From model to method: what each piece is, what it is for and how they fit together. Organized in four layers, from what thinks to what guarantees quality.
A neural network that, given input text, generates the most likely continuation. It reasons, writes and decides, but on its own it remembers nothing between calls and cannot act on the world.
A brilliant expert in a windowless room: they only know what you slide under the door.
Claude, GPT, Gemini, Llama
API call: text in → text outEverything the model “sees” in one call: instructions, history, files read and tool results. It is finite (measured in tokens) and it is the scarcest resource to manage.
The workbench: whatever does not fit on it does not exist for the model.
200K tokens ≈ a long book
Filling it with logs degrades reasoningFixed text injected at the start of every session to define role, rules and project conventions. It is always loaded, so it must be short and high-value.
A new hire’s onboarding: “this is how we work here”.
CLAUDE.md
- Use pnpm, never npm
- Run the tests before committingInformation that outlives the conversation: decisions, preferences, project status. The model does not store it; the system does, and puts it back into the context when needed.
The expert’s notebook: what they write down today, they read tomorrow.
memory/decisions.md
“ADRs live in /docs/adr”A function with a name, a description and a parameter schema that the model can ask to run. The model does not run it: it emits the call, the harness executes it and returns the result.
The hands: the brain decides “open that file” and the hand does it.
read_file(path)
bash(cmd)
create_ticket(title, desc)An open standard for connecting AI applications to external systems. An MCP server exposes tools, resources and prompts; any compatible client can use them without a custom integration.
The USB-C of AI: one connector for any peripheral.
MCP servers for Jira, GitHub,
Postgres, Slack, Zendesk…A folder with a SKILL.md (instructions) and, optionally, scripts and templates. Only its description takes up context; the full content is loaded when the task needs it.
The procedures manual on the shelf: you only open it when you need it.
skills/create-adr/
SKILL.md
template.md · validate.pyBundles several pieces (skills, commands, subagents, hooks and MCP servers) so they can be installed in one go and shared, versioned, with a whole team.
A toolbox assembled for a specific trade.
plugin “engineering”:
code-review + deploy-checklist
+ Jira MCPThe software that turns a model into an agent: it runs the loop, calls the tools, manages context and permissions, loads skills and fires hooks. The same model performs very differently depending on the harness.
The car around the engine: chassis, steering wheel and brakes.
Claude Code · Cursor · Codex CLI
Agent SDK to build your ownA system that receives a goal and decides by itself which steps to take, which tools to use and when it is done. Unlike a workflow, the path is not programmed in advance.
An employee you give a goal to, not a list of clicks.
“Fix the failing test in CI”
→ reads logs, edits, runs, verifiesThe cycle the agent repeats: think → act (call a tool) → observe the result → decide the next step. It ends when the goal is met, a limit is reached or a human is needed.
A pilot constantly correcting course, not a single shot.
edit → tests → 2 failures
edit → tests → green → doneAn agent launched by another one for a subtask. It works in its own context window and returns only the conclusion, which allows parallel work and keeps the main context clean.
A project manager who delegates to specialists and gets summaries back.
Explorer: searches 400 files
Reviewer: audits the diff without biasScripts the harness always runs on lifecycle events: before or after a tool is used, when a task ends, when a session starts. They do not rely on the model “remembering”: they enforce rules.
Automatic emergency braking: it acts even if the driver is distracted.
PostToolUse → format on edit
PreToolUse → block changes to .envA set of cases with expected outcomes to measure whether a prompt, skill or agent works, and to catch regressions when the model or the instructions change.
Unit tests, but for something non-deterministic.
50 real, labeled tickets
Does the agent classify them correctly?
Measure before and after the changeA methodology where you first agree on what to build and why; then the agent implements against that spec and is verified against it. It replaces “vibe coding” with reviewable work: intent is reviewed before code.
Blueprints before construction: you sign off the plans instead of improvising with a brick in your hand.
GitHub Spec Kit · Kiro · BMAD
spec.md → plan.md → tasks.md → codeA tool is a specific function. MCP is the standard protocol for publishing tools (and data) from an external system. MCP is how they arrive; the tool is what gets called.
MCP gives access: connect to Jira. A skill gives know-how: how to triage an incident using Jira. They do not compete; they combine.
CLAUDE.md is always loaded; a skill, only when needed. What applies to everything goes in the instructions; what is specific to one task goes in a skill.
A skill is guidance the model chooses to follow. A hook is a rule the system enforces. If something must always happen, it is a hook.
The agent is the behavior: a model deciding in a loop. The harness is the infrastructure that makes it possible. Claude Code is a harness; what happens inside it is an agent.
In a workflow the steps are programmed; in an agent the model decides them. If the path is always the same, a workflow is cheaper and more reliable.