Pack workflow design

What a pack is

A pack is a cohesive set of skills for a role’s work — all the workflows one kind of specialist needs, bundled as a single installable unit. Before you write your first SKILL.md, work through the five steps below. Each one produces a concrete decision that feeds the next.

The design vocabulary throughout is the session arc: the five stages every sustained professional tool must support —

• Arrive — start a new session
• Orient — understand where you are and what comes next
• Work — drive the core workflow
• Persist — save durable artifacts across sessions
• Collaborate — hand off to the next stage or pack

The arc is the skeleton your skills hang on. Not every pack covers every stage; the decision tree in Step 1 tells you which stages apply.

Step 1 — Characterize your workflow type

Walk this decision tree to identify your pack’s type:

Does the pack maintain a work-in-progress thread across sessions —
tracking phases, a growing project, or cumulative state?
│
├─ No → Does each invocation produce a standalone artifact
│        that the user takes away?
│        ├─ Yes → Episodic
│        └─ No  → Stateless
│
└─ Yes → Does this pack create and own the project thread,
         or derive from a thread another pack owns?
          ├─ Own    → Sustained-project
          └─ Derive → Sustained-derived

Type	Shape	Typical session
Episodic	No persistent state between sessions	Each skill invocation is standalone; a completed artifact is the output
Sustained-project	Creates and manages a durable project	Session opens a thread; a status skill orients the return; pack drives the thread to completion
Sustained-derived	Reads from and builds on a project another pack owns	Session reads from a source project; adds derived artifacts
Stateless	Pure workflow transformation; no file-system output	Input → transformation → output; nothing persists

Output from this step: your pack type. Write it down before moving to Step 2.

Step 2 — Map the arc for your pack

For each arc stage, answer the guiding question. Your answers define which skills your pack needs and which it can safely skip.

Arrive — “What does a user type to open this pack’s first session?” Name the skill that kicks things off. For a sustained-project pack, this is your *-project-start skill. For an episodic pack, it is the first substantive skill — there is no session state to initialize.

Orient — “When a user returns after a break, what do they need to know?” If the answer is non-trivial (where a project stands in its lifecycle, what was decided, what comes next), your pack needs a *-status skill. Episodic packs skip this stage: there is no persistent thread to orient to.

Work — “Which one or two skills are the core workflows this pack enables?” Name them. These are the skills that do the pack’s primary work — the reason a user installs it.

Persist — “Does the pack accumulate a project directory across sessions?” If yes, design the vault-path shape in Step 4. An episodic pack may write a standalone artifact per invocation, but it does not build up a shared project directory over time — skip Step 4 for episodic and stateless packs.

Collaborate — “What does the next pack in the user’s workflow consume from this one?” Name the artifact type and file shape another pack might read. A sustained-derived pack depends on this being well-defined by the pack it derives from.

Output from this step: an arc map — one row per stage, naming the skill (or “none”) and the artifact (or “none”).

Step 3 — Name your skills

Skill activation is description-driven — users invoke skills by natural language, not exact names. But names appear in the catalogue and in SKILL.md frontmatter, and a consistent verb vocabulary lets users predict what to type.

Use the five-verb taxonomy to pick the right name:

Verb	Meaning	Activation phrasing
status	Orient — “where am I / what’s next?”	Cold-start phrases, “what’s on today”, “orient me”
start	Create/begin a sustained project	“start a research project”, “kick off an investigation”
check	Quality/health read — “is it good / saturated / done?”	“is this ready”, “should I keep gathering”
init	Repo-scaffold only	init-project, adapt-to-project; cf. git init
resume	Return to prior work	Activation phrase — not a skill name

Banned as skill names: arrive, orient, onboard, return, onboarding — these are UX-stage labels, not user-facing commands. Using them as skill names causes activation collisions with the arc vocabulary users already know.

For the full taxonomy table and naming rules, see How to author a skill.

Output from this step: a name for each skill in your arc map.

Step 4 — Decide your vault-path shape

If your pack maintains a persistent project directory (sustained-project or sustained-derived), design the vault-path now. Skip this step if your pack is episodic (standalone output per invocation, no shared project state across sessions) or stateless.

One output_dir base per pack. Configure it once in agentbundle-layout.toml:

[your-pack-section]
output_dir = "~/your-output-dir"

Each pack has its own config section. Do not share a section with another pack.

