AI Features in Software: Four-Week Online Course

A four-week online course on putting LLMs inside a piece of software, safely, taught on the Doodle of Smarter’s web console: the sign-in page’s illustration, which a Smarter LLMClient, the Doodle Editor, chooses each day, and another, the Doodle Artist, draws. The whole feature is open source, in apps/doodle/, with its history in three commits that grow it from a calendar. There are twelve 30-minute recorded lectures and a 90-minute live session each week. Week 1 covers the feature and its AI parts, week 2 prompting from code and the contract with the model, week 3 judges, sanitizers and tests, and week 4 people, operations and shipping. See Course Formats and Grading for the format, the grading scale and the course policies, and AI Features in Software for an overview of the subject.

At a Glance

Students

Python developers who want to add LLM features to software they build.

Deployment

Each student runs Smarter from source (make init), with the example manifests loaded and a superuser login. Students change code and approve doodles in the Django admin, so a shared instance doesn’t suit this course. Week 1 includes time for setup.

Models

The models that the Doodle’s manifests name, through the instructor’s OpenAI key, and OpenAI’s moderation model.

Workload

3 hours of instruction and 4 to 6 hours of assignments each week.

Cost

Under $10 per student in model charges, capped with a Budget. Assignment 4 and Capstone B prompt the most.

Learning Objectives

By the end of the course, students can:

  1. Decide which parts of a feature a model should do and which code should do, and build the feature’s safe answer first.

  2. Declare a feature’s AI parts as Smarter LLMClients and Guardrails, and prompt them from Python.

  3. Design a model’s request and answer as a validated contract with a menu of choices, and check every answer, and every input to the prompt, in code.

  4. Evaluate a judge model’s accuracy on labeled examples, and decide whether a feature fails open or closed.

  5. Defend a feature against generated markup and a model’s mistakes with an allowlist sanitizer and tests in which the model is patched.

  6. Design the human review, audit trail, switches, disclosure and cost controls that a feature needs before its AI output reaches users.

Technical Prerequisites

  • Required: Python: functions, classes, exceptions and type hints, and JSON.

  • Required: Git, GitHub and pull requests, with a fork of smarter.

  • Required by the end of week 1: Docker, and Smarter installed from source with a superuser login, showing the sign-in page’s doodle. See the Smarter Instructor Guide.

  • Helpful: Pydantic models, Python’s unittest.mock, reading a Django app, YAML, and having called an LLM API.

  • Not needed: JavaScript or React, machine learning, or SVG. The course reads the Doodle’s React app and SVG, but students don’t write either.

Schedule

The files that the lectures refer to are listed in full under Smarter in the Source Code.

Week

Item

Content

1

Lecture 1

The newsroom. A feature as desks: the calendar, the editor, the platform’s checks, the tone judge, the artist, the sanitizer and a superuser. Which jobs suit a model (taste, judgment, drawing), which suit code (rules, limits, security), and which need a person. The blog post.

1

Lecture 2

The safe answer first. Phase 1 (commit 0df9280ff): the catalog and its schedules, regions from a time zone and nothing else, the planner, the models and the switches. What visitors see when every AI part fails.

1

Lecture 3

AI resources as configuration. The editor’s and artist’s manifests: model, temperature, system prompt, example prompts, staffOnly and output Guardrails. Why a prompt belongs in YAML, not in code: who changes it, how, and what comes with it. Smarter LLM Client.

1

Live session 1

Setup triage. The editor in the workbench, live, with its example prompts and a request the class writes together. Start Assignment 1.

2

Lecture 4

Prompting from code. prompting.py: single_use_api_key(), prompt_llmclient() and assistant_text(). Why the platform prompts its own chat api rather than a provider’s SDK: the journal, cost accounting and Guardrails. Extracted from verify_stackademy, not copied (ADR 033); phase 2 (commit d07b91f67). Then the third copy: when the Orchestrator prompted LLMClients in-process too, the doodle moved to its send_message(), and only verify_stackademy, which tests the chat api itself, kept the http path.

2

Lecture 5

A menu, not a blank page. editor_request() and the Pydantic contract, DoodleEditorRequest and DoodleEditorDecision (ADR 031). Always offering default, and asking for a reason. parse_answer() and what happens to an answer that doesn’t validate.

