Model Selection
| Task Type | Recommended Model | Why |
|---|---|---|
| Quick fixes, small edits | Claude Haiku 4.5 | Fastest, lowest credit cost |
| General development | Claude Sonnet 4.6 | Best speed/quality tradeoff |
| Complex architecture, deep reasoning | Claude Opus 4.6 | Strongest reasoning |
| Parallel routine tasks | Claude Haiku 4.5 | Resource-efficient across many Workers |
Execution Mode Selection
| Scenario | Mode | Why |
|---|---|---|
| Clear, well-defined task | Agent Mode | Direct execution, no planning overhead |
| Unfamiliar codebase, exploration | Plan Mode | Review approach before acting |
| Production changes, architectural work | Plan Mode → Agent Mode | Validate plan, then execute |
| Fast iteration loops | Agent Mode | Minimal friction |
Context Management
For large projects, keep the main context window lean:- Delegate to subagents —
@Verifier,@Fast Context, and custom subagents run in their own context; only their results return to the main agent. - Reference files explicitly — Use
@filenameto pull only what’s needed instead of loading whole directories. - Chunk large tasks — Start a new session at natural boundaries (per feature, per module).
- Plan Mode first — Validate the approach before burning execution context on the wrong path.
Configuration Priority
When rules conflict, Verdent applies this priority order:- Project Rules (
AGENTS.mdin project root) — highest priority - Workspace settings — application workspace-level settings
- User Rules (
~/.verdent/VERDENT.md) — global defaults - Verdent defaults — lowest priority
See Also
Subagents
Delegate work with isolated context windows
BYOK
Use your own provider API keys
BYOA
Run Workers on external Agent runtimes
Rules
Project and user rule files