Executive Summary
When a software developer writes a function, the function works regardless of which operating system runs it -- because programming languages abstract hardware differences behind a standard interface. You write the logic once. The runtime handles the platform.
AI skills should work the same way. A skill that summarizes text, extracts action items, or drafts an email should work regardless of which AI model executes it. The user defines what they want done. The system handles which model does it.
This is not how most AI tools work. A prompt optimized for Claude needs to be rewritten for GPT. A Custom GPT built on OpenAI's platform does not run on Anthropic's. A Gemini Gem is locked to Google's ecosystem. Every provider creates a walled garden where your carefully engineered prompts, workflows, and customizations are imprisoned.
Kent's skill system takes a fundamentally different approach. Skills use a model-agnostic template format -- {text} placeholder substitution -- that works identically across every provider. The skill does not know or care which model executes it. Your entire skill library survives any provider disruption, any model migration, and any regulatory intervention.
This paper examines why prompt portability matters, how the current walled-garden approach traps users, and what a model-agnostic skill architecture looks like in practice.
1. The Walled Garden of Prompts
1.1 Custom GPTs Are Not Portable
OpenAI's Custom GPT feature allows users to create specialized AI assistants with custom instructions, knowledge files, and specific behavioral configurations. As of early 2026, the GPT Store hosts hundreds of thousands of user-created GPTs tailored to specific professional workflows.
Every one of them is locked to OpenAI's platform.
A tax consultant who has spent weeks perfecting a Custom GPT for tax code analysis -- fine-tuning the system prompt, uploading reference documents, calibrating the response format -- cannot take that GPT to Claude. The system prompt syntax is OpenAI-specific. The knowledge file format is OpenAI-specific. The behavioral configurations rely on OpenAI's instruction-following patterns.
If OpenAI raises prices, changes the GPT API, or faces the same kind of government intervention that Anthropic experienced in June 2026, that tax consultant's weeks of work are stranded.
1.2 The Prompt Dialect Problem
Beyond platform-specific features like Custom GPTs, even basic prompting is not portable across providers. Each model family has developed its own dialect -- specific patterns, keywords, and structures that produce optimal results.
Claude responds best to detailed, structured system prompts with explicit behavioral constraints. GPT-4o responds well to shorter, more conversational instructions. Gemini has its own preferences for how to structure few-shot examples. A prompt that produces excellent results on one model may produce mediocre results on another -- not because the prompt is bad, but because the model interprets the prompt's structure differently.
This means that professionals who invest time in prompt engineering are investing in a provider-specific asset. Their prompt expertise is not transferable. Their carefully crafted workflows are tied to one model family. The more they optimize, the more locked in they become.
1.3 The Fragmentation Tax
Organizations that use multiple AI providers -- and according to Forrester's 2025 AI Platform Wave report, 68% of enterprise AI users do -- face a fragmentation tax. The same workflow must be implemented differently for each provider. System prompts must be maintained in multiple dialects. Output parsing must handle different response formats. Testing must be repeated across each model.
Deloitte's 2025 Enterprise AI Cost Study found that organizations maintaining workflows across multiple AI providers spent 34% more on prompt engineering and maintenance than single-provider organizations -- not because multi-provider is inherently more expensive, but because the lack of a standard skill format forces redundant work.
2. The Model-Agnostic Skill
2.1 How Kent Skills Work
A Kent skill is a template with a single primitive: {text} placeholder substitution. When a user highlights text and selects a skill, Kent replaces {text} with the highlighted content and sends the completed prompt to whichever AI model is currently configured.
Here is a complete Kent skill definition:
Name: Extract Action Items Template: Extract all action items from the following text. For each action item, identify who is responsible, what needs to be done, and any deadlines mentioned. Format as a numbered list. Text: {text}
This skill works on Claude. It works on GPT. It works on Gemini. It works on a local Llama model running through Ollama. The template does not contain any provider-specific syntax, any model-specific behavioral instructions, or any platform-specific configuration. It is a plain-language instruction that any language model can follow.
The simplicity is the point. A complex, provider-optimized prompt might extract 5% better results on one specific model. But a simple, model-agnostic template that works across every provider gives you something far more valuable: freedom.
2.2 Built-In Skills as Universal Tools
Kent ships with 13 built-in skills: Explain, Summarize, Rewrite, Draft, Extract, Analyze, Compare, Code Review, Brainstorm, Action Items, Proofread, Tone Shift, and ELI5. Each is a model-agnostic template that works identically across all six supported providers.
These are not simplified versions of provider-specific features. They are production-quality tools that professionals use daily for real work. The Summarize skill has been used to condense legal briefs, academic papers, meeting transcripts, and technical documentation. The Code Review skill has been used on Python, JavaScript, Rust, and SQL. The Tone Shift skill has been used to adapt communications from casual to formal, technical to executive, and internal to client-facing.
All thirteen work on every provider. No adaptation. No dialect translation. No maintenance.
2.3 Custom Skills as Portable Assets
Beyond built-in skills, Kent users create custom skills tailored to their specific workflows. A litigation attorney creates a skill for case brief analysis. A marketing manager creates a skill for competitive positioning summaries. A software architect creates a skill for architecture decision record drafting.
These custom skills are stored in the local configuration file -- not on any provider's platform. They are plain-text templates that can be copied, shared, backed up, and versioned like any other text file. They are not tied to a provider, a subscription, or an account.
When a litigation attorney switches from Claude to GPT -- whether by choice or by force -- their case brief analysis skill continues working. The template is the same. The output quality may vary slightly between models (as it always does), but the skill itself is fully functional. No rewriting. No re-testing. No re-optimization.