2

Lecture 6

Checks in code, on the way out and the way in. check_decision() enforcing what the prompt asks; has_blocked_word() and the language matching in nlp.py, and why an allowlist beats a blocklist for news. Phase 3’s trends (commit 2e6cb0806): region_trends(), is_a_person() and is_an_occasion(), fetched with safe_http.py and cached.

2

Live session 2

Review of Assignment 1. Writing a management command that prompts the editor, together, and reading its prompt in the journal. Start Assignment 2.

3

Lecture 7

A second model as judge. The llm_judge strategy and upbeat_occasion_input. tone_check() fails closed, and declined a Diwali trend; why that trade is right here and wrong elsewhere. Model quirks and the judge’s retry in model_quirks.py. Smarter Guardrail.

3

Lecture 8

Generated markup is untrusted. What an SVG can carry. sanitize.py: allowlisted elements and attributes, CSS limited to the drawing’s own classes, size caps. compose_svg() in artwork.py, <img>, and the artwork’s Content-Security-Policy. The chat widget’s same problem, in DOMPurify and Markdown.

3

Lecture 9

Testing a feature whose model is patched out. test_editor.py: a test for every refusal and failure. Saved drawings in tests/data/artist/ and saved Wikimedia responses, so no test needs the network or a model. Running the app’s tests in the container.

3

Live session 3

Review of Assignment 2. A red-team workshop: the class attacks the editor and the artist, and traces each attempt to the layer that stopped it. Capstone proposals discussed. Start Assignment 3.

4

Lecture 10

People and paper trails. Proposed drawings, Preview the week, Approve and Reject in admin.py; DoodleEditorRun; a person’s choice never replanned; enable_doodle_auto_approve off by default; and the credit line that tells visitors an AI chose or drew the doodle (commit f00ac5cac).

4

Lecture 11

Running it. Celery Beat every six hours, EDITOR_MAX_ATTEMPTS, a day asked once; caching fallbacks that keep the page whole while a drawing awaits approval; the api contract and the React app; and a platform without the manifests, where the calendar does the whole job.

4

Lecture 12

Cost and design review. Reading a feature’s cost per day from the journal, capping it with a Budget, and choosing cheaper models for judges. The blog post’s takeaways as a checklist for reviewing any AI feature, applied to a feature the students use every day.

4

Live session 4

Capstone presentations (about 6 minutes each, with 2 minutes of questions), and a course retrospective.

Assignments 1 to 3 are due before the next live session. Assignment 4 is due in the middle of week 4. The capstone proposal is due before Live session 3, and the capstone at the start of Live session 4.

Smarter in the Source Code

Every lecture, lab and assignment in this course rests on the files and commits below. Every reference is to the main branch of github.com/smarter-sh/smarter, unless another repository is named, so that students can open each one during class. Python paths are under smarter/smarter/. The case that runs through the course is the web console’s Doodle, the illustration on the sign-in page that two Smarter LLMClients choose and draw each day, described in How Two Smarter LLMClients Create the Sign-in Page’s Daily Doodle and in Doodle React Component.

The Feature Without the AI

Topic

What Smarter does

Where to look

The app’s map

The Doodle is an ordinary Django app. Its package docstring lists each module: regions, schedules, the catalog, the planner, the editor, the sanitizer, caching and the api.

apps/doodle/, __init__.py

The catalog

Nineteen occasions with artwork drawn and reviewed ahead of time, each a short YAML file and an SVG of its decorations. A date is fixed, the nth weekday of a month, or a country’s holiday by name, worked out by the holidays package. themes.py validates every file before it saves any.

data/themes/, README.md, schedule.py, themes.py

Regions

The browser sends its time zone and nothing else. Smarter maps it to one of 12 regions and the visitor’s local date. No IP address is looked up, and nothing about the visitor is stored.

regions.py

The planner

The calendar alone chooses each region’s doodle a week ahead: the due theme of the highest priority, or the default. A doodle that a person chose is never replanned. This is the safe answer that the AI parts improve on, and fall back to.

planner.py

The models

DoodleArtwork stores an SVG under its SHA-256. Doodle is one region’s doodle of one day, with its status, its source (calendar, default, human or LLM) and its reason. DoodleEditorRun keeps every prompt to the editor, its answer and its refusals.

models.py

The schedule

A Celery task plans the doodles every six hours from Celery Beat, then asks the editor about new days. manage.py plan_doodles --editor does the same by hand.

