Skip to the article
The Ontology AI Field notes / 01 7 min read

A field guide to model control

Where can you change an AI?

Skills steer a model from above—through context, instructions and tools. Open weights let builders reach below the interface—into generation, parameters and, sometimes, training.

Explainer 7 min read No maths degree required
One AI system, several places to intervene
01
Application layer Skills, context & tools
change often
API boundary closed models usually stop here
02
Generation layer Sampling & token constraints
change carefully
03
Model layer Weights & architecture
change deeply

“Lower” means closer to the machinery—not automatically better.

Start with the simple version

The short version

A skill changes the model’s working conditions.
Open weights let you change more of the machinery.

Picture an excellent chef in someone else’s kitchen. You can hand them a recipe, point out the ingredients and let them use the oven. That is the role of a skill: it gives a capable model a method, relevant material and approved ways to act.

If you own the kitchen, you can also recalibrate the oven, replace the equipment or rebuild the room. That is the extra control that comes with open weights and a self-managed model stack.

Higher-level change

Skills change how the system works on a task

They add instructions, examples, scripts, references and access to tools. They are quick to edit and can work with both open and closed models.

weights stay fixed

Lower-level change

Open weights expand how the model itself can be run

Builders can host the parameters, inspect intermediate scores, add decoding rules, fine-tune behaviour and choose the infrastructure around the model.

more of the stack is yours

First, fix the vocabulary

Three labels that are often collapsed into two.

The everyday phrase “open model” hides an important difference between having the weights and having the whole recipe used to produce them.

01

Closed model

You send input to a service and receive output. The provider owns the model runtime and chooses which controls the API exposes.

  • Fast to adopt
  • Little infrastructure to run
  • Internals remain behind the service boundary
03

Open-source AI

A stronger claim: freedoms to use, study, modify and share, plus access to the preferred form for modification—not just a checkpoint.

  • Weights and model code
  • Training and data information
  • Terms that preserve the relevant freedoms

The distinction follows the Open Source AI Definition 1.0. Source 1

01 Work above the model

A skill is a reusable way of working.

It packages the know-how an agent needs for a particular kind of task, then loads that know-how when it is relevant.

refund-review /
MD
SKILL.md when to use it + the working method
required
>_
scripts/ repeatable, executable operations
optional
§
references/ policies and detailed knowledge
optional
assets/ templates and resources for the output
optional

A small package, loaded progressively

The model does not need every manual in its head all the time.

An agent can first see a short catalogue of skill names and descriptions. When the task matches, it loads the relevant instructions and pulls in deeper references or scripts only as needed.

What changes Context, procedure and available actions
What does not The model’s learned parameters
01

Discover

The task matches a skill’s description.

02

Load

The method and relevant references enter context.

03

Work

The model follows the method and calls approved tools.

04

Act

Files, systems or drafts change within permissions.

The folder anatomy and progressive-loading pattern follow the Agent Skills specification. Source 2

Move the point of control

Same system. Different intervention.

Select a layer to see what a builder can change there. The highlighted line is the intervention point—not a claim that every product exposes the same controls.

Outcome answer · file · action
service / API boundary
04 Creation layer Data, objective & architecture full model development

Application layer · works with open or closed models

Skill: change the context and available actions

Give the model a procedure, supporting material and approved tools. The model reasons with that temporary context, then the application carries out permitted actions.

Good fit
Procedures, current knowledge, integrations and repeatable work.
Boundary
The model weights remain unchanged.
one task · three control points
Load refund-review → check policy → draft decision → save note

Open weights mostly expand the lower half of this picture. Skills live in the top half—and remain useful no matter which model sits underneath.

02 Work closer to the model

Open weights move the boundary down.

When you can run the model yourself, the score for every possible next token is available inside your own process. That creates new intervention points.

Interactive probability lab

The model proposes. You can move the odds.

Every possible next token receives a score—a logit. Softmax turns those scores into probabilities. Change the intervention and watch the same model arrive at a different distribution.

Illustrative values · not a benchmark
Prompt + predicted token Base weights
The cat in the hat

No extra context and no runtime intervention. The model’s learned association supplies the starting scores.

Next-token probability Softmax total · 100%
hat logit 4.4
73.1%
mat logit 2.8
14.8%
box logit 2.2
8.1%
car logit 1.5
4.0%
Current winner Four-token toy vocabulary
base logit + no modifier softmax probability

“hat” wins because the base scores encode a strong learned association.

A concrete example of runtime-level control is a logits processor in Transformers, which receives token scores and returns processed scores. Source 3

Choose the intervention, not the ideology

Use the highest layer that reliably solves the problem.

Higher layers are faster to change and easier to inspect. Lower layers offer deeper control, but usually demand more data, evaluation, infrastructure and care.

Need Best first move Why Model access
Teach a repeatable procedure Skill Readable, versionable and quick to update Open or closed
Use live systems or change files Skill + tools Puts action behind explicit permissions Open or closed
Guarantee an output grammar Decoding constraint Prevents invalid tokens during generation Self-hosted or supported API
Shift persistent model behaviour Fine-tune / adapter Changes learned behaviour across prompts Weights or provider service
Own, inspect and replace the runtime Open weights Moves infrastructure and generation under your control Downloadable parameters
Ask 01

Does this knowledge change often?

Keep changing policies and procedures in context, references or tools—not baked into weights.

Ask 02

Must the output obey a hard rule?

Use validation or constrained decoding. A beautifully written instruction is still an instruction.

Ask 03

Do you need to own the boundary?

Open weights matter when deployment, privacy, latency or deep inference control must be yours.

The idea to keep

Skills are editable working methods.
Open weights are editable machinery.

They are not rivals. A strong AI system often uses both: a controllable model underneath, and clear, auditable skills above it that connect intelligence to real work.