creating-processes

name	creating-processes
description	Use when creating a new PAS process from a user's goal description. Invoked by the PAS router, not directly by users.

Creating Processes

Create a complete process definition from a user's goal. A process defines WHAT needs to happen, in WHAT ORDER, to achieve a specific GOAL. It assigns work to agents, defines phase gates, and manages flow.

Execution Framing

This skill IS the execution framework. When generating plans for process creation:

• Do NOT produce a standalone task list. Every step is a step within THIS skill's workflow.
• If you are in plan mode, exit plan mode first — this skill requires interactive brainstorming with the user via AskUserQuestion, which plan mode prevents.
• If a step requires work not covered by this skill, flag it as a PAS gap rather than a standalone manual step.

Workspace Convention

When creating a process, the workspace reflects the running process (PAS), not the artifact being created:

• Correct: .pas/workspace/pas/create-songwriting-process/
• Incorrect: .pas/workspace/songwriting/create-process/

The workspace slug should describe the task: create-{name}-process. The process field in status.yaml is always pas (since the PAS orchestrator is running).

What Is a Thin Launcher?

A thin launcher is the .claude/skills/{name}/SKILL.md file generated by pas-create-process. It contains no logic — only Read directives pointing to the real process definition, lifecycle protocol, and orchestration pattern. Its purpose is to register the process as a Claude Code slash command (e.g., /songwriting) that loads the full process on invocation.

Workflow

1. Clarify the Goal

Never assume you understand what the user wants. Ask clarifying questions until the user confirms.

• Ask one question at a time, brainstorming-style
• Probe for: scope, quality expectations, input format, output format, audience
• Continue until you can state the goal back in a single sentence the user confirms
• If the goal maps to an existing process, suggest modifying it instead of creating new

2. Prepare Reference Material (if applicable)

If the process requires domain knowledge from raw source material (transcripts, documentation, course content):

1. Create .pas/processes/{name}/reference/ directory
2. Store the original source material in reference/source/ — this is the authoritative knowledge base
3. Analyze the source material to determine the best reference format:
   • If already well-structured: use directly, no distillation needed
   • If raw/unstructured (e.g., transcripts): distill into a structured methodology doc alongside the source
   • Match the format and depth to the material — do not impose arbitrary length limits
4. Any distilled reference supplements the source material — it does not replace it
5. Skills must trace techniques back to the source. When a reference doc is insufficient, agents consult the original source material directly.

Skip this step if the process is based on general knowledge or user-provided specifications.

3. Design Phases

Break the goal into sequential phases. For each phase define:

• Input: what files/data this phase needs (from user or previous phases)
• Output: what files/data this phase produces
• Gate: what review point exists (user approval, orchestrator check, or none)

Parallelism: infer from I/O dependencies. Phases sharing the same input but not depending on each other can run in parallel. Phases listing another phase's output as input must wait. No explicit depends_on needed. Optional sequential: true at process level to force linear.

4. Determine Agents

Start with the minimum viable set. Every process needs an orchestrator. Add specialist agents only when:

• A phase requires distinct expertise (research vs writing vs verification)
• Quality feedback suggests a specialist would outperform the orchestrator
• The phase is complex enough to warrant a dedicated agent

For simple processes (1-3 phases, similar skills), the orchestrator handles everything (solo pattern).

5. Select Orchestration Pattern

Read ${CLAUDE_PLUGIN_ROOT}/library/orchestration/SKILL.md for the decision matrix. Then apply it:

Agents	Discussion needed?	Parallel phases?	Pattern
1	N/A	N/A	solo
2+	Yes	N/A	discussion
2+	No	Yes	hub-and-spoke
2+	No	No	sequential-agents

Default to hub-and-spoke when unsure.

6. Generate Process

Run the generation script with all decisions from steps 1-5:

bash ${CLAUDE_SKILL_DIR}/scripts/pas-create-process \
  --name {process-name} \
  --goal "{one-sentence goal}" \
  --orchestration {pattern} \
  --phase "{name}:{agent}:{input}:{output}:{gate}" \
  --input "{name}:{description}" \
  --description "{brief description}" \
  --sequential {true|false} \
  --base-dir {directory}

Repeatable flags: --phase (required, at least one), --input (required, at least one).

Optional: --base-dir sets the root directory for output (default: current directory). Use for test isolation to avoid generating into the project's real .pas/processes/ directory.

This creates the process directory (process.md, mode files, references/, feedback/), thin launcher, and changelog.

7. Create Agents

For each agent determined in step 4, use creating-agents/SKILL.md:

• Read creating-agents/SKILL.md from the same skills directory as this skill
• Follow its workflow for each agent
• The orchestrator agent is always created first

8. Determine Hooks

Evaluate whether the process needs lifecycle hooks:

• Feedback hooks: If .pas/config.yaml has feedback: enabled, the PAS plugin's hooks handle self-eval and routing automatically. No per-process hooks needed for feedback.
• Domain-specific guards: Does any phase need pre-conditions checked before tool use? (e.g., block destructive commands, validate inputs)
• Lifecycle automation: Should anything happen automatically at session start, agent stop, or task completion?

If hooks are needed, use creating-hooks/SKILL.md from the same skills directory as this skill.

Most processes do not need custom hooks — the PAS plugin's global hooks cover feedback lifecycle. Only add hooks when the process has domain-specific automation needs.

9. Verify Against Source Material

If Step 2 (Prepare Reference Material) was used, cross-check every created skill against the reference doc:

1. For each skill, list every technique, tactic, metric, and number it contains
2. Verify each one exists in the reference material — flag any that don't as potential fabrication
3. Check each section of the reference doc is covered by at least one skill — flag uncovered sections as omissions
4. Remove fabricated content. Add skills or skill sections for omissions.

This is a mandatory step when source material exists. Do not skip it.