A manual-invoke orchestration gate for AI coding agents. Before you spin up a multi-agent workflow, it answers a prior question: should this even be a workflow?
workflow-triage returns a structured verdict on the lightest credible execution mode
for a task — and stops there. No scripts, no launches. Its whole value is restraint:
knowing when not to orchestrate.
Agentic harnesses make it cheap to fan out dozens of subagents. That's powerful for genuinely broad work and wasteful — sometimes harmful — for everything else. Over-orchestration burns tokens and removes human judgment from decisions that needed it. Under-orchestration leaves big work cramped in a single context. This skill is the gate that picks the right altitude, biased toward the lighter option.
| Verdict | When |
|---|---|
inline |
Small, sequential, low-risk; one context handles it. |
few-subagents |
Small multi-perspective pass (2–4 angles), merged once. |
dynamic-workflow |
Breadth (5+ independent units), repeatable per-unit stages, or enforced verification. One-shot fan-out now. |
loop |
Depends on external state changing over time (CI, deploy, queue, prices, inbox). |
hybrid |
A sequenced mode — one mode safely unlocks another (e.g. scope-lock → workflow). Often safer, not heavier. |
do-not-automate |
A protection verdict — decision-heavy, risky, financial, production-sensitive, or taste-led work where the human judgment is the work. |
Gates first, then the lightest credible mode.
- Apply hard gates. If one trips, return the gate-appropriate verdict (usually
do-not-automate,inline, orhybrid). - If none trips, score five dimensions and pick the lightest mode that fits:
inline → few-subagents → loop / dynamic-workflow → hybrid.
Gates override scores. A high breadth score does not earn a workflow if the task mutates production, is a financial decision, has an undefined "done", or is being orchestrated to dodge a hard call. Scores inform; gates decide.
DECISION ORDER
1. gates → if tripped, stop at the gate verdict
2. else → lightest credible mode on the ladder
LADDER (lightest → composed)
inline → few-subagents → loop / dynamic-workflow → hybrid
(do-not-automate is OFF the ladder — it's a protection verdict)
DIMENSIONS (0–5, inform only)
breadth · stage-depth · verification-need · external-state · human-sensitivity
GATES (override scores)
single-file · sequential · undefined-"done" · production-mutation
financial-decision · strategy/taste · unbounded-scope · dodging-a-decision
The skill always responds in a fixed structure: Verdict · Confidence · Why · Fit Scores (table) · Recommended Execution Shape · Human Checkpoints · Token/Complexity Estimate · Stop Rule · Next Step. Verdict and shape only — never a runnable script.
Copy this folder into your agent's skills directory:
cp -r workflow-triage ~/.claude/skills/
Then invoke it manually when scoping work — e.g. "triage this", "should this be a workflow?", "inline or subagents?". It does not trigger proactively, scan for opportunities, or run on a schedule.
SKILL.md— always-loaded core: procedure, verdicts, rubric, gates, output template.REFERENCE.md— on-demand depth: scoring anchors, gate reasoning, tie-break logic, and worked examples with filled scorecards.
Early release. Eval-validated against its designed case set; real-world dogfooding is ongoing and improvements land gradually.
MIT.