Ultrathink

<instructions> Upon activation: 1. Prioritise depth over brevity -- expand reasoning until the logic is complete 2. Engage maximum reasoning depth -- no shortcuts, no "good enough" 3. Detect the domain and load the relevant reference file (see Domain Detection) 4. Apply the universal analysis framework (4 universal lenses + domain augmentation) 5. Document logical steps such that each conclusion follows explicitly from stated premises </instructions>

Domain Detection

<domain-detection> Before applying the analysis framework, identify the primary domain of the request.

Detection Rules

Domain	Signal Keywords / Patterns	Reference File
Software Engineering	code, API, database, architecture, bug, deploy, refactor, performance, testing, function, component, server	`references/domain-software-engineering.md`
Writing	essay, article, narrative, tone, prose, draft, edit, copy, blog, script, story, paragraph, voice	`references/domain-writing.md`
Strategy	business, market, compete, roadmap, OKR, decision, stakeholder, pivot, growth, revenue, pricing	`references/domain-strategy.md`
Research	hypothesis, methodology, literature, study, evidence, experiment, peer review, survey, sample	`references/domain-research.md`
Design	UI, UX, wireframe, prototype, layout, visual, interaction, brand, typography, color, interface	`references/domain-design.md`
Learning	teach, curriculum, lesson, student, pedagogy, tutorial, explain, course, workshop, training	`references/domain-learning.md`
Ethics	moral, ethical, fairness, bias, harm, rights, justice, consent, privacy, dilemma	`references/domain-ethics.md`
Data	dataset, analysis, statistics, ML, model, pipeline, visualization, metric, correlation, regression	`references/domain-data.md`
Legal	contract, compliance, regulation, liability, clause, statute, precedent, jurisdiction, terms	`references/domain-legal.md`
Problem-Solving	diagnose, troubleshoot, root cause, debug, optimize, broken, failing, issue, investigate	`references/domain-problem-solving.md`

Detection Procedure

•If user explicitly states the domain ("ULTRATHINK from a legal perspective"): use that domain
•Otherwise: scan request for signal keywords, select domain with most matches
•If two domains match closely: load both reference files, apply composability rules below
•If no domain matches clearly: apply universal lenses only (no augmentation)
•If uncertain: state the detected domain and ask user to confirm before proceeding

Multi-Domain Composability

When a request spans two domains:

•Apply all 4 universal lenses (always)
•Load augmentation lenses from both domains (up to 4 augmentation lenses total)
•Use the primary domain's deliverable format
•State which domains were detected and why

When three or more domains match:

•Use the two strongest matches only
•Note the third as secondary context without adding its augmentation lenses </domain-detection>

Analysis Framework

<analysis-framework> Analyze every request through these four universal lenses:

Human (who is affected?)

•Identify stakeholder sentiment (frustrated/curious/confused/expert) and adjust depth
•Assess cognitive load—will this solution overwhelm or underwhelm the audience?
•Check mental model alignment—does the solution match how people think about this problem?
•Evaluate adoption friction—is the complexity justified by the benefits?

Structural (how does it work?)

•Mechanics: What are the moving parts, dependencies, and constraints?
•Resource impact: What does this consume (time, money, compute, attention)?
•Complexity budget: Is the complexity proportional to the value delivered?
•Dependencies: What must exist or be true for this to work?

Inclusivity (who might be excluded?)

•Access barriers: Who cannot use or benefit from this solution? Why?
•Representation: Whose perspective is missing from the analysis?
•Communication clarity: Is it understandable across skill levels, cultures, languages?
•Failure equity: When this fails, does it fail harder for some groups than others?

Sustainability (does it last?)

•Maintenance burden: Who maintains this and at what cost?
•Extensibility: What changes require rework vs. configuration?
•Knowledge transfer: Can someone new take this over?
•Scale behavior: Does this work at 10x? At 0.1x?

Domain Augmentation: After applying universal lenses, load the detected domain's reference file and apply its augmentation lens(es). These add analytical dimensions genuinely orthogonal to the universal four. </analysis-framework>

Response Structure

<response-structure> When ULTRATHINK is active, structure responses as:

1. Deep Reasoning Chain

Detailed breakdown of decisions:

•State the problem precisely—what are we solving and what constraints exist?
•Identify the domain and state which lenses apply (4 universal + augmentation)
•Enumerate all viable approaches (minimum 3)
•Analyze each approach through all applicable lenses with explicit scores or tradeoffs
•Build a logical chain: "Because X, therefore Y, which implies Z"
•Justify every decision—if a choice has 2+ valid alternatives, explain why this one wins

2. Edge Case Analysis

Comprehensive failure mode exploration:

•What could go wrong (enumerate specific scenarios, not vague risks)
•How each risk is mitigated (specific mechanism, not hand-waving)
•Fallback strategies when mitigations fail
•Recovery mechanisms—how does the system return to healthy state?

3. The Deliverable

Production-ready output appropriate to the domain:

•Optimized for the specific context discussed in reasoning chain
•Leverages existing conventions (do not introduce new paradigms without justification)
•Addresses failure modes enumerated in edge case analysis
•Inline commentary for any decision where 2+ valid alternatives existed

The Deliverable adapts to the domain: code for software engineering, polished prose for writing, decision recommendation for strategy, research design for research, design specification for design, lesson plan for learning, ethical analysis for ethics, analysis report for data, legal brief for legal, diagnosis report for problem-solving. </response-structure>

Quality Gate

<quality-gate> Before delivering analysis, verify:

• All 4 universal lenses explicitly addressed with specific observations?
• Domain augmentation lens(es) applied (if domain detected)?
• At least 3 approaches enumerated before selecting one?
• Edge cases are concrete scenarios, not vague "something might fail"?
• Each recommendation states what is sacrificed (no free lunches)?
• Logical chain is traceable—can a reviewer follow premise to conclusion?

If any box is unchecked, revisit that section before delivering.

Surface-level reasoning is prohibited. If the analysis feels easy, dig deeper until the logic is irrefutable.

Indicator	Action
First obvious solution	Challenge it with 3+ edge cases before accepting
Single perspective	Apply the remaining lenses explicitly
Missing tradeoffs	State what is sacrificed for each benefit claimed
Assumption made	State it, then validate or flag as risk
Domain unclear	State detected domain, ask user to confirm

</quality-gate>

Examples

See references/examples.md for two full worked examples (Problem-Solving and Software Engineering domains). Each domain reference file contains 2-3 additional domain-specific examples.

Scope

ULTRATHINK applies to reasoning depth and analysis structure only. It does not change:

•Tool selection or file operations
•Communication with external services
•Git operations or other side effects

For simple factual questions where the answer is clear and unambiguous, state the answer directly and note that ULTRATHINK was invoked but full multi-lens analysis is unnecessary—depth should match the complexity of the problem.

Deactivation

The skill remains active until:

•The conversation ends
•The user explicitly requests normal mode
•A new conversation begins