Create

Create

Unified creator for skills, commands, agents, and specs. Parses --type from $ARGUMENTS (or asks) and runs the matching workflow.

Shared references: references/creator-conventions.md, references/prompt-engineering-guide.md, references/interaction-patterns.md.

Dispatch

Parse $ARGUMENTS for --type skill|command|agent|spec (or leading positional token skill|command|agent|spec). If ambiguous, AskUserQuestion (header "Type"):

• "Skill" — new Claude Code skill
• "Command" — new slash command
• "Agent" — new subagent
• "Spec" — technical specification (delegates to spec-writer agent)

Remainder of $ARGUMENTS is the subject/name hint.

---

Shared Workflow

All types follow this skeleton; per-type sections add fields and validation.

1. Gather Requirements

AskUserQuestion for discrete choices; free-form for open-ended.

Common fields:

• Purpose (free-form): what it does, when to invoke it.
• Location — AskUserQuestion header "Location":
  • "Project" — .claude/<kind>/ (versioned with repo)
  • "User Global" — ~/.claude/<kind>/ (across projects)
  • "Plugin" — <kind>/ (prove-style plugin root)

Type-specific fields follow (see per-type sections).

2. Generate

Apply references/prompt-engineering-guide.md techniques. Derive hyphen-case name from purpose when not supplied.

3. Quality Self-Check

Before presenting, verify:

• Primacy: critical directives appear early.
• Constraint pairing: every "never X" has an "instead, do Y".
• No preamble bloat, no redundant restatement.
• Model calibration: Opus < Sonnet < Haiku in scaffolding.
• Token budget fits consumption context.

Fix failures before presenting.

4. Validate

Per-type checklist (see sections below). Must pass before writing.

5. Review Gate

AskUserQuestion header "Review": "Create" / "Revise".

6. Commit Delegation

On commit request, delegate to the commit skill. Do not craft ad-hoc commits.

Plugin Scope Registration

When adding to a plugin, ensure .claude/.prove.json has the scope:

"scopes": { "commands": "commands/", "agents": "agents/", "skills": "skills/" }

---

Type: Skill

Path: skills/<name>/SKILL.md (plus optional references/, assets/, scripts).

Gather (skill-specific)

• Interaction pattern — AskUserQuestion header "Interaction":
  • "Autonomous" — no user input
  • "Interactive" — uses AskUserQuestion gates
  • "Delegating" — thin wrapper over an agent
• Resources — AskUserQuestion header "Resources" (multiSelect):
  • "Reference docs" (references/<topic>.md)
  • "Templates/assets" (assets/<template>.md)
  • "Scripts"
  • "None"

Frontmatter

---
name: <hyphen-case>
description: <action-first; 2-3 trigger scenarios; natural trigger phrases>
---

Description tuning (highest-leverage field):

• Front-load the primary action verb.
• Include 2-3 concrete trigger scenarios.
• List natural trigger phrases the user would say.

Body by Interaction Pattern

• Autonomous: numbered workflow steps.
• Interactive: gather → execute → validate → review gate.
• Delegating: before delegating → delegation prompt → after delegation.

Validation Checklist

• Directory name matches name frontmatter
• name and description present
• Description front-loads action and includes trigger scenarios
• Body uses numbered steps for workflows
• Interactive skills follow references/interaction-patterns.md
• Token budget appropriate for skill context
• Referenced resources exist or are in plan
• Plugin skills: skills scope in .claude/.prove.json

Optional Command Wrapper

If the skill should be user-invocable via /prove:<name>, offer a thin command:

---
description: <short skill description>
argument-hint: "[input]"
---

# <Name>: $ARGUMENTS

Load and follow `skills/<name>/SKILL.md`.

Commit example: feat(skills): add <name> skill.

---

Type: Command

Path: <location>/commands/<name>.md (or commands/<ns>/<name>.md for namespaced /ns:name).

Gather (command-specific)

• Arguments — AskUserQuestion header "Arguments":
  • "No arguments"
  • "Single argument" — use $ARGUMENTS
  • "Multiple arguments" — use $1, $2, ...
• Tool restrictions — AskUserQuestion header "Tools":
  • "Full access (Recommended)" — omit allowed-tools
  • "Read-only" — Read, Grep, Glob
  • "Git-only" — Bash(git *)
  • "Custom" — user specifies
• Skill delegation (plugins only, free-form): which skill does it wrap? Plugin commands are thin entry points.

Frontmatter

---
description: <action-first; shown in /help and autocomplete (~250 char cap)>
argument-hint: "[expected args]"
allowed-tools: <if restricted>
---

Body Templates

Standalone:

---
description: <action>
argument-hint: "[input]"
---

<Command prompt content>

Plugin wrapper (thin — logic lives in skill):

---
description: <action>
argument-hint: "[input]"
---

# <Name>: $ARGUMENTS

Load and follow `skills/<name>/SKILL.md`.

