Stackbone ships three Agent Skills that teach a coding agent how to scaffold, build, publish and operate durable agents and workflows on Stackbone. The skills are published as an open-source repository at github.com/stackbone/stackbone-skills and follow the Agent Skills Open Standard.

Every stackbone init and stackbone link installs them into the project for every coding agent the upstream skills registry can detect, so the agent driving the terminal knows the surface without the creator having to explain anything.

Why a skill?

The CLI is designed for a coding agent to drive — --json, semantic exit codes, stackbone metadata, inline stackbone docs. But for the agent to actually use those primitives it has to know they exist.

Dropping a skill in the project tree is the standard discovery mechanism: Claude Code (and the equivalent in other coding agents) reads the frontmatter description, decides whether the user's task matches, and only loads the body when it does.

This is the progressive-disclosure / find-skills pattern. It avoids both extremes:

• Don't dump the manual into every prompt — that's token-expensive and pollutes context.
• Don't make the agent guess — that wastes turns rediscovering the exit-code table and the --json envelope.

The three skills

The split is by surface area, not by coding agent target — the same three skills ship for every supported agent. Together they cover the modern Stackbone runtime: a workspace discovered by convention from its files (one folder per agent under deep-agents/, one workflows/<name>.workflow.ts per workflow), made of durable deep agents and durable Workflow SDK workflows, written against the ambient stackbone client. An optional stackbone.config.ts can override that scan, but most projects need none.

Skill	Loads when the user is…
stackbone	Writing the inside of a workspace with @stackbone/sdk: authoring a deep agent (model + tools) under deep-agents/<name>/, and writing durable workflows under workflows/<name>.workflow.ts. Knows the ambient stackbone client (database, storage, ai, rag, secrets, config, prompts), callDeepAgent() from @stackbone/sdk/workflow to call a sibling agent, stackbone.connection(id) to call a connector, and durable human-in-the-loop with requestApproval() from @stackbone/sdk/workflow.
stackbone-cli	Driving the stackbone CLI: scaffold a workspace (init) and add pieces to it (add agent, add workflow, add workflow-agent), link an existing project (link), the local runtime + Studio (dev), publish a workspace bundle (publish), and operate a running install — durable workflows (workflows), runs (runs), approvals (hitl), fake connector triggers (trigger), config + local type codegen (config), connectors (connectors), db migrations (db), whoami / current, metadata, docs. Knows the command tree, the --json envelope, exit codes and the device-flow login.
stackbone-debug	Triaging a failure: SDK { data, error } envelope codes, HTTP 4xx/5xx from the control plane, durable workflow/agent runs failing or stuck, HITL approvals parked on a hook, agent chat session/stream issues, connector (Connect) failures, slow queries, secrets decryption errors, and publish / workspace-bundle build failures. Guides diagnostic command execution, does not propose fixes.

Each skill points the agent at the next one when the topic mismatches — for example, the stackbone skill says "for CLI tasks use stackbone-cli instead". You do not pick which one to load; the coding agent does it from the prompt.

For the concepts these skills build on, see Agents & sessions, Workflows and Connect.

The skills track the published packages. Each skill teaches the surface that ships with the current @stackbone/cli and @stackbone/sdk — re-installing the skills always pins them to the versions you have. The SKILL.md linked in the table is the exact bytes a coding agent loads, so open it to see what your CLI version documents.

Source of truth

The skills are maintained at github.com/stackbone/stackbone-skills. That repository is the single source of truth — this page only explains the wiring. To audit the exact bytes a coding agent will load, open the SKILL.md linked in the table above; to file a correction, open a PR on that repository.

The published @stackbone/cli package pulls the same skills through the upstream skills registry, so re-running stackbone init or stackbone link refreshes the bundle to the latest published version.

The agent skills build on the public frameworks behind the Stackbone runtime. Read their docs to go deeper: LangChain's deep-agent framework (the durable-agent layer), the Workflow SDK (durable execution), and the Vercel AI SDK (the model layer).

Installation

You normally don't install the skills yourself — stackbone init and stackbone link do it for you (see Auto-install via the CLI below). The two standalone paths are for ad-hoc setups (e.g. importing the skills into an existing project that wasn't scaffolded with the CLI).

Using the skills registry

pnpm dlx skills add stackbone/stackbone-skills

Add -a claude-code / -a cursor / -a windsurf / -a cline / -a codex to target a specific coding agent. With no -a flag the registry installs into every supported agent it detects in the project tree.

Claude Code marketplace

/plugin marketplace add stackbone/stackbone-skills
/plugin install stackbone

Auto-install via the CLI

stackbone init and stackbone link both call the upstream skills registry as a best-effort step after the scaffold. Every coding agent the registry recognises gets the three skills, and .gitignore is patched so the per-agent caches do not get committed.

The call is intentionally best-effort: an offline machine or a broken pnpm dlx only emits a warning — init / link always complete. Re-run either command at any time to refresh the bundle.

Where the skills land in your project

The exact tree depends on which coding agents the registry detects. For Claude Code:

<project root>/
└── .claude/
    └── skills/
        ├── stackbone/SKILL.md
        ├── stackbone-cli/SKILL.md
        └── stackbone-debug/SKILL.md

Cursor, Windsurf, Cline and Codex use their own respective directories; the registry handles the placement. Refer to the stackbone-skills README for the per-agent layout.

Updating

To pick up the latest skill bundles, re-run:

stackbone link --force          # an existing project
stackbone init                  # a new project
pnpm dlx skills add stackbone/stackbone-skills  # ad-hoc, outside the CLI

Do not hand-edit the installed copies — the next install overwrites them. File corrections against the stackbone-skills repository so every consumer benefits.