Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →jpoindexter/adaptation-strategies
Decision framework for making a model do a task — choosing between in-context learning, RAG, fine-tuning (LoRA/QLoRA/DPO), and distillation. Maps the gap you have (knowledge / behavior / capability / cost) to the right lever, names when each is the WRONG tool, and covers data requirements, eval, catastrophic forgetting, and combining methods.
Passed 21 of 22 test cases with the skill, and 20 of 22 without it.
1 fixed20 passed either way1 still failing0 broke
Tokens, turns and time describe resource use, not a dollar price.
A test case the skill fixed. Same model, same prompt, with and without the skill.
A medical assistant must query fluctuating clinical trial databases while strictly adhering to a concise, standardized hospital report style. How should these two requirements be handled together?
| Checked | Without skill | With skill |
|---|---|---|
| Recommends combining retrieval-augmented generation for trial data with fine-tuning for report style | Missed | Passed |
The instructions your model reads when the skill loads.
You have a base model. It doesn't do the task well enough. You have four levers to change that, and they are not interchangeable — each changes a different thing. Picking the wrong one is the most expensive recoverable mistake in applied AI: weeks of labeling, a training run, and an eval suite, all to solve a problem a 200-token prompt would have fixed.
The whole discipline reduces to one question: what kind of gap do you have?
| Gap | Symptom | Right lever |
|---|---|---|
| Knowledge | Model doesn't know a fact, doc, or current state | RAG |
| Behavior / format | Model knows enough but won't reliably do it the way you need (tone, structure, JSON, domain register) | Fine-tune (after prompt+RAG exhausted) |
| Capability | Model literally can't — reasoning depth, hard task it fails at any prompt | Bigger model, then ICL to steer it |
| Cost / latency at scale | Quality is fine, the bill or the p95 isn't | Distill or fine-tune a small model |
Memorize that table. Everything below is detail on each lever and the failure modes of using one in the wrong row.
Prompt engineering, few-shot examples, instructions, in-prompt context. Changes behavior at inference time only — zero weight updates.
Retrieve relevant chunks from a knowledge source and inject them into the prompt. For KNOWLEDGE, freshness, and attribution — not for behavior. (Full retrieval/chunking/eval detail lives in the rag skill; here we only place it among the levers.)
Update the model's weights on task data. For BEHAVIOR, format, style, domain register, and latency-by-shrinking-the-prompt. Variants:
| Variant | What it does | When |
test cases passed, vs 20 / 22 without
More skills from the same source and nearby in the registry, with their headline benchmark lift.