From worker to builder. Leave with builder instincts.

start here. turn weekly work into a reusable tool: learn the six-move builder loop, write a five-field brief, and save it for next week. finish by building an interactive webpage from a real sales deck and sending it to your team.

read small ai-written scripts: follow values, check function returns, and trace if branches. if you are starting here, learn these reads first; if you already ran a studio, connect them to its Python.

read nested api, config, and llm data; walk lists and dicts with loops, and catch wrong-layer bugs.

read Python tracebacks to locate failures. catch errors you can handle without hiding unexpected bugs.

trace shared lists and dicts to catch unintended mutation. learn when references, shallow copies, and deep copies preserve your data.

follow files and api calls through a script. learn with-blocks, status families, and bounded retries.

call model APIs, construct messages, and read responses using the basic four-line pattern.

define output schemas, request JSON, and validate with Pydantic before malformed data reaches your application.

connect agents to tools and data through MCP. learn servers, tool discovery, and the shared protocol behind integrations.

build the agent cycle: request a model response, execute tool_use, return results, and repeat until end_turn.

an AI coding tool writes and edits with you; Cursor is one brand, Claude Code another. write prompts that clarify the task before work begins.

review agent-written commits with git’s three-state model, daily commands, and GitHub CLI. learn to undo mistakes.

keep credentials out of git: use .env, os.getenv, and ignore rules, then practice responding to a leak.

read tool calls, results, and stop reasons to locate failures in an agent’s execution trace.

same input, check the output, pass or fail. don't ship on a demo that felt right.

use chunking, embeddings, and vector search to ground model answers in your data. evaluate retrieval and avoid unnecessary infrastructure.

Train one consistent job with reviewed examples and a QLoRA adapter; compare held-out evals with prompt-only without claiming broader intelligence gains.

balance cost, latency, and quality with token math, caching, streaming, batching, and model selection.

when the model lies to your customer. isolate what went wrong, classify the four common failures, and ship a verified fix instead of blaming the model.

connect the prompt, call, validation, trace, eval, and MCP tool in the smallest end-to-end llm feature you could ship. on the developer route, harnesses, image, video, and production tooling follow. later retrieval and prompt-cache cost chapters extend this checkpoint beyond demo inputs.

claude code, cursor, aider, codex cli: learn the four layers around the model api, then build a small harness you can run and break. for actions that change the world, write a repo-based rollback runbook a junior can execute at 3am.

engineer the harness around a model: turn failures into rules, manage context, design long-running loops, enforce hooks, and choose what to build or buy as runtime SDKs evolve.

drive your computer through a text box: type a command and press enter. move around files, make folders, and read files without the mouse — the foundation for later tool chapters.

claude cli brings the chat model to your actual files. install it, sign in, and run your first command. watch ai read, plan, and change files, and learn when to use the cli instead of chat.

codex cli brings another company's ai to your terminal and files. install it, sign in, and run your first command. apply familiar claude cli patterns while learning the differences and when to use each tool.

a claude skill packages your team's playbook for loading when relevant, saving repeated explanations. for team managers: define, share, and provision skills; explore hr, legal, and ops examples; choose skills over one-off prompts; and establish governance before real use.

choose from the 2026 image lineup, write effective prompts, and turn one idea into a hundred ready-to-ship images. compare uses and costs: nano banana 2 for volume, nano banana pro for fidelity, flux, midjourney, ideogram, and gpt-image-2.

learn the 2026 video lineup, camera-control patterns, and cost math for this difficult, expensive, strategically useful medium. decide whether your idea is viable.

ai generates assets; code assembles them. compare and combine hyperframes, remotion, and claude design in data-driven workflows that turn one template into a hundred videos.

tables are the working surface of applied ml. learn rows, columns, missing values, joins, aggregates, and the dataframe habits ai-generated notebooks assume.

most training data starts in a database. learn the select, join, filter, aggregate, and leakage traps that decide whether a model is learning signal or nonsense.

a dataset is a product surface. build ingestion, validation, partitions, manifests, and checkpoints so the next run is not a mystery.