tasks.py, celerybeat.py, plan_doodles.py

AI Resources as Configuration

Topic

What Smarter does

Where to look

The editor

An LLMClient manifest: a model, a temperature of 0.2, a long system prompt that describes the request and the answer, example prompts that are real requests, and two output Guardrails. Its taste is configuration, changed by editing YAML, not by a release. Smarter LLM Client.

llmclient-smarter-doodle-editor.yaml (example-manifests)

The artist

A second LLMClient, at temperature 0.6, whose system prompt describes the base scene in coordinates, the palette, a limit on shapes, and what it must never draw.

llmclient-smarter-doodle-artist.yaml (example-manifests), base.svg

Output Guardrails

The editor’s answer passes OpenAI’s moderation model and a Guardrail that strips a stray code fence so the JSON parses. Smarter Guardrail.

json-code-fence-output.yaml, moderation-output.yaml (example-manifests)

Staff only

staffOnly: true hides an LLMClient from everyone but the platform’s staff, who can still open it in the workbench and try it by hand.

llmclient-smarter-doodle-editor.yaml (example-manifests)

Prompting from code

The platform prompts its own LLMClients through the chat api, as any client would. single_use_api_key() creates an api key and always deletes it, prompt_llmclient() sends one prompt in a new session and validates the result, and assistant_text() returns the answer. The journal and the account’s cost accounting see every prompt.

prompting.py

Extracted, not copied

verify_stackademy had the same logic first. When the Doodle needed it, it was moved to prompting.py and both use it. ADR 033.

verify_stackademy.py

Finding the resources

smarter_llmclient() returns the editor or the artist, or None when the example manifests aren’t loaded, in which case the calendar does the whole job.

editor.py

Trust Nothing the Model Says

Topic

What Smarter does

Where to look

A menu, not a blank page

editor_request() builds a JSON document that lists, for each region, exactly what the editor may choose: the catalog’s due themes, the region’s other holidays, and its trends. default is always a safe choice.

editor.py

The contract with the model

Pydantic models describe the request (DoodleEditorRequest) and the answer (DoodleEditorAnswer, DoodleEditorDecision), with a reason for every choice. parse_answer() refuses anything that doesn’t validate. ADR 031.

editor.py

Every choice checked again

check_decision() enforces in code what the system prompt asks: the choice must be on the region’s menu, of an allowed category, with a short caption and alt text, and a Wikipedia link or none. A refused choice is recorded, and the region keeps the calendar’s doodle.

editor.py

Blocked topics

has_blocked_word() refuses mourning, war, politics, accidents, crimes and the like, with the same language matching as the Plugins’ selectors: plurals, typos and WordNet synonyms, never a word inside another.

editor.py, nlp.py

Inputs filtered before the model

Trends come from Wikipedia’s most viewed articles. region_trends() drops people, dates and blocked words, and keeps only articles whose English description names an occasion, an allowlist. They are fetched with the Plugins’ hardened HTTP client and cached for 12 hours.

trends.py, safe_http.py

Judges, Sanitizers and People

Topic

What Smarter does

Where to look

A judge of tone

Word lists can’t tell sad news from a festival. upbeat_occasion_input is an llm_judge Guardrail that asks a small model whether an occasion is plainly cheerful.

upbeat-occasion-input.yaml (example-manifests), llm_judge_strategy.py

Failing closed

tone_check() declines an occasion when the judge says no, when it fails, and when the Guardrail doesn’t exist on the platform.

editor.py

Asking a Guardrail directly

evaluate() runs one Guardrail on one text and returns whether it triggered, without a chat. tone_check() calls it in-process, and the Guardrail api’s evaluate view calls it over HTTP.

guardrail/api/v1/views/views.py

Models that differ

A model that rejects a parameter, as one rejects the judge’s temperature, is retried once without it. The chat provider and the judge share one list of such quirks.

model_quirks.py, clients.py

Generated markup

The artist’s SVG is untrusted. The sanitizer keeps an allowlist of shapes and attributes, strips handlers, links, ids, url(), javascript: and data:, limits CSS to the drawing’s own prefixed classes, and caps its size.

sanitize.py, artwork.py

Two locks on one door

The page shows artwork with <img>, in which a browser runs no scripts, and the artwork’s url sends a sandboxing Content-Security-Policy and nosniff.

views.py

A person decides

A drawing is saved as proposed. In the Django admin, Preview the week shows every region’s doodle for every planned day, with who chose it and the editor’s reason, and Approve and Reject buttons. Doodle Editor Runs are the paper trail.

