Domain Research

Domain Research

Before you build an agentic system for a domain, you research the domain the way a senior practitioner would brief a new hire. Output: a single DOMAIN.md that later agents consume.

Research layers

1. Taxonomy: core concepts, sub-fields, canonical workflows. (e.g., 3D modeling → modeling / UV / rigging / animation / shading / lighting / rendering / post.)
2. Tools: the 5-10 dominant programs, libraries, APIs. Note open-source vs. commercial, CLI vs. GUI, scriptable vs. not.
3. MCPs / programmatic interfaces: what's already exposed via MCP? What's scriptable via SDK? What requires GUI automation?
4. Idioms & conventions: how do practitioners name things, structure projects, version files, share work? (e.g., Blender → bpy API, collections, linked libraries.)
5. Best practices: top 10 rules practitioners would cite unprompted. Sourced from docs, style guides, community consensus.
6. Anti-patterns: top 10 ways novices and AI go wrong. Common failure modes.
7. Eval seeds: 10 concrete tasks that an agent in this domain should ace, and 10 harder ones it should survive.

Research methodology

• WebSearch for current-year docs, "best practices ", "[tool] Python API", "anti-patterns in ".
• WebFetch the official docs, top 3 community style guides, one textbook if available.
• MCP registry search (search_mcp_registry) for each major tool.
• arxiv / scholar (if connected) for recent methodology papers.
• Cross-check sources: a practice is only "consensus" if it appears in 3+ independent sources.

DOMAIN.md structure

# Domain: <name>

## One-line definition

## Taxonomy
- Concept A — what it is, when it matters.
- ...

## Dominant tools
| Tool | License | Interface | Notes |
|---|---|---|---|

## Available MCPs
| MCP | Registry ID | Capabilities | Auth |
|---|---|---|---|

## Idioms
- Naming: ...
- Project layout: ...
- File versioning: ...

## Best practices (prioritized)
1. ...

## Anti-patterns
1. ...

## Eval seeds
### Core (must pass)
1. ...

### Stretch
1. ...

## Sources
- [Doc title](url)
- ...

Scope discipline

• Time-box: 15-20 min of research per pass. Go deeper only when the user specifies a narrow sub-domain.
• Breadth before depth: taxonomy + tool list first. Drill into the top 2-3 tools only.
• Cite sources. Every best-practice bullet has a source link.

Output

Produce DOMAIN.md in the architect session's working directory. Return a 10-line summary to the orchestrator: scope, top 3 tools, recommended MCPs, top 3 anti-patterns to encode as red-lines, eval seed count.