block-creation

block-creation — Skill instructions

When to invoke

Invoke this skill when the builder expresses intent to create a new block. Trigger phrases include:

• "build me a block that ..."
• "scaffold a block for ..."
• "create a new block named ..."
• "I need a block that does ..."

Do NOT invoke when the builder is asking to:

• Install an existing block into a workspace (that's the SeldonFrame MCP's install_* tools).
• Modify an existing block (refer to direct file edits).
• Generate UI components or pages (out-of-scope for this slice; future SLICE 4 scope).

Framing — this is code-authoring, not workspace-admin

The scaffold writes files to the builder's repo working tree. It does NOT:

• commit to git
• push to a remote
• install the block into a running workspace
• modify any workspace's database

Those are the builder's follow-up actions (git add / git commit / git push, then install_block against a workspace). Claude Code is a code-gen assistant here, not a platform admin.

Workflow

1. Intent → BlockSpec

Claude Code reads the builder's natural-language intent and constructs a BlockSpec JSON object matching the shape below.

Before constructing the spec, Claude should Read:

1. Reference anatomies — the composition-contract shape of two existing blocks. These show the canonical layout for produces / consumes / tools / subscriptions:
   • packages/crm/src/blocks/notes.block.md (simple tool-only block — the PR 1 C7 smoke-test)
   • packages/crm/src/blocks/crm.block.md (real-world anatomy with ## Subscriptions declaring a reactive handler)
2. Canonical NL → BlockSpec examples (these are hand-curated translation lessons in code form):
   • packages/crm/src/lib/scaffolding/nl/example-specs.ts — EXAMPLE_SPECS export. Two examples: tool-only (contact-notes)
     • reactive-with-subscription (auto-activity-log).
3. Hard constraints on BlockSpec shape (the schema Claude must honor):
   • packages/crm/src/lib/scaffolding/spec.ts — BlockSpecSchema
     • field patterns (slug, tool name, handler name, event name, subscription event).

Reading order: anatomies first → schema → examples. Each reads in under 30 seconds; together they give Claude the concrete pattern needed to translate NL confidently.

Required fields:

• slug: kebab-case lowercase, e.g. "notes", "client-satisfaction". Cannot collide with reserved core blocks: crm, caldiy-booking, email, sms, payments, formbricks-intake, landing-pages.
• title: human-readable title, e.g. "Notes", "Client Satisfaction".
• description: one-line builder-facing description.
• triggerPhrases: 2-5 natural-language phrases that would activate this block.
• frameworks: array of framework strings, e.g. ["universal"].

Optional fields (defaulted if absent):

• produces: array of { name, fields } event declarations.
• consumes: array of { kind, ... } discriminated-union entries (kind: "event" | "soul_field" | "trigger_payload").
• tools: array of tool definitions: { name, description, args: [...], returns: [...], emits: [...] }.
• subscriptions: array of { event, handlerName, description, idempotencyKey } reactive handlers. event is fully-qualified: "<source-block>:<event.name>".

2. Pre-classify the intent (deterministic safety check)

Before translating NL → BlockSpec, call the deterministic classifier once to sanity-check the intent:

import { classifyIntent } from "@/lib/scaffolding/nl/intent-classifier";

const classification = classifyIntent(nlIntent);
// { tier: 1 | 2 | 3, issues: string[], suggestedAction: string }

Behavior by tier:

• Tier 3 — the intent is destructive / modifies an existing core block. Refuse by default. If the builder explicitly confirms (e.g., "yes, this is an admin-only tool with safeguards"), the scaffold may proceed; otherwise relay the specific risk and stop.
• Tier 1 — the intent is empty, trivially short, or self-contradictory. Ask ONE focused clarifying question (suggestedAction points at what to ask). Wait for a reply. Re-classify after the reply; proceed only when classification is no longer tier 1.
• Tier 2 — the common case. Proceed with scaffold generation. Fill sensible defaults for under-specified fields; mark each default with TODO (scaffold-default) comments the builder can grep post-scaffold.

The classifier's heuristics are loose — false-negative tier-2 intents are acceptable (downstream BlockSpec validation catches structural problems). False-positive tier-3 decisions are preferred over false-negative — better to refuse once and require confirmation than scaffold a destructive tool silently.

