blitz

name	blitz
description	Maximum parallelism burst — all tasks execute simultaneously with no coordination overhead

When this skill is invoked, you MUST follow this exact output sequence.

Step 1: Header (print IMMEDIATELY)

[BW] blitz — maximum parallelism, zero overhead...

Step 2: Task decomposition (aggressive splitting)

Split the user's request into the MAXIMUM number of independent sub-tasks. Blitz splits more aggressively than trio:

• Trio: 1-5 tasks with agents per task
• Blitz: up to 10 tasks, 1 agent per task, all parallel

[BW] blitz: {N} parallel tasks identified
  1. "{task}" → bestwork:{agent}
  2. "{task}" → bestwork:{agent}
  ...
  N. "{task}" → bestwork:{agent}
[BW] launching all...

Step 3: Launch ALL agents simultaneously

Print each launch line, then spawn immediately. Do NOT wait between spawns:

[BW] ▶ blitz: bestwork:{agent1} (task 1)
[BW] ▶ blitz: bestwork:{agent2} (task 2)
[BW] ▶ blitz: bestwork:{agent3} (task 3)
[BW] ▶ blitz: bestwork:{agent4} (task 4)
[BW] ▶ blitz: bestwork:{agent5} (task 5)

ALL agents use run_in_background: true. Every agent runs independently with no cross-communication.

For EACH agent, call TaskCreate so the user sees live spinners:

TaskCreate: subject="bestwork:{agent} (task N)", description="{task}", activeForm="bestwork:{agent} working..."

When done, TaskUpdate to completed.

Step 4: Results (print as each completes, in completion order)

[BW] ✓ task 3 done — bestwork:{agent3} — {summary}
[BW] ✓ task 1 done — bestwork:{agent1} — {summary}
[BW] ✓ task 5 done — bestwork:{agent5} — {summary}
[BW] ✓ task 2 done — bestwork:{agent2} — {summary}
[BW] ✗ task 4 failed — bestwork:{agent4} — {error}

Step 5: Conflict resolution (if needed)

If multiple agents edited the same file, detect and resolve:

[BW] blitz: conflict detected in {file}
[BW]   task 2 wrote lines 10-20
[BW]   task 4 wrote lines 15-25
[BW]   resolving: merging changes (task 2 priority — completed first)

Step 6: Quick verify

Run a single verification pass over all changes:

[BW] blitz verify: type check... {PASS|FAIL}
[BW] blitz verify: tests... {PASS|FAIL}

If any verify fails, fix inline (no retry loop — blitz is about speed).

Step 7: Summary

[BW] ═══════════════════════════════════
[BW] blitz complete: {N} tasks, {M} parallel agents, {K} conflicts resolved
[BW] ═══════════════════════════════════

Step 8: Summary Report

After printing the summary line, call report-writer (src/observe/report-writer.ts) to save detailed output:

writeReport("blitz", {
  total: {N},
  done: {M},
  agents: [list of all agents used],
  tasks: [list of task descriptions],
  durationMs: {elapsed ms},
  decisions: [conflict resolutions and key choices]
}, projectRoot)

Then call summarize (src/observe/result-summarizer.ts) and print the 1-liner:

[BW] ✓ {M}/{N} done — details: .bestwork/reports/blitz-{timestamp}.md

When to use blitz vs trio

• Blitz: many independent changes (batch fixes, multi-file refactors, style updates, lint fixes, test additions). No coordination needed.
• Trio: tasks that need quality gates (PM review, critic feedback). Coordination matters.

Rules

• ONE agent per task — no PM, no critic, no lead
• ALL agents launch simultaneously — maximum parallelism
• NO meeting logs — blitz skips all ceremony
• NO feedback loops — if something fails, fix it once and move on
• NO confirmation prompt — blitz is fire-and-forget (fire-and-forget)
• Conflicts resolved by completion order (first to finish wins, later merges)
• Agent selection: tech specialists only, best-fit per task domain
• Up to 10 parallel tasks (Claude Code agent limit)