vectors, probability, distributions, correlation, and uncertainty are not trivia. they are how you read model behavior without worshipping it.

labels, splits, baselines, training, prediction, and evaluation. the supervised workflow is the first complete model loop.

not every useful model has labels. cluster, compare, retrieve, and recommend by turning examples into useful neighborhoods.

accuracy is a blunt instrument. learn confusion matrices, precision, recall, thresholds, slices, and failure notes so model quality has evidence.

read tensor code without flinching. tensors, shapes, broadcasting, gradients, and autograd are the grammar of modern deep learning scripts.

the training loop is where models change. learn loss, gradients, optimizer steps, schedules, checkpoints, and the bugs ai ships there.

architecture literacy for builders: convolution, attention, tokens, decoding, kv cache, quantization, and what those choices do to cost and behavior.

features, runs, configs, artifacts, and registries are how ml work becomes repeatable instead of a lucky notebook.

serving a model means handling inputs, versions, routes, batch jobs, ci gates, rollback, and production failures deliberately.

the last mile: logs, drift, alerts, retraining decisions, cloud cost, gpu constraints, architecture docs, demos, and role stories.

Run image models with creative briefs, structured brand kits, interface design tokens, seed and reference locks, and QA filters for fifty consistent assets.

Run Fieldnote's fall relaunch from an evidence-backed content record through approved creative, an export-verified report, and a checked short-clip plan. Every rejection, result, and next decision keeps a receipt.

Audit how approved buyer questions are answered and cited, then build one source-backed asset for a real evidence gap. Manual observations only, with dated receipts and no causal claims from one change.

Check support drafts against policy and banned promises, evaluate macros per intent using fictional cases, and retain human approval and escalation for money, legal concerns and emotional requests. Passing fixtures never authorizes deployment.

Measure client voice from approved copy, check variants and channel budgets, log rejections, verify factual claims before shipping, and record corrections.

Account for dropped rows, verify aggregations on hand-checked samples, trace memo figures, challenge recommendations using the SR 11-7 triad, and require identical reruns.

Define specs and acceptance checks before AI builds, verify results, derive status and go/no-go decisions from evidence, and record one accepted function’s value using comparable hours, finance’s loaded rate and a named owner.

Practice employee-data boundaries, resume-linked screening briefs, named legal review for policies and synthetic decision-record checks. Local fixtures do not establish privacy, compliance or legal defensibility.

Capture versioned SOPs, catch invented steps, distinguish auto-run work from named approvals, surface exceptions and verify report figures. Practice literal identifier masking locally; fixture checks do not authorize disclosure.

Verify every cited authority at its source, pinpoint contract claims, route sensitive documents under counsel-defined rules and record verification and sign-off.

Build a triage agent’s tool contract, metered loop, golden evals, diagnostic trace and incident ratchet. Practice the harness controls that make failures inspectable.

start here. turn weekly work into a reusable tool: learn the six-move builder loop, write a five-field brief, and save it for next week. finish by building an interactive webpage from a real sales deck and sending it to your team.

read small ai-written scripts: follow values, check function returns, and trace if branches. if you are starting here, learn these reads first; if you already ran a studio, connect them to its Python.

read nested api, config, and llm data; walk lists and dicts with loops, and catch wrong-layer bugs.

read Python tracebacks to locate failures. catch errors you can handle without hiding unexpected bugs.

trace shared lists and dicts to catch unintended mutation. learn when references, shallow copies, and deep copies preserve your data.

follow files and api calls through a script. learn with-blocks, status families, and bounded retries.

drive your computer through a text box: type a command and press enter. move around files, make folders, and read files without the mouse — the foundation for later tool chapters.

claude cli brings the chat model to your actual files. install it, sign in, and run your first command. watch ai read, plan, and change files, and learn when to use the cli instead of chat.

codex cli brings another company's ai to your terminal and files. install it, sign in, and run your first command. apply familiar claude cli patterns while learning the differences and when to use each tool.

