hermes-webtop: Skill for Claude Code

Parallel Delegation

Always use as much subagent parallelism as possible: the main agent's job is orchestration, not execution. Every piece of work that a subagent can carry should be handed off, batched into the same turn, so the main agent never sits idle and the conversation keeps moving.

Core Principles

1. Question every task first: can a subagent do this? If yes and it is independent of what you're doing, delegate it.
2. Maximize parallelism: N independent tasks → N delegate_task calls in the same assistant turn. Never spawn one at a time and never serialize independent work — sequential delegation multiplies wall-clock time for no benefit. Prefer many small, focused subagents over one big one.
3. Stay unblocked: between spawning and results, keep doing work that doesn't depend on the subagents. Do not hold your turn waiting for a result you don't need yet.
4. Consume, don't redo: when results arrive, use them — don't re-execute the delegated work yourself.

When to Delegate

Delegate anything mechanical, long-running, parallelizable, or context-heavy:

• Batch independent audits — e.g., dependency/security/lint audits split per package or module, one subagent per scope (see Examples).
• Script runs — independent test scripts, formatting checks, or reproducible CLI commands that can run concurrently.
• Repo scans — search_files / codebase inspection over disjoint directories or areas, per-area.
• Research & extraction — web searches, page fetches, summarization, transcript-to-summary jobs.
• Anything that doesn't need your in-flight judgment — offload it so a long-running subagent finishes by the time you need its output.

Rule of thumb: if you can enumerate the units up front, fan out one subagent per unit in a single turn.

When NOT to Delegate

• Trivial one-liners and single tool calls — a file read, a simple grep, a one-command lookup. Delegation costs tokens (~20K per spawn) and a round-trip; beating it with a single tool call is faster and cheaper.
• When you already have full context — if the answer is already in your context or you're mid-way through exactly this work, finish it yourself. Re-delegating means re-transmitting everything for nothing.
• When the task is the main agent's own judgment call — decisions, tradeoffs, or steps the conversation's next move depends on immediately.
• When spawn cost exceeds the work — one quick web_search stays in-thread; ten independent searches go to subagents.