Learn what matters. Build what you can prove.

The modules teach the subjects. These six evidence-gated rungs sequence the work. You keep one product, ship it in the first night, and make its environment more capable as you progress.

60 guided hours12 hours freeRequired excerpts includedOptional depth has no clock

The product contract

- one identifiable user
- one repeated job
- one primary input
- one valuable output
- one deterministic thin slice that can ship in the first week

Do not build for this course

- generic chatbots
- social networks or marketplaces
- multi-agent companies
- medical, legal, or financial decision systems
- anything needing more than a week to become useful

Designed for interrupted nights

Relative deadlines, 45/90-minute sessions, saved commits, a seven-day recovery protocol, scaffold removal, and a help ladder keep one missed week from becoming a restart.

Practice capability	Foundations	Prompting	Retrieval & RAG	Context Engineering & Prompt Injection	Agents & Tools	Evals	Production AI	Safety & Governance
1. Turn intent into a bounded, visible change	Orient	Apply now	Save for later	Orient	Save for later	Orient	Apply now	Apply now
2. Inspect and verify the change	Orient	Apply now	Save for later	Apply now	Orient	Apply now	Apply now	Orient
3. Make the repository self-checking	Save for later	Orient	Orient	Apply now	Apply now	Apply now	Apply now	Apply now
4. Isolate and orchestrate work	Save for later	Orient	Orient	Apply now	Apply now	Apply now	Orient	Apply now
5. Run bounded feedback loops	Save for later	Apply now	Orient	Apply now	Apply now	Apply now	Apply now	Apply now
6. Own and defend a public system	Apply now	Orient	If used	Apply now	Apply now	Apply now	Apply now	Apply now

Rung 1 / 6 12 hoursFree in full

Turn intent into a bounded, visible change

Ship a deterministic thin slice to a real URL, then prove what changed from evidence.

You know it works whenA stranger can open the URL, complete one useful job, and see the intended output.

You should be able to

name one user, repeated job, primary input, and valuable output
turn that contract into a bounded agent task
use voice, screenshots, and basic terminal commands to supply evidence
check plan limits and distinguish a usage ceiling from a software failure
ship a deterministic thin slice to a public URL

Start with

Prompt Engineering Interactive TutorialComplete the exercises on clear instructions, examples, and output shape.
From Experiments to DeploymentsRead the sections that separate a demo from a deployable product slice.
Running Codex Safely at OpenAIRead the sandbox, approval, and network-boundary sections.

Save for later

- neural-network training and transformer math
- retrieval systems and agent architecture
- authentication, billing, and full observability
- formal eval harnesses

Practice · free in full

Make: Choose one real product idea. Write AGENTS.md before the first feature, ask the agent for a plan, constrain it to one thin slice, and deploy that slice to a URL someone else can open.

Codex path

Use AGENTS.md, plan mode, a workspace-write sandbox, scoped approvals, /status, and the repository's own build command.

Claude Code path

Import @AGENTS.md from CLAUDE.md, review /memory, use /doctor to diagnose context problems, and match the same permission boundary.

Seeded failure

The first deployment intentionally receives a missing environment value. Diagnose the hosting log before asking the agent to edit code.

Budget and stop condition

Check /status before starting. Stop after 90 minutes or at 60% of the current usage window. If a limit resets mid-task, save a commit and one next executable action; the agent stopping may be an access event, not a code failure.

You know it works when

another person can open the URL and complete the one promised job; the saved log identifies the seeded failure; the final diff stays inside the written scope

Save this

product contract, AGENTS.md, public URL, before-and-after screenshot, deployment log, final diff, and commit hash

Codex is the reference harness. Tushar holds an OpenAI Codex Ambassador role. The course teaches mechanisms, uses Claude Code as a first-class adapter, and names real differences instead of claiming parity. Tool-specific instructions are versioned and reviewed every 4–6 weeks.

Rung 2 / 6 8 hoursPaid guided lab

Inspect and verify the change

Read the diff, trace the actual error, and reject plausible-looking code that does not satisfy the contract.

You know it works whenA reviewed change with a written claim-to-evidence map.

You should be able to

explain a diff in plain language
trace a failure from logs
identify hallucinated packages and deprecated patterns

Start with

Prompt EngineeringRead the sections on inputs, constraints, examples, and structured output.
Effective Context Engineering for AI AgentsRead the sections on selecting and pruning context.
Evaluation Best PracticesRead the sections on task-specific tests and failure analysis.

Save for later

- embeddings
- multi-agent systems
- unattended loops
- full system-card analysis

The knowledge stays open. Payment unlocks the seeded failures, adapter labs, grading, executable verification skills, scaffold removal, and graduation evidence path.