Build a triage agent’s tool contract, metered loop, golden evals, diagnostic trace and incident ratchet. Practice the harness controls that make failures inspectable.

review agent-written commits with git’s three-state model, daily commands, and GitHub CLI. learn to undo mistakes.

keep credentials out of git: use .env, os.getenv, and ignore rules, then practice responding to a leak.

call model APIs, construct messages, and read responses using the basic four-line pattern.

define output schemas, request JSON, and validate with Pydantic before malformed data reaches your application.

connect agents to tools and data through MCP. learn servers, tool discovery, and the shared protocol behind integrations.

build the agent cycle: request a model response, execute tool_use, return results, and repeat until end_turn.

an AI coding tool writes and edits with you; Cursor is one brand, Claude Code another. write prompts that clarify the task before work begins.

read tool calls, results, and stop reasons to locate failures in an agent’s execution trace.

same input, check the output, pass or fail. don't ship on a demo that felt right.

use chunking, embeddings, and vector search to ground model answers in your data. evaluate retrieval and avoid unnecessary infrastructure.

Train one consistent job with reviewed examples and a QLoRA adapter; compare held-out evals with prompt-only without claiming broader intelligence gains.

when the model lies to your customer. isolate what went wrong, classify the four common failures, and ship a verified fix instead of blaming the model.

balance cost, latency, and quality with token math, caching, streaming, batching, and model selection.

connect the prompt, call, validation, trace, eval, and MCP tool in the smallest end-to-end llm feature you could ship. on the developer route, harnesses, image, video, and production tooling follow. later retrieval and prompt-cache cost chapters extend this checkpoint beyond demo inputs.

claude code, cursor, aider, codex cli: learn the four layers around the model api, then build a small harness you can run and break. for actions that change the world, write a repo-based rollback runbook a junior can execute at 3am.

choose from the 2026 image lineup, write effective prompts, and turn one idea into a hundred ready-to-ship images. compare uses and costs: nano banana 2 for volume, nano banana pro for fidelity, flux, midjourney, ideogram, and gpt-image-2.

learn the 2026 video lineup, camera-control patterns, and cost math for this difficult, expensive, strategically useful medium. decide whether your idea is viable.

ai generates assets; code assembles them. compare and combine hyperframes, remotion, and claude design in data-driven workflows that turn one template into a hundred videos.

engineer the harness around a model: turn failures into rules, manage context, design long-running loops, enforce hooks, and choose what to build or buy as runtime SDKs evolve.

start here. turn weekly work into a reusable tool: learn the six-move builder loop, write a five-field brief, and save it for next week. finish by building an interactive webpage from a real sales deck and sending it to your team.

Run Fieldnote's fall relaunch from an evidence-backed content record through approved creative, an export-verified report, and a checked short-clip plan. Every rejection, result, and next decision keeps a receipt.

call model APIs, construct messages, and read responses using the basic four-line pattern.

define output schemas, request JSON, and validate with Pydantic before malformed data reaches your application.

an AI coding tool writes and edits with you; Cursor is one brand, Claude Code another. write prompts that clarify the task before work begins.

same input, check the output, pass or fail. don't ship a campaign on a demo that felt right.

Audit how approved buyer questions are answered and cited, then build one source-backed asset for a real evidence gap. Manual observations only, with dated receipts and no causal claims from one change.

choose from the 2026 image lineup, write effective prompts, and turn one idea into a hundred ready-to-ship images. compare uses and costs: nano banana 2 for volume, nano banana pro for fidelity, flux, midjourney, ideogram, and gpt-image-2.

learn the 2026 video lineup, camera-control patterns, and cost math for this difficult, expensive, strategically useful medium. decide whether your idea is viable.

ai generates assets; code assembles them. compare and combine hyperframes, remotion, and claude design in data-driven workflows that turn one template into a hundred videos.

start here. turn weekly work into a reusable tool: learn the six-move builder loop, write a five-field brief, and save it for next week. finish by building an interactive webpage from a real sales deck and sending it to your team.

read small ai-written scripts: follow values, check function returns, and trace if branches. if you are starting here, learn these reads first; if you already ran a studio, connect them to its Python.

read nested api, config, and llm data; walk lists and dicts with loops, and catch wrong-layer bugs.

read Python tracebacks to locate failures. catch errors you can handle without hiding unexpected bugs.

follow files and api calls through a script. learn with-blocks, status families, and bounded retries.