admin.py

Switches

The editor, the artwork, the trends and the credit each have a waffle switch. enable_doodle_auto_approve, which skips review, is off by default.

switches.py

Into the Product

Topic

What Smarter does

Where to look

The api

GET /api/v1/doodles/today/?tz= returns the day’s doodle as a Pydantic contract, with aiCredit and learnHowUrl. Caching falls back from the region’s approved doodle to the calendar’s choice, so a drawing that awaits approval never leaves a gap.

contract/models.py, caching.py

The page

The single-page console’s sign-in pages show the doodle, its caption, and “Chosen by” or “Drawn by a Smarter AI application” with a link that explains how; until release v0.42.10, a React app that Django hosted showed it, with a template tag. Django-React Integration.

smarter-doodle/src/App.tsx, AuthCard.tsx, react_doodle.py

Testing with the models patched

The editor’s and artist’s answers are patched in tests, with a test for every refusal and every failure. The app has well over a hundred tests.

test_editor.py

Dangerous input on file

The sanitizer is tested against saved drawings, one safe, one crowded, one empty and one dangerous; trends against saved Wikimedia responses, so no test needs the network.

tests/data/, test_sanitize.py, test_trends.py

The Doodle in the History

The Doodle was built in three phases, each a commit whose body explains what it adds, how it was tested and what it ran into, so students can watch an AI feature grow from a calendar.

Commit

Why it is worth studying

Where to look

Phase 1: a calendar

The whole feature with no LLM at all: regions, schedules, the catalog, the planner, the api, the React app and the switches. The safe answer comes first.

commit 0df9280ff

Phase 2: an editor and an artist

Two LLMClients, the prompting helpers extracted from verify_stackademy, the checks, the sanitizer, the editor runs and the admin’s Approve and Reject. Its body records what the models did when they were run for real.

commit d07b91f67

Phase 3: trends and a judge

Wikipedia trends, the tone check that fails closed, and a fix to every llm_judge Guardrail on a model that rejected its temperature.

commit 2e6cb0806

Credit where it’s due

The “Learn how” link under an AI-chosen doodle points to the article that explains it.

commit f00ac5cac

Graded Assignments

Assignment 1: The Editor’s Taste

Students apply a copy of the Doodle Editor’s manifest with an editorial policy of their own, such as science and education only, and run the same five requests through both editors, including two that tempt the model off the menu. They compare the choices and reasons, and say which of them check_decision() would refuse.

Turn in: the manifest, and a three-page Markdown report with each request, both answers and the analysis.

Criterion

Pts

Exemplary

Developing

A working manifest

4

Applies cleanly; the policy is stated in the system prompt, not scattered.

Edits that contradict the rest of the prompt.

Evidence

6

Real answers from both editors, pasted as they came.

Paraphrased, or one editor only.

Analysis

6

Explains each difference by a line of the prompts, and reads the reasons.

Lists differences without explaining them.

The off-menu attempts

4

Shows what the model did, and names the check that would catch it.

Assumes the prompt is enough.

Assignment 2: Prompting From Code

Students write a management command in their fork that prompts their copy of the editor for a day, with editor_request() and the helpers in prompting.py, and prints each decision with the verdict of check_decision(). The api key it uses must be deleted even when the prompt fails, and the command must be tested with the model patched.

Turn in: a pull request against their own fork, the command’s output for two days, and its test output from the container.

Criterion

Pts

Exemplary

Developing

Uses the platform

5

Prompts through prompting.py; the prompt shows in the journal.

Calls a provider’s SDK directly.

Parses and checks

5

parse_answer() and check_decision() on every decision, refusals printed with why.

Prints the raw answer.

Cleans up

4

The api key is deleted on success and failure, and a test proves it.

A key left behind when the prompt fails.

Tested

6

Tests with the model patched, in the style of test_commands.py.

Untested, or tested by prompting a model.

Assignment 3: Trust Nothing

Students write at least eight tests that feed crafted answers through parse_answer() and check_decision(), and crafted trends through region_trends(), with the model and the network patched. They look for one input that passes but shouldn’t, and propose a fix.

Turn in: a pull request against their own fork with the tests, their output from the container, and a paragraph on the gap they found, or why they found none.

Criterion

Pts

Exemplary

Developing

Cases covered

7

Answers and trends, eight or more distinct ways to fail, each asserting the reason given.

Only malformed JSON and off-menu.

Isolation

4

No real prompts or requests; saved responses, as in test_trends.py.

