Skills DirectorySkills Directory
SkillsLearnSecurityCategoriesDocsBlogPro
Sign InSubmit Skill
Skills Directory

Security-tested agent skills for Claude, coding agents, and AI workflows.

Directory

  • Browse Skills
  • All Skills A–Z
  • Claude Skills
  • Claude Code Skills
  • Agent Skills
  • Categories
  • Authors
  • Submit a Skill

Learn

  • Learn Hub
  • Install Claude Skills
  • Write SKILL.md
  • Skills vs MCP
  • Directories Compared

Security

  • Security
  • Methodology
  • Secure Claude Skills
  • Security Badges
  • Chrome Extension
  • Skill Manager

Company

  • About
  • Community
  • Blog
  • API Docs
  • Advertise

2026 Skills Directory. All rights reserved.

ProTermsPrivacyRefunds
Back to skills

Swarm

ASecurity

Run bounded parallel coverage or races, drain all started workers, verify their evidence, and consolidate outcomes and gaps into one report.

2 stars
0 votes
0 copies
1 views
Added 10/1/2026
ai-agentsgo

Works with

terminal

Security Analysis

A100/100

Scanned 10/6/2026

$npx -y skills add williamwue/oh-my-stack --skill swarm --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Swarm?

Add the live security badge to your README — it updates automatically with every re-scan.

Security grade badge for Swarm
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/williamwue-swarm/badge)](https://www.skillsdirectory.com/skills/williamwue-swarm)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
Files
SKILL.md
---
name: swarm
description: "Run bounded parallel coverage or races, drain all started workers, verify their evidence, and consolidate outcomes and gaps into one report."
disable-model-invocation: true
---

# Swarm

## Child session handoff

Read [the handoff contract](../poteto-mode/references/subagent-handoff.md).
New tasks, repair rounds, retries, and queue items use fresh child sessions
with the original brief, every later directive, prior findings and responses,
and unresolved objections. Reuse only for required costly live state, and
only when the host allows it. Stop and fence active writers before replacement.
A host-owned orchestrator's model catalog, workspace binding, child tools,
and review-round rules take precedence over the native binding above.
Keep its task handles and attribution receipts. Do not use a backing child
conversation as a new delegated review, or claim native-record verification
for a host-owned child. Report attribution evidence gaps explicitly.

Coordinate one bounded fan-out and return one report. This is not a standing
program, background monitor, or permission to publish. Keep a checklist: frame,
fan out, aggregate, report.

## Frame

State the done predicate, required result, and allowed actions. Choose coverage
(distinct required slices), race (same brief), or mixed (races within slices).
For each race declare first pass, rank all, or best-of plus measurable selection
criteria before dispatch. Define how a PASS is verified, not just self-reported.

Set total worker count and maximum concurrency separately, with an attempt or
time budget. Every required slice must have an owner. For mixed work, define
the selection rule per slice and require all slices for overall completion.
Use `swarm.workers` from the current resolution manifest as the default
worker route; an explicitly named model for a race arm takes precedence after
runtime validation. Resolve available roles without inventing models or hosts.
Local-only dependencies remain local; do not upload workspace data to a remote
worker merely because the runtime supports it.

Give each writer its own isolated output and explicit resource ownership. Check
the starting revision and any necessary local changes. If isolation is missing,
use read-only reports or patch proposals for overlapping scopes; only serialize
writes when the requested task permits sequential work. Never run conflicting
writers on the same resource.

## Fan out and drain

Each standalone brief states goal, shared grounding, exact slice or race arm,
ownership, allowed tools/actions, verification, and a PASS / ISSUES / BLOCKED
report with evidence. When a worker verifies a change, name its exact base and
head revisions. When it measures a result, also freeze the method: sample
count, what constitutes one sample, and sample order. Require the worker to
repeat those coordinates and the method in its result. Dispatch within the
concurrency limit and track native worker handles. Do not count a launch
acknowledgment as a terminal result.