Run image models with creative briefs, structured brand kits, interface design tokens, seed and reference locks, and QA filters for fifty consistent assets.

call model APIs, construct messages, and read responses using the basic four-line pattern.

define output schemas, request JSON, and validate with Pydantic before malformed data reaches your application.

an AI coding tool writes and edits with you; Cursor is one brand, Claude Code another. write prompts that clarify the task before work begins.

same input, check the output, pass or fail. don't trust a demo that felt right.

choose from the 2026 image lineup, write effective prompts, and turn one idea into a hundred ready-to-ship images. compare uses and costs: nano banana 2 for volume, nano banana pro for fidelity, flux, midjourney, ideogram, and gpt-image-2.

learn the 2026 video lineup, camera-control patterns, and cost math for this difficult, expensive, strategically useful medium. decide whether your idea is viable.

remotion and hyperframes turn a kit and a data file into a batch of clips. one template, many products, one render loop.

start here. turn weekly work into a reusable tool: learn the six-move builder loop, write a five-field brief, and save it for next week. finish by building an interactive webpage from a real sales deck and sending it to your team.

read small ai-written scripts: follow values, check function returns, and trace if branches. if you are starting here, learn these reads first; if you already ran a studio, connect them to its Python.

read nested api, config, and llm data; walk lists and dicts with loops, and catch wrong-layer bugs.

read Python tracebacks to locate failures. catch errors you can handle without hiding unexpected bugs.

follow files and api calls through a script. learn with-blocks, status families, and bounded retries.

Check support drafts against policy and banned promises, evaluate macros per intent using fictional cases, and retain human approval and escalation for money, legal concerns and emotional requests. Passing fixtures never authorizes deployment.

call model APIs, construct messages, and read responses using the basic four-line pattern.

define output schemas, request JSON, and validate with Pydantic before malformed data reaches your application.

build the agent cycle: request a model response, execute tool_use, return results, and repeat until end_turn.

connect agents to tools and data through MCP. learn servers, tool discovery, and the shared protocol behind integrations.

an AI coding tool writes and edits with you; Cursor is one brand, Claude Code another. write prompts that clarify the task before work begins.

same input, check the output, pass or fail. don't send a reply on a demo that felt right.

when the model lies to your customer. isolate what went wrong, classify the four common failures, and ship a verified fix instead of blaming the model.

start here. turn weekly work into a reusable tool: learn the six-move builder loop, write a five-field brief, and save it for next week. finish by building an interactive webpage from a real sales deck and sending it to your team.

read small ai-written scripts: follow values, check function returns, and trace if branches. if you are starting here, learn these reads first; if you already ran a studio, connect them to its Python.

read nested api, config, and llm data; walk lists and dicts with loops, and catch wrong-layer bugs.

read Python tracebacks to locate failures. catch errors you can handle without hiding unexpected bugs.

follow files and api calls through a script. learn with-blocks, status families, and bounded retries.

Measure client voice from approved copy, check variants and channel budgets, log rejections, verify factual claims before shipping, and record corrections.

call model APIs, construct messages, and read responses using the basic four-line pattern.

define output schemas, request JSON, and validate with Pydantic before malformed data reaches your application.

an AI coding tool writes and edits with you; Cursor is one brand, Claude Code another. write prompts that clarify the task before work begins.

use chunking, embeddings, and vector search to ground model answers in your data. evaluate retrieval and avoid unnecessary infrastructure.

same input, check the output, pass or fail. don't publish a claim on a demo that felt right.

when the model ships a claim nobody sourced. isolate what went wrong, classify the four common failures, and ship a verified fix instead of blaming the model.

