From havn
Use when output can't be validated by tests or lint alone — prompt engineering, skill design, documentation, system prompts, API contracts. Triggers on "review and fix loop", "evaluate until clean", "optimize until no issues", "review cycle", "iterate until passing", "quality gate".
How this skill is triggered — by the user, by Claude, or both
Slash command
/havn:evaluate-optimize-loopThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
Builder produces work. Reviewer evaluates against criteria. Builder fixes. Repeat until clean.
Builder produces work. Reviewer evaluates against criteria. Builder fixes. Repeat until clean.
Uses havn's builder and reviewer agents.
Dispatch a builder agent with run_in_background: true and isolation: "worktree". The builder reads project memory, implements the task, and updates memory with stable patterns discovered.
If a builder already completed work (you have its agent ID), skip to step 2.
Dispatch a reviewer agent with run_in_background: true. The reviewer prompt MUST contain:
The reviewer outputs structured findings. Instruct the reviewer to end with PASS or FAIL verdict.
Review these files against [spec]:
- [absolute/path/to/file1]
- [absolute/path/to/file2]
- Fetch [spec URL] for current requirements
Criteria:
- [specific criterion 1]
- [specific criterion 2]
Resume the original builder agent (not a fresh one — it has context and project memory from the first pass). Inject the reviewer's findings verbatim. The builder fixes issues, reports what changed, and updates project memory.
Resume the reviewer agent (not a fresh one — it remembers prior findings). Tell it which files changed. It re-checks previously-failing criteria plus any new issues introduced by fixes.
npx claudepluginhub phibkro/vevx --plugin havnOrchestrates a build-review-improve loop using sub-agents: a worker produces artifacts, a panel of reviewers critiques in parallel, and the worker iterates until passing. For explicit invocation only.
Runs multi-agent verification loop post-implementation, dispatching specialized agents for review with autonomous subagent fixes and retries until unanimous approval.
Orchestrates implement-analyze-fix loops: implements code, AI-reviews changes, fixes issues, repeats until clean or max iterations. For iterative development with quality checks.