stelow-workflow-architecture-alternatives

Architecture Alternatives

Tools: See ../stelow-workflow-orchestrator/references/cli-tools/subagents.md for subagent patterns.

Overview

This skill executes the Architecture Alternatives phase: diverging construction directions before any commitment. It is the structural sibling of interaction alternatives. Interaction asks how it behaves. Architecture asks how it is built.

How to Load

Via Orchestrator (recommended)

The orchestrator reads this file directly when needed.

Standalone

This skill works standalone. Use the Input Detection section below to tell the skill what construction decision to explore. Follow the instructions inline.

Standalone awareness: when inside stelow, reads exploration breadth from the workflow config (exploration.count, exploration.hybrid). When standalone, defaults to 3 directions + hybrid. If the config is not reachable, pass the count explicitly.

Process

Step 0: Read exploration breadth and choose direction count.

Exploration breadth controls how many construction directions are compared. A hybrid synthesis is produced whenever more than one direction exists, independent of the count.

Count	Directions explored	Hybrid
1	1 direct direction only. No divergence. Trivial changes by explicit override only.	skip
2	2 most-differentiated directions	yes
3	3 directions	yes
4	4 directions	yes
5	5 directions	yes

# Inside stelow, read from workflow config; standalone defaults to 3 + hybrid.
COUNT="${EXPLORATION_COUNT:-3}"
HYBRID="${EXPLORATION_HYBRID:-yes}"
if [ "$COUNT" -le 1 ]; then
  HYBRID="skip"
fi

Direction seeds (pick the COUNT most-differentiated for the problem; never pad with near-duplicates to hit the count):

• A — Conventional build. The safest established construction for this space: boring splits, synchronous calls, single ownership. Optimize for predictability and onboarding ease.
• B — Decoupled flow. Reframe the construction around events, queues, or streams. Optimize for independence of parts and evolution speed.
• C — Intelligence-led. Use models, predictions, or realtime sync to collapse work the user would otherwise do. Accept implementation complexity only where it meaningfully improves the outcome.
• D — Radical reduction. Fewest moving parts that still deliver the committed scopes. Collapse services, merge stores, defer abstractions. Optimize for clarity and reversibility.
• E — Power-operator build. Optimize for expert throughput: CLIs, SDKs, batch paths, dense contracts, minimal ceremony. Assume the operator knows the domain.