Chain Configurator

Chain-configurator — interactive YAML generator for the chain orchestrator

This skill writes a chain YAML; it does not execute one. Lead the user through batched Q&A to produce a draft chain spec under drafts/sample-chains/<name>.yaml, then hand off to the chain skill (or python3 -m orchestrator for the workspace-side runtime) for execution and validation.

Use only the closed chain fields documented by the installed chain skill. Do not require or disclose an external private schema.

When to use

Use this skill when one or more of the following are true:

• The user explicitly asks for it — "help me configure a chain," "generate a chain for X," "set up a chain," "create a new chain," "configure audit-," "build a chain from scratch," "scaffold a chain for me."
• The user wants to formalize a repeating workflow but has not written chain YAML before.
• The user wants to adapt an existing chain template (audit-cross-systems, validate-release-candidate, validate-calculation, validate-decision, etc.) to their specific domain.
• The user has a workflow described in natural language and wants it captured as a verifiable chain.
• A first-time chain author needs scaffolding + sanity-checking.

When to skip

• The user already has a chain YAML draft and just wants validation — read the file directly + reference the chain-spec format doc; no Q&A overhead needed.
• One-shot work that won't be repeated — use ad-hoc skill invocation; don't formalize as a chain.
• The user is editing an existing chain — use the Edit tool directly; no Q&A regeneration needed.
• The user knows the spec well — Q&A becomes friction; let them write the YAML directly and review.

Procedure (overview)

The full Q&A scripts for each template + the from-scratch flow are documented inside the skill at runtime invocation. The high-level shape:

1. Identify intent (one batched question)

Ask the user ONE consolidated question that determines the path:

1. Specialize existing template vs build from scratch?
2. What workflow does the chain run (one sentence)?
3. Roughly how many steps?

If the user says "I don't know" on #1, default to (a) — show the existing templates and let them pick the closest:

When no current-project template exists, build the minimal YAML from scratch.

2. If specializing an existing template

Ask the template-specific questions in a batched form. For each template, the questions correspond 1:1 with the template's inputs: section. Walk through them in order.

Use the existing template's structure as the scaffold; the user's answers populate the inputs. Leave the template's steps:, gates:, outputs: untouched unless the user explicitly asks to modify them.

3. If building from scratch

Walk through the chain-spec sections one at a time, batching related questions:

• Batch A — metadata + inputs: chain name (lowercase-hyphen), one-line description, declared inputs (per input: name, type, required, default, description).
• Batch B — per step (iterate for each step the user identified): step ID, runner (skill vs <primary>), skill reference (for skill runner) or instruction text (for <primary> runner), input map (referencing prior steps or chain inputs), output schema (text / json / jsonl / structured-text with anchor), retry policy, on_failure policy, gates (optional but recommended for any step that asserts a side effect).
• Batch C — outputs: which step's output(s) become the chain's final output? Render templates.
• Batch D — audit + cost budget (optional): audit log path, cost budget limit + on_exceed policy.

After each batch, summarize what the user said back as YAML-ready notes. Only render the final YAML after all batches complete.

4. Render the YAML

Emit the chain YAML to the operator-selected current-project path. Validate its closed fields against the installed chain skill and surface specific errors.

5. Hand off

Present the user with:

• The path to the new draft.
• A one-line command to validate + dry-run the chain: e.g., python3 -m orchestrator validate drafts/sample-chains/<name>.yaml && python3 -m orchestrator run drafts/sample-chains/<name>.yaml --mock.
• A note that the chain is in drafts/sample-chains/, not chains/, until the user has validated it on real inputs and is ready to promote it.

Anti-patterns

• Skipping the batched-Q&A discipline. Asking 12 separate questions one by one is friction; batch related questions and wait for the user to answer all of them in one turn.
• Writing the YAML before the user has answered the metadata + inputs batch. Premature rendering misleads — the user sees a YAML they haven't committed to.
• Inventing unsupported chain fields. Stay within the installed skill's closed format and report when the requested feature is unavailable.
• Promoting the draft to chains/ automatically. Drafts go to drafts/sample-chains/; promotion to chains/ is a separate operator-mediated step after real-input validation.
• Failing to validate the generated YAML before handing off. A malformed draft wastes the user's first run. Validate before presenting.
• Skipping cross-family routing on cross-check steps. When the user requests a cross-check step in the chain, ensure the spec routes it to a different-family verifier (per the chain-spec's family-aware routing rule). Don't let the user accidentally configure a same-family cross-check.
• Forcing the user into the "from scratch" path when an existing template is a 90% match. Specialize the template; save the user time.
• Asking the user to specify gates / output_schema / retry policy when they're a first-time author and don't yet have a workflow that surfaces those choices. Default to sensible values (e.g., output_schema: text if uncertain, retry: 1, no gates: initially) and prompt the user to revisit after the first chain run surfaces the need.

Examples across domains

Limitations

This skill writes a chain YAML; it does not execute one. Run the result via the installed chain skill.

The draft is saved to drafts/sample-chains/<name>.yaml — not chains/. Promotion to chains/ is a separate operator-mediated step that happens after the chain has been validated on real inputs.