Read and drain every started worker, including losers in a first-pass race.
A verified first-pass result may select a winner early, but cannot justify
abandoning active writers. If safe cancellation is supported, request it and
confirm terminal state. Otherwise wait within budget, or report outstanding
handles and incomplete drainage without claiming completion. Never fabricate
cancellation or terminate unrelated tasks.

If parallel dispatch is unavailable, use bounded sequential workers and say so.
If delegation is unavailable, run separated root passes only for work that can
honestly be done that way. A requested genuine race or independent coverage
remains unverified; do not relabel root passes as workers.

## Aggregate and report

Inspect actual terminal artifacts and root-check critical evidence before
accepting PASS. A dropout leaves a coverage gap; assign a bounded retry or mark
that slice BLOCKED. Race winners must meet the predeclared rule; no passing arm
means no winner. A result missing required revisions or measurement method is
not a PASS: start a fresh worker once with the same frozen contract, including
prior findings and unresolved objections, then record a
gap if it still omits them. A worker with a proven defect reports ISSUES with
every defect it can substantiate, not just the first. Keep failed or
contradictory evidence visible.

Apply [the handoff contract](../poteto-mode/references/subagent-handoff.md)
when retrying or replacing a worker. Fence active writers before replacement.

Return one table of slice/arm, actual worker identity, status, evidence, and
remaining gap, followed by concise findings and the applied selection rule.
Report overall completion only when every required slice passes and all started
workers are drained. Do not paste raw worker dumps or integrate competing code
implicitly; use [arena](../arena/SKILL.md) when synthesis is the requested task.

Attribution

williamwuewilliamwue
View sourceSee grades on GitHubMore from williamwue →
SSkills DirectorySkills Directory

Ship a skill? Prove it's safe.

Free 120-pattern security scan, letter grade, and an embeddable README badge.

Submit a skill

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments (0)

No comments yet. Be the first to comment!

SSkills DirectorySkills Directory

Ship a skill? Prove it's safe.

Free 120-pattern security scan, letter grade, and an embeddable README badge.

Submit a skill

Related Skills

Caveman

Terse caveman voice: answer first, fluff gone, every technical fact kept. Use for /caveman, "caveman mode", "talk like caveman", "be brief", "less tokens". Stays on until "stop caveman" or "normal mode".

1100021 votes

Hyperplan

Adversarial multi-agent planning skill. Self-orchestrates 5 hostile category members (unspecified-low, unspecified-high, deep, ultrabrain, artistry) via team-mode for ruthless cross-critique debate, distills only the defensible insights, then MANDATORILY hands the distilled insight bundle to the `plan` agent for executable plan formalization. Use when planning needs maximum rigor and surfacing of weak assumptions, blind spots, and over-engineering. Triggers: 'hyperplan', 'hpp', '/hyperplan', ...

698431 votes

Writing Skills

Create and manage Claude Code skills in HASH repository following Anthropic best practices. Use when creating new skills, modifying skill-rules.json, understanding trigger patterns, working with hooks, debugging skill activation, or implementing progressive disclosure. Covers skill structure, YAML frontmatter, trigger types (keywords, intent patterns), UserPromptSubmit hook, and the 500-line rule. Includes validation and debugging with SKILL_DEBUG. Examples include rust-error-stack, cargo-dep...

3931 votes

Mcp Code Execution

Routes multi-tool workflows through MCP servers for large datasets and pipelines. Use when Bash tool overhead is limiting throughput on data-heavy tasks.

3421 votes

catchup

Recovers the conversation and failed tool calls of a previous Codex, Amp, Claude Code, Antigravity, Cline, Copilot CLI, Cursor, DeepSeek Harness, Grok Build, Kimi, OpenCode, Pi Agent, or ZCode session. Use when the user says "catch up", "what did the last session do", "get me up to speed", "I switched agents", asks to recover/summarize a previous session before continuing, or asks to diagnose or report a catchup failure. Do NOT use for the current conversation, git history, or any non-agent log.

741 votes
View all in ai-agents →