A conference talk, a thesis defense, or a work presentation usually gets rehearsed the same way most people rehearse anything: read the slides out loud once or twice, roughly on time, and hope it holds together in front of the real audience. AI can tighten that loop considerably — but only for the parts of a rehearsal that are actually visible in a transcript: structure, pacing, clarity of the thesis in each section, and filler-word density. It cannot watch you deliver the talk, read a real room’s attention, or verify that the specific numbers and claims in your slides are correct.
This is a rehearsal workflow that uses AI for the transcript-visible half of practice and is explicit about the half that still needs a human room and your own verification.
Step 1: Write the outline and first draft yourself
Before any AI involvement, write your own outline: the thesis of the talk, the thesis of each section, and the one thing you want the audience to remember. This matters for the same reason covered in what deteriorates when you outsource thinking — if AI drafts the structure and content from scratch, you are rehearsing someone else’s argument in your own voice, which is noticeably weaker under real questions than an argument you actually built and can defend on the spot.
AI is useful after this point, for tightening what you have already written — not for generating the substance of what you know.
Step 2: Run it once, unaided, and time it honestly
Rehearse out loud, phone timer running, no notes beyond what you would actually have on stage. Do not stop to fix mistakes mid-run; let the run be honestly bad if it is bad. This is your baseline, and it is the only way to know your real pacing before any feedback loop starts adjusting it.
Step 3: Transcribe the run and build the rubric before reading it
Transcribe the recording (most phones and several free tools do this automatically) and, separately, write the rubric you want feedback against — before you look at the transcript, so the rubric reflects what actually matters to the talk rather than being shaped by what you notice in this specific run.
A rehearsal transcript is often more sensitive than the finished talk. Unpublished research, a thesis defense, an internal roadmap, a client name, or an embargoed result can all be sitting in a draft that has not been cleared for an audience yet, and pasting it into a consumer AI tool is a disclosure — some tools retain inputs or use them for training unless you are on a plan and setting that says otherwise. Before the transcript leaves your machine: check whether your institution, employer, or co-authors have a policy or an embargo covering the material, confirm what the tool does with your input, and strip what the structural feedback does not need — participant names, unpublished figures, client identifiers, and anything a collaborator has not agreed to share. If the talk covers work that is not solely yours, the decision to upload it is not solely yours either.
I am rehearsing a presentation. My rubric for the transcript:
1. Each section opens with a clear one-sentence thesis.
2. Transitions between sections are explicit, not assumed.
3. Filler words ("um," "so," "basically," "kind of") are minimal.
4. The opening line is a concrete hook, not a throat-clearing setup.
5. The close states a specific takeaway or action, not a generic summary.
Score my transcript against each criterion with a specific example from the text.
Do not rewrite my talk — point to the exact sentence or gap for each issue.
This mirrors the rubric-first discipline from deliberate practice with AI: the rubric is fixed before feedback, so the feedback stays specific rather than flattering whatever this particular run happened to produce. Fixing it in advance also guards against a measured tendency in these models: when a user reveals a preference about their own work, assistants tailor feedback to match it, producing more positive assessments of the same passage than they give without the preference stated (Sharma and colleagues, Anthropic, 2023). Since you cannot help having a stake in your own talk, the practical defense is to commit the criteria to writing before you can see which ones this run would fail. The underlying loop — a narrow target, an attempt, feedback against a standard, and another attempt — is the structure deliberate-practice research has long described for building performance skills (Ericsson and colleagues, Psychological Review, 1993).
Step 4: Check pacing against your actual time limit
My time limit is [X] minutes. This transcript ran [Y] minutes at my natural pace.
Given the word count and my actual limit, tell me:
- Which sections are disproportionately long relative to their importance in my outline.
- Where cutting would cost the least to the argument.
Do not rewrite the cuts for me — just identify where the disproportion is.
Cutting your own content, once you know where the disproportion is, keeps you as the person who knows what stays and what goes — a judgment call a model has no basis for making on your behalf, since it does not know which parts matter most to your actual audience or argument.
Step 5: Revise, re-rehearse, and compare
Make your own edits based on the specific gaps identified, rehearse again unaided, and transcribe again. Compare the two transcripts on the same rubric:
Here is my rubric from before. Here is my second transcript.
Score it the same way, and tell me specifically which of the five criteria improved, which didn't, and which is now the weakest one.
This is the retry step that makes the feedback loop worth running more than once — a single round of feedback without a re-attempt tells you what was wrong; a second scored attempt tells you whether you actually fixed it.
Rehearse the questions, not just the talk
A talk rarely ends when the slides do — the Q&A is where an unprepared presenter is most exposed. AI is useful for generating a set of plausible audience questions grounded in your own content, which you then answer unaided, out loud, before the real session:
Here is my presentation outline and key claims: [paste your outline].
Generate 8 questions an informed audience member might reasonably ask, based only on what is in this outline.
Do not answer them - I will answer out loud myself.
Answering these unaided, ideally recorded so you can hear yourself, surfaces the gaps in your own argument that a scripted read-through never reveals, and it is far less stressful to discover a weak answer in private rehearsal than live in front of the actual audience.
What this loop cannot check
A transcript-based rehearsal loop cannot assess eye contact, posture, gesture, vocal tone, pacing that depends on live audience reaction, or how you handle an unexpected question. Do at least one full run in front of an actual person — a colleague, a friend, a mentor — before the real presentation. Delivery quality that a transcript cannot capture is often the difference between a technically well-structured talk and one that actually lands with a room.
The reason is structural rather than motivational. A transcript records the words you said and nothing about how they landed, so every signal that comes from a listener — the confused expression at the third slide, the question that reveals your setup was ambiguous, the point where attention visibly drops — is absent from the artifact you are getting feedback on. A live run is not a more serious version of solo practice; it is the only step that generates that class of information at all.
Verify your facts before you present them
If your talk includes statistics, quotes, or specific claims, verify each one against a primary source before the actual presentation, following the same discipline in why AI sometimes gives confident wrong answers. A rehearsal loop that polishes the delivery of an inaccurate number does not fix the number — it just makes the wrong claim sound more confident when a real audience member fact-checks it afterward.
An illustrative rehearsal cycle
Here is what a full pass through the loop looks like. The numbers are an illustrative scenario, not a measured case. A researcher preparing a 12-minute conference talk runs the baseline unaided and transcribes it. The rubric pass shows that most sections open with a setup sentence rather than a thesis, and the timing check shows the talk running well over the limit. She rewrites the openings herself, cuts a sub-point the pacing check flagged as long relative to its importance, and re-rehearses. The second transcript scores better on the structural criteria — but filler words have not moved, because filler-word density under real time pressure is a delivery habit, not a scripting problem. That is the useful outcome of the loop: it separated the problems a rewrite can fix from the one that needed the live run with a colleague, rather than leaving all of them in one undifferentiated sense that the talk “needs more work.”
Your next rehearsal
Write your outline and first draft yourself. Rehearse once, unaided, and transcribe it. Build your rubric before reading the transcript, score it, revise, and re-rehearse. Book one live run with a real person before the actual presentation. The presentation rehearsal card tracks your rubric scores across rehearsals and has a place to log the live-run feedback the transcript loop cannot generate on its own.
