Home / Workshop, capstone and exam
The graduation checklist, the final exam
A written self assessment reveals gaps that a simple reread does not show, and both model choice and agentic loops with checkpoints mark the line between practitioner and beginner.
Rereading your own end of course checklist and telling yourself yes, I know how to do that, proves nothing: a reread almost always confirms what you already believe you know, because you recognise each line without needing to redo it. The version that reveals a real gap is the one that asks you to cite the precise action that proves each point, not to recognise it in a list.
A checklist that tests the action, not the memory of having read it
Go through your checklist alone, before the exam, and for each point ask the same question: what concrete action, what command, what role file, what prompt sent, proves you have already done this yourself. If the answer is I remember reading it in a lesson, do not tick the box. If the answer names a precise action you could redo right now, tick it.
point : choisir un modèle adapté au chantier
geste qui le prouve : le 2026-08-30, j'ai comparé Sonnet 5 et Fable 5.1
sur le même prompt de revue de code, et retenu Sonnet 5 pour son coût
statut : coché, geste citéThis requirement turns a checklist for rereading into a checklist of proof. The format hardly matters, a table, a list, a text file; what matters is that every ticked box carries a dated action beside it, and every box with no action stays empty rather than being ticked from memory.
Two markers that sort the practitioner from the beginner
The first marker is the choice of model. A practitioner who has done the work can name the exact identifier of the model used on their last piece of work and the reason for that choice, speed, cost or quality of reasoning, rather than only remembering having used Claude. They also know that this identifier, like the size of the available context window, varies from one model to another and must be checked against the documentation on the day, never from memory, because the range changes several times a year. The lesson choosing a model, switching it, absorbing a refusal details this action.
The second marker is the agentic loop with checkpoints: the ability to place an explicit stop before any action that would change something real, a file, a database, a submission, rather than letting an agent chain through without an intermediate check. A beginner starts a loop and hopes the final result will be good. A practitioner knows exactly where they placed their stopping point in that loop, what they check at that precise spot, and what they do if the check fails there.
These two markers are not reread, they are cited. That is the only difference that matters on exam day.
Reread from memory versus proven by an action
| What you think you have mastered versus what you can prove you have mastered | Choosing a model for a given project | Agentic loop with a checkpoint | Rereading an output before believing it |
|---|---|---|---|
| Reread and I remember | I recall having compared two models | I know a stopping point is possible | I reread the replies that seem odd to me |
| Action I can cite right now | I can name the prompt and the criterion that tipped my choice last time | I can cite the exact stopping point where I validated the plan before execution | I can cite the last time a reread changed my decision, and what it changed |
Before the final exam, a practitioner rereads their graduation checklist, ticks each point that feels familiar, then pauses on the item about model choice and searches their own project history for the exact name of the model used last time.
Write in one sentence what this situation establishes, and in one sentence what it does not establish.
What this establishes: This starting point establishes that the practitioner identified the checklist item that requires citing a precise action rather than a simple memory of having read it.
What this does not establish: It does not establish that the practitioner will actually find the exact name of the model in their history, nor that they will therefore be able to tick this checklist point.
The three most common miscalibrations
- Too broad This situation establishes that the practitioner will pass the final exam, since they have already started checking the points on their checklist.
- Too narrow This situation establishes nothing, since rereading a checklist has no bearing on passing an exam.
- Beside the point This situation establishes that the model used on the practitioner's last project was well suited to the task performed.
- Rereading a checklist almost always confirms what you already believe you know, only the obligation to cite an action reveals a real gap.
- A practitioner names the exact identifier of the model used on their last project and the reason for that choice, a beginner only remembers having used Claude.
- The size of the context window varies from one model to another and must be checked against the documentation on the day, never from memory.
- A well built agentic loop places an explicit stopping point before any action that would change something real.
- A box ticked with no action to cite is worth less than a box left empty, because it hides a gap instead of showing it.
Go through your own graduation checklist unaided, and for each point, note the precise action, the command, the file or the prompt, that proves you have already done it; if you cannot cite anything, leave the box empty rather than ticking it from memory.
Every datable claim in this lesson links here to the public text behind it. A source that does not open proves nothing.
- Anthropic, Claude model comparison, identifiers and status consultée le 2026-09-02