Skip to content

Latest commit

 

History

History
96 lines (69 loc) · 4.4 KB

File metadata and controls

96 lines (69 loc) · 4.4 KB

Example: The Diverge-Converge Pipeline

This walkthrough shows how the four workflow skills chain together to go from "vague idea" to "informed implementation" — with multi-agent research, structured debate, and parallel prototyping along the way.

Scenario

You're building a caching layer for a multi-service backend. You're not sure whether to use Redis, an in-process cache, a CDN-based approach, or something else. The requirements are fuzzy and the trade-offs are real.

Step 1: Research with /diverge

Start by exploring the problem space from multiple angles:

/diverge 4 "What caching architecture should we use for a multi-service backend with mixed read/write patterns, 50ms p99 latency target, and eventual consistency tolerance?"

What happens:

  • Claude frames a research brief and asks you to confirm
  • 4 independent agents launch in parallel, each with a different lens:
    • Prior art: surveys existing caching patterns (write-through, write-behind, read-aside, etc.)
    • Technical design: analyzes the specific constraints (50ms p99, eventual consistency)
    • Failure modes: explores what breaks — cache stampedes, invalidation bugs, cold start
    • Scale & evolution: considers how this grows as services multiply
  • After all 4 return, Claude synthesizes findings and drafts a PRD

Output: prd_caching_architecture.md with requirements, trade-offs, and open questions.

Step 2: Debate with /converge

The /diverge synthesis revealed tensions — the "prior art" agent recommended Redis, the "failure modes" agent flagged single-point-of-failure risks, and the "scale" agent suggested a tiered approach. Time to debate:

/converge 3 prd_caching_architecture.md

What happens:

  • Claude extracts the key tensions from the PRD
  • 3 teammates are assigned positions:
    • Centralized cache (Redis): argues for simplicity, proven tooling
    • Distributed/tiered cache: argues for resilience, locality
    • Hybrid approach: argues for layered caching with fallbacks
  • 2-3 rounds of structured debate with steel-manning and concessions
  • Claude moderates and produces a convergence report

Output: Resolved consensus points, refined trade-offs, and a recommended direction (e.g., "tiered: in-process L1 + Redis L2, with circuit breaker on Redis failures").

Step 3: Prototype with /diverge-prototype

Now build it multiple ways and see what the code actually looks like:

/diverge-prototype 3 prd_caching_architecture.md

What happens:

  • Claude designs 3 implementation strategies based on the (now-refined) PRD:
    • Minimal: simple Redis wrapper, ~100 lines
    • Tiered: in-process LRU + Redis with fallback, ~300 lines
    • Event-driven: pub/sub invalidation with local caches, ~400 lines
  • You confirm the strategies
  • 3 agents launch in isolated git worktrees, each building a real prototype
  • After all return, Claude compares: files changed, requirements met, trade-offs encountered

Output: comparison matrix, per-prototype highlights, recommended path forward.

Step 4: Pick and Ship

Claude presents the comparison and asks how you want to proceed:

  1. Adopt one — merge the best prototype's branch
  2. Cherry-pick — combine the tiered prototype's fallback logic with the event-driven prototype's invalidation
  3. Iterate — refine the PRD and run another round
  4. Park it — save learnings for later

Why This Works

Each step adds signal that the previous step couldn't:

Step What it adds
/diverge Breadth — surfaces options and trade-offs you wouldn't find alone
/converge Depth — stress-tests each option through adversarial debate
/diverge-prototype Reality — reveals what the code actually looks like, not just what it sounds like

The pipeline is also flexible — you can enter at any point:

  • Already know the options? Skip /diverge, go straight to /converge
  • Already have a PRD? Skip to /diverge-prototype
  • Just need ideas? Use /brainstorm standalone

Cost Awareness

This pipeline spawns many agents. Rough token usage:

  • /diverge with 4 agents: ~40-80k tokens
  • /converge with 3 debaters x 3 rounds: ~30-60k tokens
  • /diverge-prototype with 3 agents: ~60-120k tokens (these write code)

Use fewer agents (2 instead of 4) for smaller decisions. The pipeline is most valuable for decisions that are expensive to reverse — architecture choices, API contracts, data model design.