Frameworks & Glossary

Quick Reference

Every named framework, model, and vocabulary item across both courses, in one place. Use this as the cheat sheet; use the lesson pages for the full treatment.

The Core Four

The atomic unit of agentic coding. The big three of AI coding (Context, Model, Prompt) plus one new dimension, Tools, which turns a prompt into an agent. Everything in both courses reduces back to these four.

Element	What it is
Context	What the agent knows about the codebase, the problem, and the constraints.
Model	The reasoning engine. Use the most capable model and turn on thinking when it matters.
Prompt	How intent, constraints, and expectations are communicated.
Tools	The actions the agent can take. The bridge from reasoning to real-world impact.

If something fails, the root cause is in one of these four.

The 12 Leverage Points of Agentic Coding

Split into in-agent levers (always present, the Core Four) and through-agent levers (external, they flow through the context window).

#	Leverage point	Category	One-liner
1	Context	In-agent	What the agent has to work with.
2	Model	In-agent	The reasoning engine; use the best, enable thinking.
3	Prompt	In-agent	How you communicate the work.
4	Tools	In-agent	The ability to act; view them with a /tools command.
5	Standard out	Through-agent	The agent can only see what you let it see. Log everything.
6	Types	Through-agent	Types tell the story of how information travels; they are information-dense keywords.
7	Documentation	Through-agent	Keep an ai_docs/ folder local so agents do not web-fetch every lookup.
8	Tests	Through-agent	The agent will make mistakes; tests give the leverage to correct them.
9	Architecture	Through-agent	Cuts the search space in half. Consistency beats complexity.
10	Plans	Through-agent	Plans are just prompts scaled up; they live in specs/.
11	Templates	Through-agent	Reusable agentic prompts in .claude/commands/. Three times makes a pattern.
12	ADWs	Through-agent	AI Developer Workflows: agentic prompts wrapped in code and fired by a trigger. The highest leverage point.

The 4 Agentic Coding KPIs

The scoreboard. If you do not measure it, you cannot improve it.

KPI	Direction	Meaning
Size	Up	The amount of work handed off per prompt. Push run times from minutes toward hours.
Attempts	Down	The number of re-prompts to correct. Aim for one-shot solutions. High attempts means a missing leverage point.
Streak	Up	Back-to-back one-shot successes with zero issues.
Presence	Down	How much you must sit and steer. The most important one. Drive it toward zero.

PETER — the 4 elements of AFK agents

The anatomy of an away-from-keyboard, out-of-the-loop agent system. Drop the R (Review) and PETER becomes Pete, which is Zero Touch Engineering.

Letter	Element	Default in the course
P	Prompt input	GitHub issues (title + body = the prompt).
E	Environment	A dedicated device, e.g. a Mac mini the agent fully controls.
T	Trigger	GitHub webhooks via a Cloudflared tunnel, or a local script.
E	(Environment)	(see above)
R	Review	GitHub pull requests. Dropped entirely at the ZTE level.

Mnemonic ordering: Prompt input, Trigger, Environment, Review, spelled PETER.

The closed-loop prompt anatomy

Three parts. Give the agent validators and a direction to fix, and it self-corrects until the job is done.

1. Request — the task, usually a large spec.
2. Validate — the validators: linters, tests, builds, browser drives, LLM-as-judge.
3. Resolve — the direction for fixing whatever the validators catch.

Tests are the rule of law in your codebase. Their value is multiplied by the number of agent executions.

The three-level velocity ladder

flowchart LR
  I[In-loop
hands on keyboard] --> O[Out-loop
AFK / PETER]
  O --> Z[Zero Touch Engineering
drop the review]

Progress is not in-loop versus out-loop; it is in-loop moving toward out-loop. Go chores first, then bugs, then features.

The R&D Framework (context engineering)

There are only two ways to manage a context window.

• R = Reduce the context entering your primary agent (measure it, do not preload MCP servers, prime instead of bloating CLAUDE.md, use output styles).
• D = Delegate context to sub-agents and other primary agents.

The Seven Levels of Agentic Prompt Formats

A tier list of prompt formats, rated on usefulness and skill required.

Level	Format	Adds
1	High-Level Prompt	The one-off, reusable ad-hoc ask.
2	Workflow Prompt	A sequential workflow section (the most important section).
3	Control Flow Prompt	Conditions, loops, early returns.
4	Delegation Prompt	Kicks off other agents; variables to pass to them.
5	Higher-Order Prompt	Passing prompts into prompts.
6	Template Meta Prompt	A prompt that creates other prompts in a set format.
7	Self-Improving Prompt	A dynamic expertise section; agents update agents.

Companion concept, the Stakeholder Trifecta: you write prompts for three audiences now, you, your team, and, most importantly, your agents.

The classes-and-grades maturity model

How mature is your agentic layer. Three classes, each with grades, each grade distinguished by one new dimension.

Class	Grade	Distinguishing dimension
One (in-loop)	1	Prime prompt + memory files.
One	2	Specialized plan prompts + sub-agents (specs/, ai_docs/).
One	3	Custom tools (skills, MCP servers, tool-enabled primes).
One	4	Feedback loops / closed-loop prompts.
One	5	Templates (bug, feature, chore); apps/ with Git submodules.
One	6	Prompt chains / early ADWs.
One	7	Agent experts.
Two (out-loop)	1	An out-loop trigger (webhook / HTTP endpoint).
Two	2	Full AI Developer Workflows (ADWs).
Three (orchestrated)	1	The Orchestrator Agent with CRUD on agents.
Three	2	Orchestrator Developer Workflows.
Three	3	The orchestrator running raw-code ADWs.

Core principles (Dan's rule set)

1. Plan first. 80/20 of agentic success is planning.
2. Three times makes a pattern. Then template it.
3. Close the loop. Put validators in the prompt.
4. Presence down. Reduce the human presence required.
5. AND not OR. Layer tools; do not collapse them.
6. Everything named is a keyword. Semantic naming is implicit context.
7. The prompt is the primitive. Skills, sub-agents, and ADWs all compose prompts.

The philosophy behind all of it: you build the system that builds the system.

Glossary

• Agentic Prompt — a plain prompt that reliably executes long chains of tool calls.
• ADW (AI Developer Workflow) — a reusable workflow combining deterministic code, agentic prompts, and agents, fired by a trigger.
• Agentic layer — the ring of prompts, workflows, and agents around your codebase that teaches agents to operate it; kept isolated from the application layer.
• Meta prompt — a prompt that builds a prompt.
• Higher order prompt (HOP) — a prompt you pass another prompt into.
• Template meta prompt — a prompt that builds prompts in a specific format.
• IDK (information-dense keyword) — a word that points the agent at exact meaning or location; e.g. think hard activates reasoning.
• Agent expert — a self-improving template meta prompt that acts, learns, and reuses expertise from a maintained mental-model file.
• Orchestrator agent — a custom agent that CRUDs and commands a fleet of primary agents.
• ZTE (Zero Touch Engineering) — the North Star: agents ship end to end with no human review.
• Codebase singularity — the moment your agents ship better than you and your team.
• Best-of-N — run multiple agents on one problem, then compose or select the winner. More compute equals more confidence.