effective-go

Language-agnostic code simplification — reduces complexity (nesting, long functions, duplication, dead code, unclear names) while preserving exact behavior, applying one tested change at a time within a deliberate scope.

Run expensive top-tier models (Claude Fable/Opus, GPT-5.6 Sol, or similar) efficiently — the premium model keeps judgment, architecture, synthesis, and final review, while cheaper subagents (Sonnet, Haiku, Terra, Luna) do bounded token-heavy research, coding, and testing with self-contained handoffs and verified results.

Hostile, bias-free review of a change — dispatches independent adversarial reviewers that hold the code against what was agreed and question whether what was agreed was right, plus a standalone false-positive filter, and reports only verified findings in plain language. Ships a full 18-reviewer panel and a cheaper 8-reviewer quick panel that keeps both verification gates.

Docs-first discipline for anything involving third-party APIs, packages, providers, or fast-moving product behavior — the agent web-searches for current official docs and reads primary sources before assuming from memory, grounding answers in what the docs actually say.

Plain-language communication for readers who can't see the code — the agent translates findings, answers, status, plans, and recaps out of file/function/variable names into Simplified Technical English (ASD-STE100), leads with the outcome, keeps the important point in the final message of the turn, and matches technical readers only when they ask.

Budget loop for long-running or parallel agent work under Claude's 5-hour and weekly usage limits — run bounded waves, check real usage between them, pause before the cap, and resume with self-contained prompts only when the window is actually clear.

My collection of Claude (and other AI's) skills, plugins, MCP servers, and more.

Comment discipline for code the agent writes — every comment stays within a line or two, carries information the code does not, and describes the current state only, with no iteration labels, finding numbers, wave or task IDs, project phase names, or narration of how the code got here; regressions are pinned by tests instead.

Fully local visual plans and recaps — the agent writes markdown, a bundled zero-dependency server renders it in your browser with Mermaid diagrams, rich diffs, styled DB migrations, API call cards, and a read-only OpenAPI view, with live reload and a comment feedback loop. No remote services.