Customize: Projects, Skills, RAG, Fine-Tuning
This content was developed with AI assistance and is regularly reviewed for accuracy.
Snapshot: October 2026. Model names, prices, and product features change monthly; What's New in AI is the maintained chronology.
Sooner or later you will hit the limit of a single brief. The agent keeps forgetting your style guide, it does not know your product catalog, or it formats the same report differently every week. The fix is customization, and there are five levels of it. This page is a short decision guide: start at the cheapest level that solves the problem, and climb only when you can name what the level below failed to do.
Learning focus
- Name the five customization levels and what each one changes
- Match a symptom ("it forgets," "it doesn't know," "it's inconsistent") to the right level
- Recognize why fine-tuning is rarely the first answer in 2026
The customization ladder
Each level costs more to build and maintain than the one before it. Most people stop at level two or three, and that is usually the right call.
| Level | What it is | Solves | Effort |
|---|---|---|---|
| 1. A better brief | Objective, context, constraints, output, and a check, written once per task | Vague or inconsistent single results | Minutes |
| 2. Projects, custom instructions, memory | Standing context the assistant carries into every chat in a workspace | "I keep re-explaining who I am and what this is" | An hour |
| 3. Skills | A SKILL.md folder with step-by-step know-how, loaded when the task matches | "It does this recurring job differently every time" | An afternoon |
| 4. Connectors and RAG | Live access to your files and systems, or retrieval over a large document set | "It doesn't know our data" | Days to weeks |
| 5. Fine-tuning | Training a model further on your own examples | "Even with all of the above, the behavior or style is wrong at scale" | Weeks, plus ongoing upkeep |
Level 1: a better brief
Before customizing anything, check the brief. Many "the AI doesn't get it" problems are missing context, a missing output format, or no way to verify. A brief with the six fields from Brief: How to Ask AI for What You Want fixes more problems than any tool on this page, and it costs nothing.
Level 2: projects, custom instructions, and memory
When the problem is repetition of context, give the assistant a home. A Claude Project, a ChatGPT project, or a Gemini Gem holds your instructions and reference files, and every chat inside it starts with them. Memory carries preferences across chats. Claude's free plan includes up to five projects and paid plans remove that limit (as of October 2026).
A good test for this level: if you could explain the context to a new colleague in a one-page document, it belongs in a project. Supply Context walks through setting one up.
Level 3: skills
When the problem is an inconsistent procedure, write a skill. A project tells the assistant who you are and what you are working on; a skill tells it exactly how to do one job, step by step, with examples and a check at the end. Because Agent Skills is an open standard, a skill also travels with you between Claude, Claude Code, and other compatible agents. This is the level that replaced most of what people used to build as Custom GPTs, which individual ChatGPT accounts can no longer create as of August 2026. Skills, AGENTS.md, and CLAUDE.md shows you how to write one.
Level 4: connectors and RAG
When the problem is that the model does not know your data, you have two options.
Connectors give the agent live access to a system: your Drive, your CRM, your database. They are built on MCP, and for most teams they are the first answer, because the data stays current and you do not maintain a copy. Scope each connector to the minimum access the job needs. The MCP sub-track covers connecting and securing them.
Retrieval-augmented generation (RAG) indexes a large document collection so the system can find the relevant passages and hand them to the model with each question. Reach for RAG when you have thousands of documents, need answers grounded in specific passages with citations, or must control exactly which sources the model may draw on. A help-center assistant over five years of support articles is a classic case. RAG: How It Works and When to Use It explains the architecture and its failure modes.
The rule of thumb: connectors for live systems, RAG for large static libraries, and project files for anything small enough to attach.
Level 5: fine-tuning
Fine-tuning trains a model further on pairs of example inputs and ideal outputs, so the behavior you want becomes the default rather than something you ask for. It changes how a model responds. It is a poor way to teach facts, because facts change and retraining is slow; RAG or connectors handle knowledge better.
Fine-tuning earns its cost in a narrow set of situations:
- Very high volume of one well-defined task, where a smaller fine-tuned model can replace a larger general one and cut cost or latency
- A house style or output format that briefs, examples, and skills still cannot hold consistently across thousands of outputs
- Specialized classification or extraction where you have a large set of expert-labeled examples
- Requirements to run a customized model on your own infrastructure, often with an open-weight model
If you do go this way, the work is mostly data. You need a substantial set of clean, expert-reviewed examples that cover the edge cases, a held-out test set the model never trains on, and a plan for re-running the whole process when the base model is replaced. Efficient techniques such as LoRA (low-rank adaptation, which trains a small add-on rather than the whole model) make training cheaper, but they do not reduce the data work. Check the current provider documentation for which models support fine-tuning, since that list changes; Development and APIs and Open-Weight Models are the starting points.
Frontier models improve every few months. A fine-tune built on last year's base model often loses to this year's base model with a good skill and a project. That is the main reason fine-tuning has moved from a common recommendation to a specialist tool.
Choosing in practice
Match the symptom to the level:
| Symptom | Start at |
|---|---|
| "Each answer is a bit off" | Level 1: rewrite the brief with a check |
| "I re-explain the same background every chat" | Level 2: a project with instructions and files |
| "The weekly report looks different every week" | Level 3: a skill with a template and examples |
| "It doesn't know what's in our CRM" | Level 4: a connector |
| "It can't find the answer in 8,000 policy documents" | Level 4: RAG |
| "It's right, but too slow and costly at a million calls a month" | Level 5: consider fine-tuning a smaller model |
Whatever level you choose, measure before and after. Twenty real tasks scored the same way tell you whether the customization helped; Evals for Agents shows you how to build that set.
Key takeaways
- Customize from the bottom of the ladder up: brief, project, skill, connector or RAG, then fine-tuning.
- Projects carry context; skills carry procedures; connectors and RAG carry knowledge; fine-tuning changes behavior.
- Fine-tuning is a poor way to add facts and is rarely the first answer in 2026.
- Measure every customization against a fixed set of real tasks before you keep it.
Next steps
- Continue to: MCP Explained: Connectors for Users, Protocol for Builders
- For level 4 in depth: RAG: How It Works and When to Use It
- For level 3 in practice: Build Your Brief and Skill Library