Paid lab

Rung 3 / 6 14 hoursPaid guided lab

Make the repository self-checking

Turn expectations into tests, logs, typed boundaries, and scripts the agent can run itself.

You know it works whenA fresh agent can change the product and verify its own work without asking how.

You should be able to

externalize a requirement as a check
write an executable verification skill
make logs and failure states agent-legible

Start with

Prompt Injection ExplainedRead the full argument and turn the boundary between instruction and untrusted input into a check the repository can run.
Writing Effective Tools for AgentsRead the sections on narrow tools and inspectable outputs.
Demystifying Evals for AI AgentsRead task, trial, grader, transcript, and outcome definitions.
Running Codex Safely at OpenAIRead the least-privilege and audit sections.

Save for later

- custom retrieval
- multi-agent fan-out
- unattended loops
- frontier training

The knowledge stays open. Payment unlocks the seeded failures, adapter labs, grading, executable verification skills, scaffold removal, and graduation evidence path.

Paid lab

Rung 4 / 6 8 hoursPaid guided lab

Isolate and orchestrate work

Split work only when isolation improves evidence, then reunite it through explicit interfaces.

You know it works whenTwo bounded work streams that do not overwrite each other or hide failed assumptions.

You should be able to

choose sequential versus parallel work
isolate changes
separate maker and checker roles

Start with

A Practical Guide to Building AgentsRead the multi-agent orchestration and guardrail sections.
Your AI Product Needs EvalsRead the product-specific eval and error-analysis sections.
Practices for Governing Agentic AI SystemsRead the operator, oversight, and accountability sections and apply them to how isolated work streams get approved and reunited.

Save for later

- fan-out before one-agent reliability
- product-irrelevant RAG
- unbounded parallel work

The knowledge stays open. Payment unlocks the seeded failures, adapter labs, grading, executable verification skills, scaffold removal, and graduation evidence path.

Paid lab

Rung 5 / 6 9 hoursPaid guided lab

Run bounded feedback loops

Design the system around the agent with budgets, external feedback, and hard stop conditions.

You know it works whenA loop that improves one measured behavior and stops safely when evidence stops improving.

You should be able to

set a cost ceiling
define a stop condition
distinguish coding, developer-feedback, and user-feedback loops

Start with

Building Effective Agents (essay)Read the workflow patterns and stopping considerations.
Evaluation Best PracticesRead the sections on tracking a measured behavior over time and knowing when it has stopped improving.
What We Learned from a Year of Building with LLMsRead feedback, operations, and iteration sections.
Practices for Governing Agentic AI SystemsRead the operator, oversight, and accountability sections.

Save for later

- multi-agent scaling
- unrestricted unattended execution
- custom embeddings without product need

The knowledge stays open. Payment unlocks the seeded failures, adapter labs, grading, executable verification skills, scaffold removal, and graduation evidence path.

Paid lab

Rung 6 / 6 9 hoursPaid guided lab

Own and defend a public system

Operate a product through real users, failures, costs, abuse cases, and model change.

You know it works whenA public demo and evidence package that can survive technical review.

You should be able to

run a launch review
monitor cost and failures
read a system card critically
defend technical decisions with evidence

Start with

Deep Dive into LLMs like ChatGPTReview the model-mechanism sections needed to evaluate claims and limits.
Prompt Injection ExplainedRead the complete argument and map it to the product boundary.
Demystifying Evals for AI AgentsRead the production-evaluation and human-review sections.
From Experiments to DeploymentsRead reliability, monitoring, and rollout sections.
GPT-5.6 System CardAudit one reported capability or safety claim against its evidence.
Anthropic Economic Index: Economic PrimitivesSeparate observed usage data from forecasts and advocacy.

Save for later

- specialist depth unrelated to the shipped product

The knowledge stays open. Payment unlocks the seeded failures, adapter labs, grading, executable verification skills, scaffold removal, and graduation evidence path.

Paid lab

How finishing works

- Set deadlines relative to your own start date and sign a commitment contract.

- Work in 45- or 90-minute sessions that end with a saved commit and one next executable action.

- After seven missed days, recover from the last passing evidence instead of restarting the course.

- Remove scaffolds rung by rung so later work proves independence.

- Use the help ladder in order: evidence, repository guidance, official documentation, agent, then human help.

- Report graduates, active learners, and auditors separately; do not count auditors as failed completers.

Free and open source forever

The catalog, eight-module map, all reading, capability statements, save-for-later guidance, and both practice briefs in every module. Community suggestions continue through GitHub.

What the paid path adds

Supervised sequencing, prepared failures, Codex and Claude adapters, executable checks, grading, loop budgets, scaffold removal, and the evidence package that proves you walked the path.

Return to the subject map ↑