AI Tutor for Homework Help: Doing the Work Yourself, With a Check Built In

Homework is where a lot of the actual learning in a STEM course happens,
and also where a lot of students quietly stop learning and start producing.
The line between those two is not visible from the outside. A well-formed
answer shows up on the page either way. The difference is in the twenty
minutes before the answer got written down.

This page is about how to use AI on homework so that the twenty minutes
stays with you. It is written for the student who has the problem set in
front of them, a deadline that is not today, and a tool that can help or
that can quietly take the work over. [You can run your actual problem set
through the loop described here starting free, no card, on your own problem set.]

The shape of a homework problem, and where the work lives

Most homework problems in the courses I see broken up have three distinct
components, and each one gets done differently:

1. Setting up the problem. Reading the problem, identifying what is
being asked, naming the quantities, deciding which of the things you
were taught this week is the right thing to apply. This is the work that
separates a student who is passing the course from one who is just
completing it. Most of the actual learning happens here, and most of
it goes unpracticed when a tool short-circuits this step.

2. Carrying out the computation. Algebra, integration,
differentiation, the arithmetic of a statistical test. This is the
part that is the most mechanical, the part where a calculator or solver
genuinely helps, and the part where the least unique skill is
exercised. Most of the frustration students feel when they "don't
know how to do the problem" is actually in step one, but the
frustration shows up in step two.

3. Interpreting the answer. Did I pick the right sign? Does this
answer make sense in the context of the scenario? What is the
question really asking me to communicate about the answer? Step three
is where the answer becomes a response instead of a numerical value,
and it is the step most students skip entirely.

When a tool short-circuits step one, the student has practiced step
two and skipped step three. That is what a homework session looks
like from the outside when the work is being done by the tool.

What a good homework-adjacent tool actually does

A tool like Fennie is not a homework solver. It is structured around
asking you what you think the setup is, before it offers any help
further along. That is a small distinction and it matters in practice,
because it means that the tool's first move in a session is a
question to you, not an answer to the problem.

In practice that looks like a workflow. You read a problem, you write
down your setup (even if it is wrong), and you show it to the tool.
The tool tells you where the setup breaks, or confirms it. Then you
do the computation. Then you state what the answer is doing in the
context of the problem, and the tool checks that too.

The point is that every step in that sequence, you generated. The
tool is not running the sequence. That is the difference between
using a tool as a check and using a tool as a shortcut.

The specific failure patterns I see, and how to break each one

"I can't tell if my answer is right." This is the classic one.
Student finishes a problem, answer is written, and now what? The
honest answer is that most students have not been taught how to
self-verify, and verification is a distinct skill from production.
What you can do is ask the tool to walk back through your setup
and identify the weakest assumption in it. If it points at
something you had been unsure about, that is a real signal. If
it points at something you are sure about, the tool is wrong,
and that is worth a second pass.

"I did the problem, it feels right, I'm going to move on."
This is where step three gets skipped. The answer is there, the
algebra worked, and the student moves to the next problem. The
problem is that "feels right" is doing a lot of work in that
statement, and students are surprisingly bad at detecting when
their answer is right for the wrong reason. The cheapest fix is
to ask the tool to identify one assumption in your solution that,
if it were wrong, would change the answer. You do not have to
agree with it, but that conversation is where step three happens.

"I'm stuck on the first sentence of the problem." This
is the step-one block, and it is the most common type of
block students report. The tool should not give you the
setup. What you do is take the problem in pieces. What is
being asked? What am I given? What in this week's material
connects the two? Write down what you can, even in fragments,
and show it to the tool. The fragments become a full setup
after two or three rounds.

"I've never actually done this type of problem before."
A different case. This is not a block on a known problem type,
it is an unfamiliar one. The tool can point you to the
underlying concept and ask you to state it in your own words,
and then walk the connection to this problem. This is the
legitimate tutoring case, and it is the case most people
think "AI homework help" means, because it is the case
that actually requires help.

A complete session, with the work distributed correctly

A concrete version of what a session with one of these problems
looks like when the work stays with the student:

The problem. (Physics, intermediate) A block of mass m
slides down a frictional incline, then collides with a
stationary block, and the two stick together and roll to a
stop on a flat section. Find the distance the combined
mass travels after the collision, given the coefficient
of kinetic friction on the flat section.

Step one, the student's work. Read the problem. Identify
phases: (a) incline, (b) collision, (c) flat section.
Identify what is given: mass, incline, friction
coefficients. Identify what is asked: final
distance. Identify the physics tools needed:
energy, momentum (for collision),
friction. Write the connection between
phases in one or two sentences. This
is the setup, and this is where
most of the thinking happens.

Step two, the check. Show the setup to the
tool. The tool confirms the phases are
right, flags whether the collision
treatment (elastic or inelastic) is
handled correctly, and raises the
question of whether the incline phase
actually loses energy to friction.
The student revises the setup
if the tool's flag is legitimate.