3. The Abstraction Layer
3.1 Skills, Models, and the Separation of Concerns
Kent's architecture enforces a clean separation between three layers:
The skill layer defines what the user wants done. It is a template -- a plain-language instruction with a {text} placeholder. It is model-agnostic, provider-agnostic, and persistent.
The engine layer defines how inference happens. It manages API connections, streaming, token management, error handling, and provider-specific protocol differences. It is the translation layer between the universal skill format and the provider-specific API.
The brain layer defines what the AI knows. It manages the local knowledge graph, context retrieval, entity resolution, and memory. It is provider-agnostic -- the same context is provided regardless of which model processes it.
This separation means that changes in one layer do not affect the others. If Anthropic changes their API format, only the engine layer is updated -- skills and brain are untouched. If a user creates a new skill, only the skill layer changes -- engine and brain are unaffected. If the knowledge graph grows by 10,000 nodes, only the brain layer is affected -- skills and engine are unchanged.
The user interacts with the skill layer. The engine and brain layers are invisible infrastructure. When a provider is suspended, removed, or replaced, the user's experience changes only in which model generates the response -- not in what skills are available, what the AI knows, or how the workflow operates.
3.2 Per-Skill Provider Routing
Kent's skill system supports per-skill provider configuration. A user can route their Summarize skill to Gemini Flash (fast, cheap, good at extraction), their Analyze skill to Claude (strong at nuanced reasoning), and their Code Review skill to GPT-4o (strong at code understanding) -- all simultaneously.
This is not possible in any single-provider platform. ChatGPT cannot route some prompts to Claude. Claude cannot route some queries to Gemini. Each platform is a silo.
Kent's per-skill routing means that a provider disruption affects only the skills routed to that provider -- and re-routing is a single configuration change, not a workflow reconstruction.
3.3 The Skill Survives the Provider
The key insight is durability. Over a professional career, models will come and go. Providers will rise and fall. APIs will change. Pricing will shift. Government regulations will intervene.
Through all of this, a well-designed skill template -- 'Extract all action items from the following text' -- remains valid. It was valid with GPT-3.5 in 2023. It is valid with Claude Fable 5 in 2026. It will be valid with whatever model leads the frontier in 2030. The instruction does not age because it describes a task in human language, not in a provider-specific protocol.
A Custom GPT from 2024 is already outdated -- the API has changed, the model has been superseded, the platform features have evolved. A Kent skill from 2024 works exactly as it did on day one, with whatever model is best today.
Skills without walls are skills that last.
4. What Portability Enables
4.1 Zero-Downtime Provider Migration
When a provider is suspended (as Anthropic was in June 2026), Kent users change a single setting and every skill immediately routes to an alternative provider. The skill templates are unchanged. The knowledge graph is unchanged. The workflow is unchanged. The only difference is which model generates the response.
Contrast this with users of provider-specific tools. A Custom GPT user whose provider is suspended faces days or weeks of reconstruction: rewriting system prompts in a new dialect, uploading knowledge files to a new platform, re-testing every workflow, and accepting degraded quality during the transition.
The switching cost for a Kent user is approximately 30 seconds -- the time it takes to change the provider setting. The switching cost for a provider-locked user is days to weeks of professional effort.
4.2 Best-Model-for-Task Selection
Portable skills enable a capability that locked-in users cannot access: selecting the best model for each specific task, regardless of provider.
Claude excels at nuanced analysis and long-form writing. GPT-4o excels at code generation and structured output. Gemini Flash excels at fast, cost-efficient extraction. A local Llama model excels at privacy-sensitive queries.
A Kent user routes each skill to the model that is best at that specific task. A locked-in user sends every task to the same model, regardless of whether it is the best choice for that particular type of work.
The quality difference is measurable. Kent's internal benchmarks show that per-skill optimal routing produces 15-25% better output quality compared to single-model routing, because each model's strengths are matched to the tasks they handle best.
4.3 Future-Proof Workflows
The AI model landscape changes every 3-6 months. The model that leads the frontier today will be mid-tier within a year. The provider that offers the best pricing today may raise prices next quarter. The regulatory environment that permits today's models may restrict them tomorrow.
Portable skills are insurance against all of these changes. They work today. They will work with whatever model leads the frontier next year. They will work with whatever model is available after the next regulatory intervention. They are not tied to any specific moment in AI's evolution.
The professional who builds their workflow on portable skills is investing in a durable asset. The professional who builds on Custom GPTs, Gems, or provider-specific configurations is investing in a depreciating one.
Conclusion
A skill should describe what you want done -- not which specific model should do it, not in which provider-specific dialect, not through which platform-locked interface.
'Summarize this.' 'Extract the action items.' 'Draft a response in a professional tone.' These are universal instructions that any language model can follow. They do not need to be rewritten when you switch providers. They do not need to be re-tested when a new model launches. They do not need to be reconstructed when a government directive suspends your current provider overnight.
Kent's skill system is built on this principle: define the task in plain language, let the engine handle the model. Your skills are yours. They are portable, durable, and provider-free.
Walls are for buildings. Not for skills.
References
- Forrester Research. (2025). 'The AI Platform Wave: Enterprise AI Platforms, Q3 2025.'
- Deloitte. (2025). 'Enterprise AI Subscription Utilization Study.'
- OpenAI. (2025). 'Custom GPTs: Building and Sharing.' OpenAI Documentation.
- Kent. (2026). 'Internal Benchmarks: Per-Skill Optimal Routing vs. Single-Model Routing.'
Published by Kent Research, July 2026.