Placement

Location	Path
Project	.claude/commands/<name>.md
User	~/.claude/commands/<name>.md
Plugin	commands/<name>.md
Namespaced	commands/<ns>/<name>.md → /ns:name

Validation Checklist

• Filename lowercase, hyphen-delimited, matches purpose
• description present (required for model invocation and /help)
• argument-hint set if command accepts arguments
• Tool restrictions match intent (read-only for analysis, etc.)
• Plugin commands delegate to a skill — no business logic in the file
• Multi-step workflows use numbered steps
• Body does not restate the description

Commit example: feat(commands): add <name> command.

---

Type: Agent

Path: <location>/agents/<name>.md.

Gather (agent-specific)

• Tool permissions — AskUserQuestion header "Tools":
  • "Read-only (Recommended for reviewers)" — Read, Grep, Glob
  • "Read + Research" — Read, Grep, Glob, WebFetch, WebSearch
  • "Full developer" — Read, Write, Edit, Bash, Glob, Grep
  • "Custom"
• Model — AskUserQuestion header "Model":
  • "opus" — complex reasoning, architecture
  • "sonnet (Recommended)" — general development
  • "haiku" — simple, repetitive, cost-efficient

Frontmatter

---
name: <hyphen-case>
description: <WHAT it does + WHEN to invoke; drives auto-delegation>
tools: <scoped permission list>
model: opus|sonnet|haiku
---

System Prompt Structure

1. Identity and purpose
2. Available tools and when to use each
3. Workflow — numbered decision loop
4. Output format and reporting
5. Constraints and guardrails
6. Failure handling

Model calibration:

• Opus: lean, few guardrails. Target ~285-636 tokens (built-in agent range).
• Sonnet: moderate detail, explicit workflow steps.
• Haiku: more scaffolding, explicit examples, precise constraints.

Validation Checklist

• Name hyphen-case, matches filename
• Description explains what + when (drives auto-delegation)
• Tools match role (read-only for reviewers; no Write/Edit/Bash unless needed)
• System prompt has clear responsibilities and workflow
• Output format specified
• Model appropriate for task complexity
• Plugin agents: agents scope in .claude/.prove.json

Commit example: feat(agents): add <name> agent.

---

Type: Spec

Orchestrates spec lifecycle (create, revise, audit). Delegates all writing to the spec-writer agent — this skill does not author spec text directly.

Step 1: Determine Mode

If not obvious from context, AskUserQuestion header "Mode" with "Research & proceed" per references/interaction-patterns.md:

• New Draft — from description, decision records, or conversation
• Revise — edit existing spec (user gives path or search specs/)
• Audit — review for completeness, ambiguity, consistency; no modifications

Defaults:

• $ARGUMENTS points to a .spec.md file → Revise
• $ARGUMENTS references .prove/decisions/ → New Draft

Step 2: Gather Context

New Draft:

1. Read referenced decision records; extract chosen option and design details.
2. Check specs/ for related specs to cross-reference.
3. Confirm scope (in/out of scope) with user.

Revise:

1. Read the spec in full.
2. Clarify desired changes (additions, removals, clarifications, restructuring).
3. Determine version bump: major (breaking) or minor (additive/clarification).

Audit:

1. Read the spec in full.
2. Optionally read related specs for cross-reference consistency.

Step 3: Delegate to spec-writer Agent

Launch the spec-writer agent with a structured prompt. The agent owns RFC conventions, section structure (Purpose, Scope, Terminology, Conformance, normative body, Change Log), normative language, and self-auditing.

New Draft:

Mode: New Draft
Subject: [title]
Version: 0.1 (Draft)
Input context: [decision records, brainstorm notes, or requirements]
Scope:
- In scope: [what the spec defines]
- Out of scope: [what it excludes]
Related specs: [existing specs in specs/ to cross-reference]
Write to: specs/[slug].spec.md

Revise:

Mode: Revise
Spec path: specs/[name].spec.md
Current version: [X.Y]
Changes requested: [what to add, remove, or modify]
Version bump: [major | minor]
Update in place.

Audit:

Mode: Audit
Spec path: specs/[name].spec.md
Run the full Completeness Checklist. Report findings as:
- Location (section number or line)
- Issue description
- Severity (Critical / Important / Improvement)
- Suggested fix
Do not modify the spec.

Step 4: Review Output

• New Draft: verify file at specs/[slug].spec.md follows expected structure. Report location and version.
• Revise: verify version bump, Change Log update, cross-reference resolution. Report changes.
• Audit: present findings by severity. AskUserQuestion header "Next step": "Fix issues" (switch to Revise) / "Done".

Rules

• Delegate writing to the spec-writer agent; this skill orchestrates only.
• Confirm scope before drafting; prefer promoting existing decision records over starting from scratch.

Commit examples:

• docs(specs): draft validation config v2 specification
• docs(specs): revise handoff protocol — add context fields