Pack № 05 · Build

Agent Builder

Write skills that fire, build MCP servers, fan work out to subagents.

Who it's forPeople who extend their agent instead of only using it. You write your own skills, wrap an API in an MCP server, keep AGENTS.md short and split big jobs across subagents. This pack covers writing a skill and measuring it, checking someone else's skill before you install it, and running long work across sessions without losing the thread.
npx skills add anthropics/skills@e4bb5d960d87126ff3b3a8af1a104b7db67560ba --skill skill-creator mcp-builder -g -y && \
npx skills add getsentry/skills@a582c51ee29d0f35e6ac321340796f77d43da733 --skill skill-scanner agents-md -g -y && \
npx skills add obra/superpowers@22f0b463e7a5e579904dd6211a520321035526f7 --skill dispatching-parallel-agents -g -y && \
npx skills add muratcankoylan/Agent-Skills-for-Context-Engineering@d9202c76d6dc8351e648badb576a6e9b5aae1e7f --skill multi-agent-patterns -g -y && \
npx skills add trailofbits/skills@149d7b21f4067765b82e67029730b42d228ef4b4 --skill goal-prompt -g -y && \
npx skills add gastownhall/beads@78624109904655d7a4609f27d34eb0982cd8963f --skill beads -g -y && \
npx skills add JuliusBrussee/caveman@a192894723a8f47e7f504c7b5041967ff1f18e55 --skill caveman -g -y

Fetches 9 skills from 7 author repos, each pinned to the version we reviewed. Nothing is mirrored. Uses the open-source skills CLI; needs Node.js 22.20 or newer.

What's inside

9 skills. Each card links to its author's repository at the exact version we reviewed.

  1. Skill creator 1/9 Official Skill authoring Needs app Drafts a skill with you, runs test prompts with and without it, shows the results side by side in a browser viewer and tunes the description so the skill triggers. @anthropics Apache-2.0

    Skill creator

    The place to start a skill of your own. It compares runs with and without the skill, so you see whether it helps before you share it.

    Needs an app or CLI Official · Anthropic Apache-2.0
    Source anthropics/skills at e4bb5d9
  2. Skill scanner 2/9 Skill security Needs app Checks a skill before you adopt it: a bundled static scan plus a guided read for prompt injection, risky scripts, broad permissions, leaked secrets and supply chain risk. @getsentry Apache-2.0

    Skill scanner

    Skills are instructions your agent will follow. This reads one the way an attacker would before you install or publish it.

    Needs an app or CLI Community · Sentry Apache-2.0
    Source getsentry/skills at a582c51
  3. MCP builder 3/9 Official MCP servers Needs app Guides the agent through building an MCP server in TypeScript or Python: plan the tools, implement them with schemas and clear errors, test, then write evaluations. @anthropics Apache-2.0

    MCP builder

    MCP servers are the other half of extending an agent. Written by Anthropic, where the protocol started, with guides for both SDKs.

    Needs an app or CLI Official · Anthropic Apache-2.0
    Source anthropics/skills at e4bb5d9
  4. AGENTS.md maintainer 4/9 Agent instructions Ready Writes and trims AGENTS.md so it stays under 60 lines: verified commands, non-obvious conventions and links to existing docs instead of copied prose. @getsentry Apache-2.0

    AGENTS.md maintainer

    Every agent you build reads the project instruction file first. This keeps it short, checked against the repo and in one place.

    Works right away Community · Sentry Apache-2.0
    Source getsentry/skills at a582c51
  5. Dispatching parallel agents 5/9 Subagents Ready Decides when failures or tasks are independent enough to split, then sends one focused subagent per problem in parallel and checks the results for conflicts. @obra MIT

    Dispatching parallel agents

    The practical side of subagents. When to split work, how to brief each agent and how to merge what comes back.

    Works right away Community · Jesse Vincent MIT
    Source obra/superpowers at 22f0b46
  6. Multi-agent patterns 6/9 Agent architecture Ready Design reference for multi-agent systems: when extra agents are justified, supervisor, swarm and hierarchical layouts, context isolation, consensus and failure handling. @muratcankoylan MIT

    Multi-agent patterns

    The design side. Whether you need several agents at all, and which layout fits, before you build it.

    Works right away Community · Muratcan Koylan MIT
    Source muratcankoylan/Agent-Skills-for-Context-Engineering at d9202c7
  7. Goal prompt 7/9 Long-running tasks Needs app Drafts a one-line /goal command for goal mode: a checkable end state, the command that proves it, what must not change and a stop bound, with missing parts flagged. @trailofbits CC BY-SA 4.0

    Goal prompt

    A long unattended run depends on its stop condition. This drafts one that names the proof and closes the cheap ways out.

    Needs an app or CLI Community · Trail of Bits CC BY-SA 4.0
    Source trailofbits/skills at 149d7b2
  8. Beads task memory 8/9 Official Task memory Needs app Teaches the agent the bd issue tracker: find unblocked work, claim it, write notes that survive context compaction and close it, with dependencies kept across sessions. @gastownhall MIT

    Beads task memory

    Work that spans sessions and agents needs memory outside the context window. Beads keeps tasks and blockers in a tracker the agent drives.

    Needs an app or CLI Official · Steve Yegge MIT
    Source gastownhall/beads at 7862410
  9. Caveman 9/9 Reply style Ready Makes the agent answer first and drop greetings, hedging and recaps for the whole session, while code, commands, numbers and error text stay exact. @JuliusBrussee Apache-2.0

    Caveman

    A voice for the whole session. Answer first, no ceremony, every fact exact. Handy on long runs, and a compact example of a skill that changes behaviour with instructions alone.

    Works right away Community · Julius Brussee Apache-2.0
    Source JuliusBrussee/caveman at a192894

Good to know

What you bring

  • Python 3 for the eval, viewer and packaging scripts
  • Claude Code for the description optimiser, which runs claude -p
  • uv to run the bundled Python scanner
  • Node.js or Python to build and run the server
  • Optional: Anthropic API key for the bundled evaluation script (paid)
  • uv to run the bundled formatter script
  • bd CLI, version 0.60 or newer
  • A one-time bd init in your project, run by you and not by the agent

Pick one, not both

  • Skill authoring: writing-skills from Superpowers and writing-for-agents by Matt Pocock overlap with skill-creator. This pack uses skill-creator for its runs with and without the skill; do not add a second authoring skill.
  • Always-on style: Caveman and Ponytail both set rules for every reply in a session. This pack uses Caveman because it shapes the replies and leaves your code alone; do not run both.
  • Session handoff: the handoff skill by Matt Pocock writes a note for the next session. It is left out because this pack takes its subagent skill from Superpowers and does not mix two workflow sets. Beads covers work that spans sessions.
  • Task memory: Beads needs the bd CLI. If you do not want to install it, drop Beads and do not put a second tracker skill in its place.
Review slipPack № 05
  • Every skill file read in full
  • Pinned to a reviewed version
  • Author and licence on every card
  • Requirements and costs disclosed
  • No sponsored skills inside
Last reviewed
Nov 2026
Reviewed