Step three, the computation. The
student carries out the algebra with
paper or a calculator. This is the
step a generic calculator
is good at, and it is
the step where students
should not be
relying on an AI
tool to do the
arithmetic.

Step four,
the
interpretation.
The student states
what the answer is
actually doing:
what does a large
answer mean in
the context
of the
scenario?
Does a
zero
answer
make
sense?
The tool
checks
the
interpretation
against
the
setup.

At no point in that session did the tool do the work. It asked for
the setup, and it flagged where the setup was weak, and it checked
the interpretation at the end. That is the shape of a tool that is
helpful to homework instead of doing it.

The line between helpful and not helpful, stated plainly

The line is not "can the tool solve the problem?" Many can. The line
is "does the tool put the student through the same sequence they
would go through when actually learning the material?" The
sequence is setup, computation, interpretation. If the tool
collapses that sequence, it is a shortcut. If it walks through
the sequence with you, it is a check.

Fennie is on the check side of that line. That is what
its "integrity by design" framing is pointing at: the
tool asks you what you are thinking before it gives
you anything, and the loop is setup-then-check-then-interpret,
not answer-then-forget.

Practical rules that work for most people

1. Never start with the problem. Start with a sentence
   describing what the problem is doing. If you can't write
   the sentence, that is the actual place you are stuck,
   and the tool can help there specifically.

2. Do the setup on paper before typing it in. A
   physical commitment to the setup catches more errors
   than typing one into a chat box, because the friction
   of writing it out surfaces what you actually know
   versus what you think you know.

3. Ask the tool to identify the weakest assumption in
   your solution, not the right answer. This single
   reframe changes what the tool can do for you, and it
   is the difference between using it as a check and
   using it as a shortcut.

4. Do step three (interpretation) in words, not
   in a symbol. If your interpretation is a formula,
   you are still in step two. Step three is a
   sentence.

The same problem, three different students

The same homework problem, worked by three different students with the
same tool, gives three different learning outcomes. The problem is
identical in all three cases, and the tool is the same in all
three cases. The difference is where each student stops asking
themselves "do I know this?" before asking the tool.

Student A reads the problem, and does not write down any setup
before talking to the tool. They ask the tool how to start, and
the tool walks them through the setup, step by step. They get the
answer. The tool did step one, and the student did step two,
which is the calculation. This is the shortcut pattern, and it
is the one that looks like learning from the student's point of
view, because the session felt like a lesson, and the answer
came out right.

Student B reads the problem, writes down what they think the
three phases are, and shows it to the tool before doing any
calculation. The tool points at the weakest assumption, and
Student B walks that assumption back. The calculation runs
after the setup is checked, and the interpretation runs after
the calculation. The tool did not do any of the three steps,
and the student did all three. This is the loop, and it is
the one that builds the skill.

Student C reads the problem, does the calculation without a
setup, and then asks the tool if their answer looks right.
The tool confirms it does, and Student C moves on. The student
skipped step one entirely, and the tool was not in the
position to catch it, because the setup was not in the
session. This is the third pattern, and it is the one where
the student has practiced neither step one nor step three,
and they have just practiced step two with a slightly
faster pen.

Same problem. Same tool. Three different learning outcomes,
and the difference is entirely in where the student stopped
and started.

The FAQ that actually comes up

Q: Can I just let the tool do the problem and check the
final answer?
Yes, and that is the shortcut. It works for producing
an answer, and it does not work for building the skill,
which is why the grade eventually catches up with the
approach. Most instructors can tell, because the
distribution of errors in a "tool-produced" solution
looks different than in a "thought-through" one.

Q: What's the difference between homework help
and a tutor?
The homework help framing implies the tool is doing the
homework for you, and that is a category most tools fall
in, because that is what they are good at and what they
are optimized to do. The tutor framing is the tool
asking you questions and checking your work. The
second one is what this page is about.

Q: When is it fine to just use a calculator or a
solver?
For the arithmetic in step two, fine. For the setup in
step one, no. That is the whole line.

Q: Do I need Fennie specifically, or would any
tool work?
Any tool can be put through the loop above. The
question is whether the tool is structured to ask
you for your setup before it gives you more. If yes,
the loop works with it. If the tool just answers,
the loop collapses into the shortcut.

Bottom line

Homework is the highest-leverage place an AI tool can
genuinely help, and also the highest-risk place one can
quietly take the work away. The line between them is
not visible in the final answer, which is why the
student has to be the one protecting it. The
concrete protection is the loop: setup, check,
computation, interpretation, with your name on
each step.

That is what the tool described here is built for, and
it is what [the free, no-card loop] is for. The
14-day Premium trial is a reasonable
time to see whether the loop works for your
specific course, and canceling it is two
clicks in settings if it does not.

Start on the free tier, no card required, run one section of your course and see whether the loop holds for your material before you commit to a second subscription. Get started free