainstorm first (/pm:brainstorm) — explore approaches and produce a spec
- No, create the issue directly — I know what I want
If the user chooses brainstorming, invoke /pm:brainstorm with the original request as arguments;
otherwise (user chose to proceed directly, or fewer than two signals present), continue with the
Create Issue Workflow below.
Do NOT block on this check — it is a recommendation, not a gate. Clear requests like "fix the
login crash" or "add a health check label" should flow straight through without asking.
Create Issue Workflow
Create structured GitHub issues optimized for LLM agent execution first, human readability second.
Every issue produced by this skill follows the Agent-Optimized Issue Format — structured sections
with consistent headers, machine-parseable acceptance criteria, explicit file paths, verification
methods, and clear scope boundaries.
Activation
This skill activates when users want to create work items for an agent team. Recognize these signals:
| Signal | Examples |
|---|
| Direct | "create an issue", "write a ticket", "plan this work" |
| Implicit | "we need to fix...", "let's add...", "can we refactor..." |
| Shorthand | "/pm", "project manager", "create task" |
Arguments
| Flag | Description |
|---|
-quick | Quick mode — propose smart defaults instead of blocking on ambiguity |
Usage: /pm -quick add a delete button to user profiles
Parse the first argument for -quick. If present, activate quick mode. Everything after the flag is the task description.
Core Workflow
1. Classify → 2. Discover → 3. Challenge → 4. Explore Codebase → 5. Draft → 6. Review → 7. Create
Step 1: Classify Issue Type
Use AskUserQuestion to determine the issue type:
Question: "What type of work is this?"
Options:
- Bug: Something is broken or behaving incorrectly
- Feature: New functionality or enhancement to existing behavior
- Epic: Large initiative requiring 3+ coordinated tasks
- Refactor: Improve code structure without changing behavior
- New Project: Build something from scratch (includes tech stack decisions)
- Chore/Research: Maintenance, dependency updates, spikes, investigations
If the user's initial message already makes the type obvious (e.g., "there's a crash when..."),
skip this step and classify automatically. State your classification and proceed.
Step 2: Type-Specific Discovery
Run the question flow for the classified type. See references/WORKFLOWS.md.
Key principles:
- Use
AskUserQuestion for structured choices (max 4 questions per call, 2-4 options each)
- Use follow-up conversation for open-ended details
- Batch related questions together to minimize round-trips
- If user says "you decide" or similar, make a reasonable choice and note it as
[AGENT-DECIDED: rationale]
- Mark gaps as
[NEEDS CLARIFICATION: question] — don't guess on ambiguous requirements
- Never accept vague requirements. Before moving to the next step, critically examine every
requirement for specificity. If a user says "add a button", ask: Where? What states? What action?
If a user says "fix the bug", ask: What exact behavior? What's expected vs actual? What triggers it?
Treat every underspecified detail as a blocker.
Step 3: Requirements Challenge
After discovery, systematically check for underspecified requirements. See the
Requirements Challenge Checklist in references/WORKFLOWS.md
for the full dimension list.
Default mode (critical):
- Analyze gathered requirements against the challenge checklist dimensions
- Identify every gap — placement, states, behavior, edge cases, error handling, accessibility
- Use
AskUserQuestion to probe each gap (batch related questions, max 4 per call)
- Do NOT proceed to codebase exploration until all critical ambiguities are resolved
- An agent executing the resulting issue should never need to guess intent
Quick mode (-quick):
- Analyze gathered requirements against the same checklist dimensions
- For each gap, propose a smart default with rationale
- Present all assumptions in a single summary: "Here's what I'll assume: [list]. Confirm or correct?"
- Proceed after one confirmation round — only block on truly ambiguous requirements where no
reasonable default exists
- Tag every assumed detail in the final issue body with
[AGENT-DECIDED: rationale]
Example — "add a delete button to user profiles":
Critical mode:
I need to clarify several details before drafting:
- Where on the profile page should the button go? (header actions, settings section, footer)
- Who can see it? (all users, admins only, own profile only)
- What happens on click? (immediate delete, confirmation dialog, soft delete)
- What are the visual states? (default, hover, disabled, loading)
- What about error handling? (network failure, permission denied)
Quick mode (-quick):
Here's what I'll assume — confirm or correct:
- Placement: profile header action bar, right-aligned
- Visibility: own profile only, admins on any profile
- On click: confirmation dialog → soft delete → redirect to home
- States: default (red outline), hover (red fill), loading (spinner), disabled (greyed, no permission)
- Errors: toast notification with retry option
Step 4: Codebase Exploration
Before drafting, explore the codebase to enrich the issue with concrete details:
- Discover: Use the Explore agent to survey the codebase for structure, conventions, and relevant areas. For large or unfamiliar codebases, use repomix-explorer (if available) to get a structural overview.
