Skills — Claude Code plugin
A toolkit for running large, multi-step tasks across a team of specialized subagents that hand off
work verbatim — the workstream workflow, a researcher that profiles real-world roles, and a
write-an-agent skill that authors new agents from a shared house template.
The core workflow is the workstream. Instead of an orchestrator that reads each subagent's output
and passes along a summary (lossy), every participant appends its full work to a shared, append-only
thread.md. The orchestrator only routes — it reads the latest "Next steps" line to decide who
runs next and passes a directory path, never a summary.
How it works
A workstream is one directory:
<WORKSPACE>/.work/<Ymd-His>-<slug>/
├── brief.md # the request, context, expected outputs
├── thread.md # append-only log — the verbatim handoff medium
├── PROTOCOL.md # the subagent contract (travels with the workstream)
└── <other files> # research, generated code, test output — referenced, never duplicated
- Start or resume by invoking the skill. Pass a task description and the agent proposes a
slug and seeds
brief.md from it; pass a bare new slug to scaffold an empty brief; pass an
existing slug to reopen a workstream (e.g. after a session restart).
- Fill in
brief.md — request, context, definition of done.
- Orchestrator loop: read the latest
### Next steps in thread.md → spawn the named
@role subagent with just the directory path → it reads PROTOCOL.md + brief.md +
thread.md verbatim, does its part, and appends one entry routing to the next @role.
Repeat until a step routes to → @done or → @user.
Each thread.md entry is ## <n> — <Title> · @<role> · <Ymd-His>, a verbatim body, then a
### Next steps list where every item ends in → @role / → @user / → @done.
A workstream is opened only by invoking the skill, and resumed by invoking it again with an existing
slug. When the user asks, parts get consolidated into durable docs (ADRs, PRDs, specs) in their
real home; .work/ stays scratch (gitignored).
Bundled agents
researcher — an expert, general-purpose research subagent: deep multi-source investigation
(codebases, libraries, APIs, web, design trade-offs) returning source-cited findings with explicit
confidence and open questions. It also carries a role / job-requirement specialization that
produces a Competency Dossier for a real-world role — the input the write-an-agent skill builds
an agent from. Usable on its own from any task, and slots into a workstream as the @researcher role
(it reads PROTOCOL.md, writes findings to their own file, and appends a thread entry).
Building the fleet
The write-an-agent skill authors new role agents from one house template
(skills/write-an-agent/templates/agent.md), so every agent shares the same structure — routing
guards, lean responsibilities, self-contained references, and workstream awareness. For a role
modeled on a real-world job, it first asks @researcher for a Competency Dossier, then writes
the agent from it.
Competency Dossiers persist as durable artifacts under research/<role-slug>.md — named by
discipline, not persona (a project extending the plugin keeps its own in .claude/research/). The
agent's ## Expertise section synthesises the dossier and references it by path, so the deep
knowledge has a single home and the agent stays lean. Routing between agents is left to Claude Code's
description-based discovery — agents declare what they don't do by capability and hand work back to
the orchestrator; they never hardcode peer names.
Install on Claude Code
# From GitHub:
/plugin marketplace add lucasvscn/skills
/plugin install skills@lucasvscn
# Or from a local clone:
/plugin marketplace add ./path/to/this/repo
/plugin install skills@lucasvscn
After install, the bundled skills are model-invoked automatically from their descriptions —
workstream (multi-agent / verbatim-handoff / cross-session-resume requests), write-an-agent
(creating a new role agent), and git-commit (before committing) — and can be run explicitly as
/skills:workstream, /skills:write-an-agent, and /skills:git-commit. The bundled researcher
agent becomes available as a subagent type.
Complementary skills
Matt Pocock's skills pair well with this plugin — in
particular it ships caveman, the ultra-compact writing mode that this plugin's git-commit skill
uses (when available) to draft a terse, filler-free commit body. Install it with:
npx skills@latest add mattpocock/skills
Then run /setup-matt-pocock-skills to finish setup.
Layout
