sprint-status
The sprint-status skill enforces structured communication across parallel Claude Code sessions by requiring each response to include session identification headers and conclude with explicit status markers. Use this skill when managing multiple concurrent Claude instances working on the same project to eliminate confusion about which session completed what work, what remains blocked, and when user input is required before proceeding.
git clone --depth 1 https://github.com/rohitg00/pro-workflow /tmp/sprint-status && cp -r /tmp/sprint-status/skills/sprint-status ~/.claude/skills/sprint-statusSKILL.md
# Sprint Status
When running multiple Claude Code sessions in parallel, confusion is the enemy. This skill ensures every session identifies itself and every step reports its state.
## Session Identification
Every response that involves a decision, plan, or significant action starts with orientation:
```text
SESSION: my-app | branch: feat/auth | task: Add JWT refresh tokens
```
This takes one line. It costs almost nothing. It prevents the user from applying feedback to the wrong session.
### Detecting Parallel Sessions
Check for sibling Claude Code processes:
```bash
pgrep -af "claude" | grep -v "$$" | head -5
```
Or check for active worktrees:
```bash
git worktree list 2>/dev/null
```
Or look for session markers (written by session-start.js / session-end.js):
```bash
ls $TMPDIR/pro-workflow/sessions/ 2>/dev/null | tail -5
```
If multiple sessions are detected, always include the session identification header. If only one session is running, include it at task boundaries and before presenting options.
## Status Lines
End every major step with exactly one status line. No ambiguity.
### STATUS: COMPLETE
All work for the current step is done. Ready to commit, merge, or move to the next task.
```text
STATUS: COMPLETE
Changed: src/auth/refresh.ts, src/auth/refresh.test.ts
Tests: 14 pass, 0 fail
Ready to commit.
```
### STATUS: COMPLETE_WITH_NOTES
Done, but flagging something the user should know about.
```text
STATUS: COMPLETE_WITH_NOTES
Changed: src/api/upload.ts
Tests: 8 pass, 0 fail
Notes:
- Upload size limit is hardcoded to 10MB, should be configurable
- No rate limiting on this endpoint yet (separate task)
```
Notes are for things that work but could be better. Not blockers — observations.
### STATUS: BLOCKED
Cannot proceed without user input or an external dependency.
```text
STATUS: BLOCKED
Blocker: Need database migration approved before writing the ORM layer
Waiting on: DBA approval for schema change in migrations/0042_add_tokens.sql
Can continue: Nothing else in this task until unblocked
```
### STATUS: NEEDS_INFO
Missing context to make a good decision. Asking before guessing.
```text
STATUS: NEEDS_INFO
Question: Should refresh tokens expire after 7 days or 30 days?
Impact: Changes token cleanup job schedule and storage requirements
Default if no preference: 7 days (more secure, standard practice)
```
Always provide a sensible default so the user can say "go with the default" without context-switching into the decision.
## Parallel Session Patterns
### Pattern 1: Feature + Tests
```text
Session 1: feat/auth → implementing JWT refresh
Session 2: feat/auth-tests → writing test suite for auth module
```
Both sessions work on the same feature but don't touch the same files. Session 2 can start writing tests against the interface before Session 1 finishes the implementation.
### Pattern 2: Independent Features
```text
Session 1: feat/upload → file upload endpoint
Session 2: feat/billing → billing webhook handler
Session 3: fix/login-bug → login redirect fix
```
Completely independent. Merge order doesn't matter.
### Pattern 3: Stacked Changes
```text
Session 1: feat/base-types → shared type definitions (must merge first)
Session 2: feat/api → API layer (depends on Session 1)
Session 3: feat/ui → UI layer (depends on Session 1)
```
Session 1 merges first. Sessions 2 and 3 rebase after. Flag the dependency in every status update.
## When to Report Status
| Event | Include Status? |
|-------|----------------|
| File edit completed | No (too granular) |
| Test suite run | Yes, if pass/fail changed |
| Phase completed (research, plan, implement) | Yes |
| Presenting options to user | Yes (with session header) |
| Hitting a blocker | Yes, immediately |
| Before asking a question | Yes (NEEDS_INFO) |
| After committing | Yes (COMPLETE) |
| End of session | Yes (final status) |
## Sprint Dashboard
When the user asks for status across sessions, compile a sprint view:
```text
SPRINT STATUS
Session 1: feat/auth STATUS: COMPLETE (ready to merge)
Session 2: feat/upload STATUS: BLOCKED (waiting on S3 creds)
Session 3: fix/login-bug STATUS: COMPLETE_WITH_NOTES (works, needs perf review)
```
## Anti-Patterns
- Skipping the session header when multiple sessions are active
- Using STATUS: COMPLETE when there are known issues (use COMPLETE_WITH_NOTES)
- Burying the status line in a paragraph (put it on its own line, at the end)
- Reporting status after every single edit (only at meaningful boundaries)
- Not providing a default with NEEDS_INFO (forces the user to context-switch fully)
- Saying "almost done" or "mostly working" instead of a concrete status
## Add to CLAUDE.md
```markdown
## Sprint Status
Every decision and plan starts with: SESSION: project | branch | task
End major steps with STATUS: COMPLETE | COMPLETE_WITH_NOTES | BLOCKED | NEEDS_INFO
NEEDS_INFO always includes a sensible default.
When parallel sessions are detected, always include session headers.
```Coordinate multiple Claude Code sessions as a team — lead + teammates with shared task lists, mailbox messaging, and file-lock claiming. Patterns for team sizing, task decomposition, and when to use teams vs sub-agents vs worktrees.
Auto-configure quality gates, hooks, and settings for a new project. Detects project type and sets up appropriate tooling. Use when onboarding a new codebase.
Decompose large-scale changes into independent units and spawn parallel agents in isolated worktrees. Use for migrations, refactors, codemods, and any change touching 10+ files with the same pattern.
Capture a user-reported defect as a durable GitHub issue written in the project's own domain language. Explores the codebase in parallel for context but never leaks file paths or line numbers into the issue. Use when the user reports a bug conversationally, runs a QA pass, or says "file an issue", "log this as a bug", "capture this".
Smart context compaction with state preservation. Saves critical files, task progress, and working state before compaction, restores after. Use before manual compact or when auto-compact triggers.
Analyzes and optimizes context window usage across sessions. Use when context feels bloated, sessions run slow, or approaching compaction limits.
Master the four operations of context engineering — Write, Select, Compress, Isolate. Manage token budgets, compaction strategies, and context partitioning to keep AI sessions sharp and efficient.
Optimize token usage and context management. Use when sessions feel slow, context is degraded, or you're running out of budget.