- Target: Use
Glob, Grep, and Read to drill into specific files identified during discovery — find files that need modification, understand current patterns, check for related work (TODOs, existing tests, related components), and verify that proposed changes don't conflict with existing code.
This step is critical — agents executing the issue will perform better with accurate file paths
and pattern-aware implementation hints.
Step 5: Draft the Issue
Use the appropriate template from references/TEMPLATES.md.
Agent-first formatting rules:
- Sections are contracts — every section header means something. Agents parse them.
- Acceptance criteria are tests — write them as verifiable assertions:
VERIFY: [condition]
- File paths are absolute from repo root —
src/auth/login.ts, not "the login file"
- Approach is sequential — numbered steps an agent follows linearly
- Scope is explicit — "In Scope" and "Out of Scope" prevent agents from over-engineering
- Dependencies are linked —
Blocked by: #N and Blocks: #N
- Constraints are non-negotiable — performance targets, backwards compatibility, etc.
Write the draft to a temp file: /tmp/issue-body.md
Step 6: Review
Present the draft to the user with a summary:
- Title
- Type and labels
- Key acceptance criteria
- File scope
Ask: "Ready to create this issue, or want to adjust anything?"
For epics: also present the sub-issue breakdown before creating.
Step 7: Create
Size label selection (before creating the issue):
Applies to: Feature, Epic sub-issue, New Project sub-issue, Refactor.
Skips: Epic parent, New Project, Bug, Chore, Research.
Determine the size label deterministically from analysis already gathered in earlier steps.
Features — derive from Step 2 breadth signals and Step 3 structural change detection:
| Analysis result | Size label |
|---|
| 0 breadth signals, no structural changes | size/S |
| 1 breadth signal, no structural dependency retained | size/M |
| Structural dependency retained in single issue (user declined split) | size/L |
| Multi-story routed to Epic | No size label on parent; each sub-issue labeled independently |
Refactors — use target area from Refactor Flow Round 1, Question 2:
| Target area | Size label |
|---|
| Single file/function | size/S |
| Single module/component | size/M |
| Cross-module or architecture-level | size/L |
Epic sub-issues — derive from the decomposition table columns (Subsystems and Structural Changes recorded during Round 2 decomposition):
| Subsystems | Structural Changes | Size Label |
|---|
| Single | No | size/S |
| Single | Yes | size/M |
| Multiple | No | size/M |
| Multiple | Yes | size/L |
New Project sub-issues — use the same decomposition-based sizing as Epic sub-issues.
size/XL is intentionally excluded — work at that scale is always decomposed into an Epic with independently sized sub-issues.
gh issue create --repo OWNER/REPO \
--title "<type-prefix>: <description>" \
--body-file /tmp/issue-body.md \
--label "<type-label>,<size-label>"
gh issue create --repo OWNER/REPO \
--title "<type-prefix>: <description>" \
--body-file /tmp/issue-body.md \
--label "<type-label>"
Title prefixes by type:
| Type | Prefix | Label | Size? |
|---|
| Bug | fix: | bug | No |
| Feature | feat: | enhancement | Yes |
| Epic | epic: | epic | No (sub-issues: Yes) |
| Refactor | refactor: | refactor | Yes |
| New Project | project: | project | No (sub-issues: Yes) |
| Chore | chore: | chore | No |
| Research | spike: | research | No |
On failure: Save draft to /tmp/issue-draft-{timestamp}.md, report error.
On success: Remove the temp file: rm -f /tmp/issue-body.md
For epics and new projects with sub-issues: create the parent issue first, then each sub-issue with Part of #PARENT_NUMBER in its body. After creating each sub-issue, formally link it to the parent via GitHub's Sub-Issues API so the parent's sidebar tracks completion progress:
sub_issue_id=$(gh api "repos/OWNER/REPO/issues/SUB_NUMBER" --jq .id)
gh api "repos/OWNER/REPO/issues/PARENT_NUMBER/sub_issues" \
--method POST -F "sub_issue_id=$sub_issue_id"
Use -F (capital F) for sub_issue_id — the API requires an integer, and -F sends it as a number while lowercase -f would send a string, causing a type error.
If either the sub_issue_id lookup or the POST to /sub_issues fails for a sub-issue (rate limit, plan doesn't support sub-issues, permissions), log the error and continue to the next sub-issue — don't abort the loop. Skip the POST if the ID lookup failed. The Part of #N markdown reference still provides a human-readable cross-reference even if the API link fails.
Clean up /tmp/issue-body.md only after all sub-issue creation and linking attempts have completed, even if some links failed.
Report all created issue URLs to the user.
Quality Checklist
Before creating any issue, verify:
Duplicate Check
Before creating, always search for existing issues:
gh issue list --search "keywords" --state all --limit 10
If similar issue exists → inform user, suggest linking instead of duplicating.
Repo Detection
Detect the current repo automatically:
gh repo view --json nameWithOwner -q .nameWithOwner
If not in a git repo or no remote → ask user for the target repo.
Templates & Workflows