Tests that call the LLM or Wikipedia.

The gap

6

A real gap with a failing test and a fix, or a convincing search that found none.

“The checks look fine.”

Runs as shown

3

Output from the container matches the pull request.

Took the tests’ passing on trust; output not included.

Assignment 4: A Judge, Measured

Students write an llm_judge Guardrail for a criterion that code can’t check, label twenty inputs by hand, and run each through the Guardrail with evaluate(), as tone_check() does. They report its false positives and false negatives, try one change to its prompt or threshold, and say whether a feature that uses it should fail open or closed.

Turn in: the Guardrail’s manifest, the labeled inputs and results as a table, and a one-page analysis.

Criterion

Pts

Exemplary

Developing

A fair test set

5

Labeled before running, with hard cases on both sides of the line.

Twenty easy cases, or labels chosen after the results.

Results

5

Every input’s verdict and score, recorded as they came.

A summary without the data.

The change

4

One change, measured on the same set, kept or reverted on the evidence.

Several changes at once, or none measured.

Fail open or closed

6

A decision argued from what each kind of error costs the feature.

“Closed, to be safe,” with no costs weighed.

Capstone Projects

Students choose one, propose it before Live session 3, and present it in Live session 4.

Capstone A: An AI Feature of Your Own

Learning objective. Build an AI feature into software, with its AI parts as Smarter resources, its safety in code, and a person in the loop.

In their fork of Smarter, or in a Django project of their own that prompts a Smarter platform, students build a feature in which an LLMClient makes a choice or writes something that users see: a tip of the day picked from the documentation, a suggested description for a new resource, or a summary of an account’s week. It must work without the model, give the model a menu or a schema, check every answer in code, judge what code can’t check with a Guardrail, record what was asked and answered, let a person approve or override what is new, have a switch, and say on the page when an AI made it. A feature meant for Smarter itself must be agreed with the maintainers in a GitHub issue first.

What students turn in:

  • A pull request against their own fork, with the LLMClient’s and Guardrail’s manifests.

  • Tests with the model patched, and their output, with at least 90% coverage of the feature’s code.

  • A two-page design note: the feature’s desks, what each check guards against, whether each one fails open or closed and why, and the cost per day from the journal.

  • A six-minute presentation, including one answer the code refuses and one a person rejects.

How it is graded (100 points):

Criterion

Pts

Exemplary

Developing

A safe answer first

15

The feature works, and says nothing wrong, with the model off.

Breaks or shows an error without the model.

Contract and checks

20

A validated request and answer, and every rule of the prompt enforced in code.

Trusts the prompt.

Judge and review

20

A judge for what code can’t check, and a person approves what is new.

The model’s output goes straight to users.

Record, control, disclosure

15

What was asked and answered is kept; a switch turns the AI off; the page credits it.

No record, no switch, or no disclosure.

Tests

20

Refusals and failures tested, with the model patched, at 90% coverage or more.

Only the happy path, or tests that prompt a model.

Design note

10

Each check tied to the failure it prevents and its cost.

A description of the code.

Capstone B: Red Team the Newsroom

Learning objective. Evaluate an AI feature’s defenses with evidence, layer by layer, and strengthen the weakest.

On their own platform, students attack the Doodle with at least forty cases: requests that tempt the editor off the menu, trend titles written to inject instructions, occasions that are sad but contain no blocked word, and briefs that ask the artist for scripts, links, external images or a drawing over the base scene. For each case they record which layer stopped it: the prompt, an output Guardrail, check_decision(), the trends filter, the judge, the sanitizer, or the person in the admin. They fix the two weakest points they find, each with a test that fails before the fix.

What students turn in:

  • A table of the cases, with the input, what each model answered, and the layer that stopped it, or none.

  • A four-page analysis of where the defenses overlap and where they are thin, citing the code and the commits.

  • A pull request against their own fork with the two fixes and their tests.

How it is graded (100 points):

Criterion

Pts

Exemplary

Developing

Range of attacks

20

Every layer is tested, with cases a careful attacker would try.

Forty variants of one attack.

Evidence

20

Real answers and editor runs, recorded as they came.

Results from memory, or assumed from the code.

Attribution to layers

20

Each case traced to the code or resource that stopped it.

“It was blocked.”

The fixes

25

Failing tests, then fixes that pass without breaking the app’s other tests.

Suggestions without tests.

Writing

15

Concise, organized and specific.

Long or general.

Further Study