robmcquade
- 5 skills
- 0 followers
- 6 hours ago last updated
- ▌ Improve · robmcquade bundleTake an artifact, idea, problem statement, thought process, or solution and return a materially stronger REWRITTEN version of it, plus a grounded account of what changed and why. An evaluator-optimizer loop — a fresh-context critic finds what's weak, the work is revised, repeated until it converges — and an independent fresh-context verifier checks the rewrite (fixes applied, no new errors, no meaning or voice drift) before it reaches you. It PRODUCES the result (the counterpart to /pressure-test, which only recommends). User-invoked only. In Claude Code it dispatches fresh-context subagents; on a surface without subagents it runs the same loop inline. Bounded so it never iterates forever — a goal confirmed up front, the evaluator owning the stop, and a hard pass cap. Pass an explicit goal as an argument, or it derives and states one. Preserves your voice and intent; for files it writes a .orig sidecar first so the original is always recoverable.
- ▌ Set Goal · robmcquadeCompose a strong condition for Claude Code's built-in /goal autonomous loop, then hand it back ready to paste. /goal keeps Claude working across turns until a fast evaluator model (Haiku by default) judges the condition met — and that evaluator sees only the main transcript, not your files and not a subagent's internal work unless Claude restates it, so the condition must be provable from what Claude surfaces. This skill turns a rough aim into a well-formed condition — the three things the docs call for (a measurable end state, a stated check, constraints that must hold) plus the turn or time cap the docs recommend separately — validates it against the transcript-only rule and the common failure modes, and explains the residual risks. User-invoked only. Adaptive — refines a draft you pass, or drafts from the project and asks only the gaps from a blank slate. Pass deep to adversarially stress the drafted condition with a fresh-context subagent.
- ▌ Pressure Test · robmcquade bundleStress-test substantive work — surface strengths, weaknesses, hidden assumptions, failure modes, and adjacent ideas, then return prioritized, actionable fixes (each weakness plus how to fix it). Recommends the fixes; does not produce the rewrite (that is /improve's job). User-invoked only. In Claude Code a deterministic orchestrator auto-escalates security, PII, and credential artifacts to deep, fans out fresh-context subprocess critics with schema-forced output, firewalls sensitive content away from web search, runs an independent verifier, and emits a verified findings manifest; without a shell it runs the same lenses inline and discloses that the enforced gate did not run. Use on durable artifacts (files, frameworks, memos, redlines, decisions, contracts, public content, prompts, proposals) and multi-step work where early choices constrain later ones. Apparent narrowness ('just review this') can mask real stakes. Skip single-fact lookups and content already endorsed and ready to execute.
- ▌ Evaluate AI Merits · robmcquadeEvaluate the merits of external content about Claude, Anthropic products, Claude Code, MCP servers, AI agents, prompt engineering, or AI tooling broadly. Adds AI-specific verification on top of evaluate-merits — routes to primary sources, applies heightened date-of-recording sensitivity (AI tooling ships weekly), flags AI-specific conflations (built-in vs custom add-ons, model-version vs universal guidance, surface differences between claude.ai/Code/API, deprecated features, opinion vs Anthropic guidance, preview vs GA). Use whenever the user shares content on Claude or AI tooling. Composes with evaluate-merits when both apply. Do NOT use for direct AI questions not referencing external content — those are web search and answer.
- ▌