Benchmark data shows agents need to be told about shard files to use them effectively. Without prompting, agents passively discover shards in grep results but don't actively use them. With prompting, results improve significantly (Lance's 14-repo test: -43% cost on complex tasks).
Add `supermodel --skill` that prints a prompt agents can use:
```
$ supermodel --skill
This repository has .graph.* shard files next to source files containing
dependency, call graph, and impact analysis data from Supermodel. When
navigating the codebase, read the .calls, .deps, and .impact files to
understand relationships between modules before making changes.
```
Usage patterns:
- `supermodel --skill >> CLAUDE.md` — append to Claude Code instructions
- `supermodel --skill >> AGENTS.md` — append to agent instructions
- Agent can run `supermodel --skill` to learn about the shards
From Grey's suggestion in Slack — portable across agent platforms, doesn't require hooks or config file injection.
Benchmark data shows agents need to be told about shard files to use them effectively. Without prompting, agents passively discover shards in grep results but don't actively use them. With prompting, results improve significantly (Lance's 14-repo test: -43% cost on complex tasks).
Add `supermodel --skill` that prints a prompt agents can use:
```
$ supermodel --skill
This repository has .graph.* shard files next to source files containing
dependency, call graph, and impact analysis data from Supermodel. When
navigating the codebase, read the .calls, .deps, and .impact files to
understand relationships between modules before making changes.
```
Usage patterns:
From Grey's suggestion in Slack — portable across agent platforms, doesn't require hooks or config file injection.