start here. turn weekly work into a reusable tool: learn the six-move builder loop, write a five-field brief, and save it for next week. finish by building an interactive webpage from a real sales deck and sending it to your team.

read small ai-written scripts: follow values, check function returns, and trace if branches. if you are starting here, learn these reads first; if you already ran a studio, connect them to its Python.

read nested api, config, and llm data; walk lists and dicts with loops, and catch wrong-layer bugs.

read Python tracebacks to locate failures. catch errors you can handle without hiding unexpected bugs.

trace shared lists and dicts to catch unintended mutation. learn when references, shallow copies, and deep copies preserve your data.

follow files and api calls through a script. learn with-blocks, status families, and bounded retries.

tables are the working surface of analysis. learn rows, columns, missing values, joins, and the dataframe habits a real export assumes.

most analysis tables start in a database. learn the select, join, filter, and quality checks that decide whether a memo is defensible.

a dataset is a product surface. build ingestion, validation, partitions, manifests, and checkpoints so the next run is not a mystery.

Account for dropped rows, verify aggregations on hand-checked samples, trace memo figures, challenge recommendations using the SR 11-7 triad, and require identical reruns.

call model APIs, construct messages, and read responses using the basic four-line pattern.

define output schemas, request JSON, and validate with Pydantic before malformed data reaches your application.

use chunking, embeddings, and vector search to ground model answers in your data. evaluate retrieval and avoid unnecessary infrastructure.

same input, check the output, pass or fail. don't cite a number on a demo that felt right.

start here. turn weekly work into a reusable tool: learn the six-move builder loop, write a five-field brief, and save it for next week. finish by building an interactive webpage from a real sales deck and sending it to your team.

read small ai-written scripts: follow values, check function returns, and trace if branches. if you are starting here, learn these reads first; if you already ran a studio, connect them to its Python.

read nested api, config, and llm data; walk lists and dicts with loops, and catch wrong-layer bugs.

read Python tracebacks to locate failures. catch errors you can handle without hiding unexpected bugs.

follow files and api calls through a script. learn with-blocks, status families, and bounded retries.

drive your computer through a text box: type a command and press enter. move around files, make folders, and read files without the mouse — the foundation for later tool chapters.

Define specs and acceptance checks before AI builds, verify results, derive status and go/no-go decisions from evidence, and record one accepted function’s value using comparable hours, finance’s loaded rate and a named owner.

call model APIs, construct messages, and read responses using the basic four-line pattern.

define output schemas, request JSON, and validate with Pydantic before malformed data reaches your application.

an AI coding tool writes and edits with you; Cursor is one brand, Claude Code another. write prompts that clarify the task before work begins.

read tool calls, results, and stop reasons to locate failures in an agent’s execution trace.

same input, check the output, pass or fail. don't sign off on a demo that felt right.

when the model lies to your customer. isolate what went wrong, classify the four common failures, and ship a verified fix instead of blaming the model.

balance cost, latency, and quality with token math, caching, streaming, batching, and model selection.

a claude skill packages your team's playbook for loading when relevant, saving repeated explanations. for team managers: define, share, and provision skills; explore hr, legal, and ops examples; choose skills over one-off prompts; and establish governance before real use.

start here. turn weekly work into a reusable tool: learn the six-move builder loop, write a five-field brief, and save it for next week. finish by building an interactive webpage from a real sales deck and sending it to your team.

read small ai-written scripts: follow values, check function returns, and trace if branches. if you are starting here, learn these reads first; if you already ran a studio, connect them to its Python.

read nested api, config, and llm data; walk lists and dicts with loops, and catch wrong-layer bugs.

read Python tracebacks to locate failures. catch errors you can handle without hiding unexpected bugs.

follow files and api calls through a script. learn with-blocks, status families, and bounded retries.