Skill-specific subdirectories under the base. Each skill that writes files owns one subdirectory under output_dir. Do not write all artifacts flat into the base — it makes a status skill’s scanning harder and creates namespace collisions as the pack grows.

The journey-mapping skill is the canonical example: it writes customer-journey artifacts to <output_dir>/journeys/. Other experience-design skills each own a sibling directory (screens/, blueprints/). The experience-status skill is the one skill that scans across all subdirectories — it reads the vault, not a single subpath.

Output from this step: a directory layout diagram for your pack’s output.

Step 5 — Register with workspace-status

If your pack has a *-status skill, register it so that workspace-status can route to it when a user cold-starts a session.

How to register:

1. Choose a shaping_queue type string for your pack. This string appears in workspace.toml entries for your pack’s shaping items (for example, design for the experience-design pack, research for the desk-research pack).

2. Add a routing entry in the workspace-status skill’s routing table: type = "<your-type>" → <your-pack>-status.

3. Add a fallback. When your pack is not installed, workspace-status skips the routing entry rather than erroring. Document the fallback message in your *-status skill body — it tells the user what to run to get started (for example, when the experience-design pack is not installed, workspace-status falls back gracefully rather than routing to experience-status).

If your pack is episodic or stateless, skip this step. There is no persistent thread to orient to, and a status skill for an episodic pack would have nothing to read.

Output from this step: a shaping_queue type string and a routing entry.

Reference: worked archetypes

Episodic — product-strategy

Each skill invocation is standalone. A product-strategy session runs a single skill — market analysis, competitive positioning, or user persona — and produces a complete artifact. There is no cross-session thread to track.

Arc map:

Stage	Skill	Artifact
Arrive	Any skill (e.g., market-analysis)	—
Orient	None	—
Work	Whichever skill the user invokes	Standalone analysis artifact
Persist	Each skill writes a per-invocation standalone artifact	Strategy document (not accumulated across sessions)
Collaborate	—	Artifact feeds adjacent workflows

No status skill needed — each invocation is complete; there is no thread to return to.

Other catalogue packs that follow the episodic pattern: architect (each architecture review is standalone) and converters (each file conversion is standalone).

---

Sustained-project — desk-research

The desk-research pack creates and manages a research project that spans multiple sessions. A project opens with desk-research-project-start, advances through capture → digest → synthesize → feedback phases, and closes when the stop signal fires.

Arc map:

Stage	Skill	Artifact
Arrive	desk-research-project-start	Creates overview.md; initializes the project
Orient	desk-research-project-status	Reads overview.md; reports phase and working hypothesis
Work	desk-research, desk-research-project-check, desk-research-project-digest	Source captures, saturation check, synthesis
Persist	All writing skills	Files under <output_dir>/ (configured via [research] output_dir)
Collaborate	—	Synthesized research feeds product-strategy or experience-design

The status skill reads overview.md at the configured output_dir. When no project exists, it surfaces a not-started message and points to desk-research-project-start.

---

Sustained-derived — experience-design

The experience-design pack reads from customer-journey artifacts and derives screen flows, service blueprints, and experience assessments. It builds on what journey-mapping creates; it does not create the initial project from scratch.

Arc map:

Stage	Skill	Artifact
Arrive	journey-mapping (creates the first journey artifact)	Customer-journey map in <output_dir>/journeys/
Orient	experience-status	Reads frontmatter across journeys/, screens/, blueprints/; reports steel-thread completion
Work	user-flow, service-blueprint, other experience skills	Screen flows, blueprints
Persist	All writing skills	Files under <output_dir>/journeys/, screens/, blueprints/
Collaborate	—	Derived artifacts feed content-design and development

When [design] output_dir is not configured, experience-status surfaces a not-configured message and points to journey-mapping, which sets the path on first run.

---

Stateless — reserved category

The stateless type is reserved for future packs whose shape is pure transformation: input arrives, output leaves, nothing persists. No current catalogue pack fits this classification.

The converters pack and the architect pack are both episodic — each invocation produces a standalone artifact the user receives as output, not a transformation that feeds back into a pipeline without producing a visible artifact.

If you are designing a pack that genuinely fits the stateless shape — pure transformation, no file output, no session state — this is the right type. Verify first: if each run produces a file the user saves, you are episodic, not stateless.