Harness documentation¶
How Dazzle runs unattended improvement without silent drift — and how to adapt the same pattern to other tasks.
This section is for humans (operators, visitors, people who want to steal the design). Executable truth for agents remains the runtime runbook:
| Audience | Start here |
|---|---|
| Curious outsider | Improve as harness exemplar; leftover-honesty ethnography |
| Operator (status, rearm, force a dig) | Operator field guide |
| Strategy map (what each playbook is for) | Strategy catalog |
| Executing agent | .claude/commands/improve.md (not duplicated here) |
| Sibling commands | Autonomous Harness (fleet overview) |
Design rule: packaging here does not change how the loop operates. Runtime stays agent-first; these docs make the structure legible.