Agent Skills

Glove ships Agent Skills that teach coding agents to build with Glove the way the framework is designed: from its primitives. The most common mistake agents make with Glove is rebuilding something it already provides, such as a job queue instead of a schedule, an approval table instead of requiresPermission, or a summariser instead of compaction. Each skill leads with the parts of an agent harness and a before you build X, use Y table, then points to detailed references only when the task needs them.

The three skills

They follow the same three layers as the docs: harness, system and production.

SkillLayerUse it forSource
gloveHarnessglove-core and every capability package: the loop, tools, models, stores, runtime context, skills, hooks, subagents, the display stack, memory, forms, goals, facts, MCP, REPLs, scratchpad, VFS, working environments, browsers and sandboxes, classifiers, voice, image and video, mesh and continuum..claude/skills/glove
glove-foundrySystemFoundry applications: definitions and instances, conversations, installable capabilities, schedules, sleep and spawn, transmissions and playbooks, approvals and steering, the client, REST and OpenAI-compatible APIs, persistence and the inspector..claude/skills/glove-foundry
kuiperProductionDeploying a Foundry app on Kuiper: the CLI, what the builder infers, request authorization, the /data disk, secrets and variable templates, the model gateway, browsers, sandboxes, the inbox, logs and the Inspect card..claude/skills/kuiper

Agents that read plain text instead of skills can use /llms.txt, /llms-full.txt and /foundry/llms-full.txt, and Kuiper's kuiper.sh/llms.txt.

Install with npx

The fastest way to install the skills is with the skills CLI:

bash
npx skills add porkytheblack/glove -a claude-code

This installs the skills into your project's .claude/skills/ directory and the agent picks them up automatically. Add -g to install them for every project:

bash
npx skills add porkytheblack/glove -a claude-code -g

Manual install

Copy the skill folders you need from the repository into your project's .claude/skills/. Every link inside a skill is either relative to that skill's folder or an absolute URL, so a copied skill works outside the Glove repository.

text
your-project/
└── .claude/
    └── skills/
        ├── glove/
        │   ├── SKILL.md                       # the parts of a harness, the primitives table, rules, decision guide
        │   └── references/
        │       ├── core.md                    # the loop, stores, subscribers, providers, server-side, inbox
        │       ├── tools-and-display.md       # tools, ToolResultData, summaries, the display stack
        │       ├── react-next.md              # GloveClient, useGlove, <Render>, createChatHandler
        │       ├── extensions.md              # hooks, skills, subagents
        │       ├── memory-and-workflows.md    # memory, forms, goals, facts, preparation
        │       ├── mcp.md                     # glove-mcp
        │       ├── execution-and-workspace.md # scratchpad, REPLs, VFS, working environment, browsers, sandboxes
        │       ├── classifier.md              # glove-classifier
        │       ├── voice.md                   # pipelines, realtime speech, avatars, LiveKit
        │       ├── media.md                   # image and video workflows
        │       ├── multi-agent.md             # mesh and continuum
        │       ├── packages.md                # every package in one table
        │       ├── api-reference.md           # exact signatures
        │       ├── examples.md                # patterns from the example apps
        │       └── gotchas.md                 # the full list of pitfalls
        ├── glove-foundry/
        │   ├── SKILL.md                       # definition / instance / run, the system primitives table, layout, persistence
        │   └── references/
        │       ├── condensed-reference.md     # fields, endpoints, CLI flags, built-in tool inputs
        │       ├── building.md                # authoring rules for apps, playbooks, schedules, guidance, voice, multi-agent
        │       └── contributing.md            # only for changing glove-foundry itself
        └── kuiper/
            ├── SKILL.md                       # deploy, production-ready Foundry, variables, models, inbox, sandboxes, operate
            └── references/
                └── cli.md                     # every kuiper command and flag

What the skills teach

  • The parts of a harness. Loop, tools, context, state, workspace, surface and team for glove; identity, time, arrival, oversight and access for glove-foundry; run, keep, reach and watch for kuiper. Each part comes with the question it answers, so an agent can place a request before it writes code.
  • Before you build X, use Y. Every primitive from the primitives map, phrased as the thing an agent is about to hand-build, with the API and package that already does it.
  • The rules the primitives follow. Everything the model does is a tool; adapters at every boundary; mount, don't wrap; live state at the tail of the prompt; work in files and code; people in the loop through primitives; credentials never in the model.
  • Decision guides. glove-core or Foundry; plain tools, a REPL, the scratchpad or a working environment; which way to collect information; which surface.
  • The rules that prevent silent bugs. Such as the Displaymanager casing, event key differences, permissions keyed on tool and input, goals living in glove-memory/goals, and every Foundry run being a fresh process that forgets an in-memory store.
  • Deep references on demand. Full API signatures, worked patterns from the examples, and per-package guides, read only when the task needs them.

Using the skills

Once installed, the skills activate when your coding assistant sees Glove, Foundry or Kuiper code, or when you ask it to build an agent. You can also load one explicitly in Claude Code:

bash
/glove
/glove-foundry
/kuiper

Example prompts

  • “Build a coding agent that works in a sandbox and asks before it writes files”
  • “Give this agent a reminder it can set for itself every weekday morning”
  • “Wake an agent per customer when a ticket arrives, and have a person approve refunds”
  • “Keep the user's preferences in front of the model on every turn without touching the system prompt”
  • “Deploy this Foundry folder on Kuiper with its state on the disk and models through the gateway”

How user-invoked skills land in history

The skills above are install-time bundles for your coding agent. The runtime defineSkill system inside glove-core — the /skill-name directives users type into chat — has a behavior worth highlighting because it shows up in your stored message history.

When a user types something like "/research-mode tell me about ribosomes", Glove parses the directive token but does not strip it from the message. Instead, the bound /name token is replaced in the persisted user text with a non-triggerable placeholder of the form [invoked_extension__skill_<name>] for skills, and [invoked_extension__hook_<name>] for hooks. The skill or hook handler runs as expected; the persisted user message records exactly which extensions fired.

text
User typed:
  "/research-mode tell me about ribosomes"

Persisted user message text:
  "[invoked_extension__skill_research-mode] tell me about ribosomes"

Skill handler receives:
  ctx.parsedText = "[invoked_extension__skill_research-mode] tell me about ribosomes"
  ctx.source     = "user"
  ctx.args       = undefined

The placeholder is intentionally non-triggerable — re-parsing it does not re-bind to the skill — so historical messages can be replayed safely without firing the extension a second time. UIs that render the transcript should detect the placeholder and either substitute a chip showing “skill: research-mode” or strip it entirely; the agent reads it verbatim and infers context from it.

Synthetic user messages that the skill itself injects (the handler's return value) are persisted with is_skill_injection: true so transcript renderers can style them differently from real user turns. See Hooks, Skills & Subagents for the full defineSkill / defineHook / defineSubAgent surface.

Updating the skills

Installing or releasing npm packages does not update installed skill copies. Re-run the install command to fetch the latest versions:

bash
npx skills add porkytheblack/glove -a claude-code