Practice employee-data boundaries, resume-linked screening briefs, named legal review for policies and synthetic decision-record checks. Local fixtures do not establish privacy, compliance or legal defensibility.

call model APIs, construct messages, and read responses using the basic four-line pattern.

define output schemas, request JSON, and validate with Pydantic before malformed data reaches your application.

an AI coding tool writes and edits with you; Cursor is one brand, Claude Code another. write prompts that clarify the task before work begins.

keep credentials out of git: use .env, os.getenv, and ignore rules, then practice responding to a leak.

a claude skill packages your team's playbook for loading when relevant, saving repeated explanations. for team managers: define, share, and provision skills; explore hr, legal, and ops examples; choose skills over one-off prompts; and establish governance before real use.

same input, check the output, pass or fail. don't send a people draft on a demo that felt right.

start here. turn weekly work into a reusable tool: learn the six-move builder loop, write a five-field brief, and save it for next week. finish by building an interactive webpage from a real sales deck and sending it to your team.

read small ai-written scripts: follow values, check function returns, and trace if branches. if you are starting here, learn these reads first; if you already ran a studio, connect them to its Python.

read nested api, config, and llm data; walk lists and dicts with loops, and catch wrong-layer bugs.

read Python tracebacks to locate failures. catch errors you can handle without hiding unexpected bugs.

trace shared lists and dicts to catch unintended mutation. learn when references, shallow copies, and deep copies preserve your data.

follow files and api calls through a script. learn with-blocks, status families, and bounded retries.

drive your computer through a text box: type a command and press enter. move around files, make folders, and read files without the mouse — the foundation for later tool chapters.

claude cli brings the chat model to your actual files. install it, sign in, and run your first command. watch ai read, plan, and change files, and learn when to use the cli instead of chat.

Capture versioned SOPs, catch invented steps, distinguish auto-run work from named approvals, surface exceptions and verify report figures. Practice literal identifier masking locally; fixture checks do not authorize disclosure.

call model APIs, construct messages, and read responses using the basic four-line pattern.

define output schemas, request JSON, and validate with Pydantic before malformed data reaches your application.

connect agents to tools and data through MCP. learn servers, tool discovery, and the shared protocol behind integrations.

build the agent cycle: request a model response, execute tool_use, return results, and repeat until end_turn.

an AI coding tool writes and edits with you; Cursor is one brand, Claude Code another. write prompts that clarify the task before work begins.

same input, check the output, pass or fail. don't trust a demo that felt right.

claude code, cursor, aider, codex cli: learn the four layers around the model api, then build a small harness you can run and break. for actions that change the world, write a repo-based rollback runbook a junior can execute at 3am.

start here. turn weekly work into a reusable tool: learn the six-move builder loop, write a five-field brief, and save it for next week. finish by building an interactive webpage from a real sales deck and sending it to your team.

read small ai-written scripts: follow values, check function returns, and trace if branches. if you are starting here, learn these reads first; if you already ran a studio, connect them to its Python.

read nested api, config, and llm data; walk lists and dicts with loops, and catch wrong-layer bugs.

read Python tracebacks to locate failures. catch errors you can handle without hiding unexpected bugs.

follow files and api calls through a script. learn with-blocks, status families, and bounded retries.

Verify every cited authority at its source, pinpoint contract claims, route sensitive documents under counsel-defined rules and record verification and sign-off.

call model APIs, construct messages, and read responses using the basic four-line pattern.

define output schemas, request JSON, and validate with Pydantic before malformed data reaches your application.

an AI coding tool writes and edits with you; Cursor is one brand, Claude Code another. write prompts that clarify the task before work begins.

use chunking, embeddings, and vector search to ground model answers in your data. evaluate retrieval and avoid unnecessary infrastructure.

same input, check the output, pass or fail. don't file a draft on a demo that felt right.

when the model ships a privilege call nobody cited. isolate what went wrong, classify the four common failures, and ship a verified fix instead of blaming the model.

keep credentials out of git: use .env, os.getenv, and ignore rules, then practice responding to a leak.

a claude skill packages your team's playbook for loading when relevant, saving repeated explanations. for team managers: define, share, and provision skills; explore hr, legal, and ops examples; choose skills over one-off prompts; and establish governance before real use.