3. Clarify rarely (three-tier policy per audit G-4)

• Tier 1 — Ask once only when the intent is genuinely meaningless (zero description, or internally contradictory type declarations). Ask ONE focused question; proceed.
• Tier 2 — Default with TODO markers for everything else. When the intent doesn't name tools, scaffold create_<entity> / list_<entity>s / get_<entity> as defaults. Every scaffolded default lands with TODO (scaffold-default) markers the builder can grep.
• Tier 3 — Fail on dangerous output. Refuse to scaffold:
  • Destructive tools without explicit confirmation.
  • Modifications to existing blocks (scope limits to NEW blocks).

4. Run the scaffold

Once the BlockSpec is constructed, write it to a temporary file and invoke the scaffold CLI:

# Claude writes /tmp/spec.json with the BlockSpec JSON
pnpm scaffold:block --spec /tmp/spec.json

The scaffold will:

1. Validate the BlockSpec against the Zod schema.
2. Render BLOCK.md, tools.ts, subscription handler stubs (if any), and test stubs.
3. Write all files to packages/crm/src/blocks/<slug>.* and packages/crm/tests/unit/blocks/<slug>.spec.ts.
4. Run the validation gate:
   • parseBlockMd round-trip on the new BLOCK.md
   • tsc --noEmit across the CRM package
   • pnpm emit:blocks:check (after the builder adds the block to the emit TARGETS registry on next step)
5. On success, print the created-files list and next-step hints.
6. On failure, print the orphan report with git clean recovery commands.

5. Relay results

After the scaffold succeeds, tell the builder:

• Files created (from the scaffold's stdout).
• Next-step checklist (also from the scaffold's stdout):
  1. Review TODO (scaffold-default) markers — they're the fill-in points.
  2. Add the block to scripts/emit-block-tools.impl.ts TARGETS list so emit:blocks:check covers it.
  3. Run pnpm emit:blocks to populate the TOOLS block in the BLOCK.md.
  4. Run pnpm test:unit — new test stubs appear as todos.
  5. git diff / git add / git commit when satisfied.

On failure, relay the orphan report verbatim so the builder has the exact recovery commands.

BlockSpec example — "notes" block

{
  "slug": "notes",
  "title": "Notes",
  "description": "Simple note-taking on contacts.",
  "triggerPhrases": [
    "Add a notes block",
    "Install notes",
    "Let me jot notes on contacts"
  ],
  "frameworks": ["universal"],
  "produces": [
    {
      "name": "note.created",
      "fields": [
        { "name": "noteId", "type": "string", "nullable": false },
        { "name": "contactId", "type": "string", "nullable": false }
      ]
    }
  ],
  "consumes": [],
  "tools": [
    {
      "name": "create_note",
      "description": "Create a note on a contact.",
      "args": [
        { "name": "contactId", "type": "string", "nullable": false, "required": true },
        { "name": "body", "type": "string", "nullable": false, "required": true }
      ],
      "returns": [
        { "name": "noteId", "type": "string", "nullable": false, "required": true }
      ],
      "emits": ["note.created"]
    }
  ],
  "subscriptions": []
}

Dry-run mode

Pass --dry-run to see what would be created without writing:

pnpm scaffold:block --spec /tmp/spec.json --dry-run

Useful for previewing before committing to the scaffold.

Error recovery

If the scaffold fails mid-pipeline:

• Files that landed before the failure remain on disk (by design — orphan detection, not transactional rollback).
• The error message lists every orphan + concrete git clean commands to remove them.
• Fix the underlying issue and re-run. The scaffold refuses to overwrite existing files, so the builder must remove orphans first (or pick a different slug).

Out of scope for this skill

• NL intent parsing without an explicit BlockSpec construction step (future PR will layer this on; this PR requires Claude Code to translate intent → BlockSpec JSON inline).
• UI scaffolding (pages, Puck components) — future SLICE 4.
• Database schema generation (builder authors Drizzle schemas manually if their block needs persistence).
• Marketplace publishing (out of slice).