Self-critique and Constitutional AI reasoning skill. Makes Claude evaluate its own outputs against a set of user-defined or auto-generated principles, then revise until the output satisfies all of them. Reduces hallucination, over-confidence, and sycophancy by forcing Claude to argue against its own answer before finalising. Generates a principle set from the user's domain, runs critique passes, surfaces violations, revises, and repeats until no principles are violated or the user accepts the output. Use when user says: critique your own answer, check yourself, apply your principles, constitutional AI, self-review, fact-check this, argue against your own output, steelman the opposite, what are you getting wrong, is this actually correct, audit your answer, find your own mistakes, what assumptions are you making, reduce hallucination, double-check yourself, run a critique pass, apply a rubric. Do NOT activate for: creative work where principles would suppress quality, requests that explicitly want a single con
2026-05-26
Constraint-based problem solving skill. Finds elegant solutions when the design space is heavily restricted — budget caps, time limits, technology mandates, team size, ethical bounds, regulatory requirements. Treats constraints as the design material, not the obstacle. Surfaces constraint violations early, ranks constraints by hardness, and finds solutions that satisfy all hard constraints while optimizing within soft ones. The tighter the constraints, the sharper the solution. Use when user says: very limited budget, we can't use X, strict deadline, only N people, regulatory constraint, tech stack is fixed, solve with these restrictions, must work within these limits, strong constraints, can't change the requirements, work within the box, solve this but we can only, tiny budget, no budget, works with these limitations, solve under constraint. Do NOT activate for: open-ended problems with no constraints, problems where the first step is to challenge the constraints themselves (use pre-mortem instead). First r
2026-05-26
Cross-domain synthesis skill. Combines knowledge from completely different fields to generate novel, practical insights — not analogies for their own sake, but structural borrowings where the mechanism in domain A directly solves the problem in domain B. Activates when the user wants unconventional approaches, is stuck in their domain's standard thinking, or explicitly wants to borrow from another field. Finds the structural isomorphism — not just a surface metaphor. Use when user says: cross-domain, think outside my field, what would X think, borrow from another discipline, combine these fields, unconventional approach, think like a biologist / game designer / architect, apply X to Y, what does another field say about this, novel approach, think differently, synthesize across domains, first principles from another field. Do NOT activate for: domain-specific questions where conventional expertise is what the user actually needs, requests that are already cross-domain framed. First response: "Cross-Domain Synt
2026-05-26
Rapid iterative refinement skill. Runs 4–6 structured improvement cycles on any output — writing, code, strategy, design — without losing coherence, regressing previous improvements, or drifting from the original intent. Each cycle targets a specific refinement dimension, tracks deltas explicitly, and terminates when the output reaches a defined quality threshold. Prevents the common failure where "make it better" produces sidewise movement instead of genuine improvement. Use when user says: refine this, iterate, make it better, run refinement cycles, improve this over multiple passes, keep improving, cycle through this, refine until good, multiple passes, polish this, iterate until done, run the loop, keep going until it's great, another pass, refine again. Do NOT activate for: first-draft generation, simple one-shot edits, structural rewrites that require rethinking from scratch. First response: "Iterative Refinement active. Paste the output to refine and describe what 'done' looks like — quality bar, inten
2026-05-26
Large document and codebase synthesis skill. Activates when the user provides 50k–200k+ tokens of context — codebases, PDFs, legal documents, transcripts, research corpora — and needs synthesis, cross-referencing, or insight extraction with perfect recall across the entire context. Not a summarizer. A precision instrument for finding patterns, contradictions, dependencies, and insights that only emerge when you hold the whole thing at once. Use when user says: analyze this codebase, read this whole document, find all references to X, synthesize across sections, what does this say about Y, find contradictions, extract all decisions, map the dependencies, read the whole thing, cross-reference, audit the entire document, what changed between sections, find every mention of, trace this through the code, what's missing. Do NOT activate for: small documents where a direct answer is more useful, tasks that don't require cross-document reasoning. First response: "Large Document Mode active. Paste or attach your docum
2026-05-26
precision-prompt-chaining
Multi-step prompt chaining skill where each step's output becomes the next step's input, with explicit rules, constraints, and reference tokens between steps. Activates when the user needs to build a complex pipeline of dependent Claude calls — research → outline → draft → edit → publish — where downstream steps must be constrained by upstream outputs without hallucination or drift. Designs chains with named output variables, explicit pass rules, validation gates between steps, and failure recovery conditions. Use when user says: build a chain, multi-step prompt, pipeline of prompts, chain these steps, step by step with outputs, connect prompts, feed output into next prompt, automate a workflow, prompt pipeline, chained calls, sequential reasoning, dependent steps, chain of thought at scale, compound prompt. Do NOT activate for: single-turn questions, simple back-and-forth conversation, cases where one prompt is sufficient for the task. First response: "Prompt Chain Builder active. Describe the full workflow
2026-05-26
Adversarial red-teaming skill for code, systems, strategies, and plans. Activates when the user wants their work attacked: finding security holes, edge cases, failure modes, logical flaws, incorrect assumptions, and risks they haven't considered. Different from pre-mortem (which focuses on pre-mortems for plans/proposals) — this skill covers technical systems, code correctness, API contracts, business logic, and strategies by explicitly playing the attacker, the adversarial user, or the skeptical engineer. Surfaces the most dangerous findings first. Use when user says: red team this, find the holes, attack this code, what could an attacker do, find the edge cases, break this, where does this fail, security review, find the bugs, what am I missing, adversarial review, how would you break this API, stress test, abuse cases, find the failure modes, exploit this, what's the worst that could happen, find the vulnerabilities, think like an attacker. Do NOT activate for: requests for improvements or feature suggesti
2026-05-26
Silent failure detection skill. Identifies when Claude is being confidently wrong — surfacing hidden assumptions, exposing overconfident claims, probing calibration, and catching hallucination before it causes damage. Uses specific questioning techniques to force the model to reveal its actual uncertainty rather than presenting plausible-sounding answers as facts. Turns overconfidence into a detectable signal. Use when user says: are you sure about this, check your confidence, how do you know that, what are you assuming, could you be wrong, probe this, test your answer, are you hallucinating, what's your actual confidence, stress test your answer, expose your assumptions, calibrate this, fact check yourself, where could you be wrong, question your own answer, is this actually true, what if you're wrong, what evidence would change your answer. Do NOT activate for: creative tasks where factual accuracy is not the concern, outputs the user explicitly wants without critique. First response: "Silent Failure Detect
2026-05-26
