Every Claude Code user is, by definition, AI-app-curious. And until v1.5, opchain had zero owners for the part of a project where an LLM is actually in the loop. You could spec, design, build, audit, ship, and monitor a CRUD app with opchain — but the moment “and then Claude summarizes it” entered the picture, you were back to vibes.
v1.5 closes that gap. It adds four AI-native skills and wires them through the pipeline you already use, so the AI part gets the same idea → spec → build → ship → evaluate treatment as the rest of the app. The theme is one sentence:
Building with AI should be an engineering discipline, not a vibe.
The four new skills
Each owns one slice of an LLM feature, and each is built to make that slice a measured artifact instead of a hope.
- oc-claude-api — the request surface. Model routing (Opus / Sonnet / Haiku / Fable by task), prompt caching on by default, tool-use patterns, and version-migration playbooks that produce a diff PR. When you move a model version, this skill gates the change instead of letting it silently drift.
- oc-rag-forge — a tri-agent retrieval harness (Designer → Builder → Evaluator). It picks the vector DB, embedding model, and chunking strategy, then scores retrieval against a labelled set instead of eyeballing three queries that happen to work.
- oc-agent-forge — a tri-agent build harness for Claude Agent SDK apps. It owns topology (single / orchestrator-worker / pipeline / hierarchical), tool budgets, and the harness loop — and gates the agent on a task-fixture suite, so “it worked when I tried it” stops being the bar.
- oc-prompt-ops — prompts as code: versioned, diffable, and gated on an eval suite the same way application code is gated on tests. A prompt change becomes a reviewable diff with a measured score delta.
The ripples
The new skills don’t sit off to the side. They extend what was already there:
- oc-stack-forge gained
kind: vector-dbpacks: pgvector, Pinecone, Turbopuffer, and Supabase Vectors. Picking a vector store is now part of the same coverage registry as picking a language or a host. - oc-code-auditor learned an AI-safety pass — prompt-injection and tool-use-safety rules that only run when an LLM is in the loop. It traces untrusted content into prompts, and tool arguments into dangerous capabilities.
- oc-app-architect
/oc-discovernow branches on “is this an AI app?” and routes through the four new skills automatically. You don’t have to know they exist; the orchestrator does.
How they chain
The point of a skillchain is that the pieces compose. A typical AI feature now flows like this:
/oc-discoverdetects the AI surface and routes to oc-claude-api for model routing + caching defaults.- If it’s retrieval-shaped, oc-rag-forge picks the vector DB (via the new stack-forge packs) and stands up an eval set.
- If it’s agentic, oc-agent-forge designs the topology and tool budget.
- oc-prompt-ops versions every prompt and gates changes on the eval suite.
- oc-code-auditor runs the AI-safety pass before you ship.
No single mega-prompt. Each skill does one job, writes a checkpoint, and hands off. (For the argument behind that design, see the engineering notes on evaluating instead of eyeballing.)
Dogfooding
opchain evaluates itself. The new prompts/opchain-eval/ set is a routing
goldset — given a dev request, does opchain pick the right skill? — published as
the worked example for /oc-prompt eval. We catch our own trigger-copy drift the
same way we’d ask you to catch yours. The
honest limits of that dogfooding
are their own post.
The through-line
Every skill in this release makes the AI part of your app an evaluated artifact — measured, versioned, and safe to ship. That’s the whole bet of v1.5: the difference between a demo and a product is whether you can prove the AI part works, and prove it still works after you change it.
Browse the skill library or install opchain to start.