Convert · the 6th AI Pair turns a plain prompt into a Skill bundle

The AI Pair now has a sixth instrument. Convert takes a plain prompt and emits a complete Agent Skills bundle · SKILL.md plus optional scripts/ and references/ files · in one shot.

What it does

Click the Convert tab on either workbench (Prompts or Skills), optionally add steering ("no scripts please", "focus on the validation logic"), and run. The AI proposes a complete bundle and opens a preview modal. Review the file list, the decomposition rationale, and the proposed SKILL.md. Apply or discard.

Two entry points, same panel:

• Prompts workbench · Convert creates a new Skill via createSkill and navigates you to /skills/<newId>. The source prompt stays untouched.
• Skills workbench · Convert mutates the current empty draft. SKILL.md is replaced with the AI-generated body and every proposed file lands in editorFiles. Save the version to persist.

Decomposition discipline

The prompt builder enforces the rules from the agentskills.io open spec and Anthropic's Skills docs:

• Long taxonomies or vocabulary glossaries go to references/<topic>.md. SKILL.md gets a pointer.
• Executable logic the agent should run goes to scripts/<verb>.{py|js|ts|sh}.
• Worked examples stay inline if they're short, or move to references/examples.md if they're not.
• Step-by-step orchestration stays in SKILL.md.
• Atomic prompts produce files: []. Convert does not fabricate decomposition where the source prompt has no content for it.

The model surfaces its structural reasoning in a decomposition_rationale field so you can sanity-check before applying.

What it validates

Convert's decoder + preview modal enforce the structural rules before the Apply button enables:

• File-count cap of 12 (typical output is 0-4 files).
• Per-file size ≤ 1 MiB, total bundle output ≤ 250 KB.
• Path-whitelist regex SKILL.md | scripts/[a-z0-9._-]+\.(py|js|ts|sh) | references/[a-z0-9._-]+\.md.
• Frontmatter name 1-64 chars [a-z0-9-], not "anthropic" or "claude".
• Frontmatter description ≤ 1024 chars.
• No duplicate paths in files[]. No SKILL.md entry in files[].

Any validation issue surfaces in the preview with a per-file label · Apply stays disabled until the bundle clears.

Model eligibility

Convert needs a model with ≥ 64k output tokens and streaming support. Honors your Workbench Model when it qualifies. Eligible: Opus 4.7, Opus 4.6, Sonnet 4.6, Sonnet 4.5, Haiku 4.5, GPT-5.5, GPT-5.4 + variants, Gemini 2.5 Pro / Flash / Flash-Lite. Excluded: GPT-4.1 family (32k output, too small), GPT-5.5 Pro (background mode only).

BYOK and tier behavior

Convert is BYOK-only. The 7-call demo budget covers Critique / Improve / Rewrite only · adding Convert to demo would open a second platform-funded inference path. The rest of the chain is standard: getLLMProviderForContext for BYOK resolution, impersonation block at the route guard, tier enforcement on Free's 250 monthly AI calls cap. Rate-limited to 10 calls per minute per user · Convert produces a full multi-file bundle so its per-call output is heavier than Critique or Improve.

What it isn't

Convert is origin-only. It does NOT participate in the critique → improve → rewrite AI Chain. The mental model: Critique / Improve / Rewrite refine an existing artifact; Convert authors a new one from a different source format.