Dag Council

/dag-council — auto-authored expert council

For wide solution spaces: the conductor reads your goal + ground truth and authors the personas itself (real, distinct expert angles — not one prompt resampled N times). Candidates generate concurrently, a judge panel scores through multiple lenses, and the champion is synthesized with the runners-up's orthogonal highlights grafted in.

Usage

bun run dag-council 
# JSON: { goal, groundTruth, conductorModel?, lensCount?, lensModel?, reasonModel? }

Stdout: first line = artifact path, then LENS CHAMPIONS / SYNTHESIS CANDIDATES / FINAL marker sections (stable format, safe to split downstream).

Grounded tier (domain-correctness forks)

When the decision is in a domain where being wrong is expensive and reality is messy (accounting/legal/ops: dirty data, partial failure, lifecycle-end) and it's hard to reverse, upgrade the run:

1. Market-first: research what real-world practice actually does first (e.g. via

/dag-research) and put that hard-evidence baseline into groundTruth — personas argue from facts, not vibes.

1. Domain-role personas: put the real roles into goal framing — the daily

operator / compliance & audit / automation first-principles / lifecycle-end.

1. Anti-happy-path judging axis: state in groundTruth that every option must be

judged against dirty data, concurrency, partial failure, cross-boundary, lifecycle-end and scale-balloon — otherwise personas quietly answer for the happy path.

1. Judge + graft: unanimous consensus is a strong signal; the champion often shrinks

the debate rather than picking a side. Domain red lines stay with the owner — council output is input, not the final call.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: AbyssCN
• Source: AbyssCN/oh-my-dag
• License: MIT

Install and usage instructions live in the source repository linked above.