Walkthrough: 12 minutes, then questions

0:00–1:30 intro. 1:30–7:30 demo story with the six boxes. 7:30–11:30 pack at steering altitude. 11:30–12:00 ask and 90-day tease. Then you stop and they ask. If you do not stop, there is no Q&A, and Q&A is where they decide. The walkthrough is a menu. The questions are the meal. Protect the last forty minutes of a real loop for their questions by not overflowing your twelve.

Honor the clocks with a timer in rehearsal. In the real loop you may not show a timer, but your body will know 90 seconds, six minutes, four minutes. If Run 1 is 16 minutes, you did not have a walkthrough. You had a tour. Cut 20 percent in this order: adjectives, vendor names, autobiography, extra boxes, platform vision, the second example. Never cut: user, refuse, eval bar, kill, honesty, ask.

Draw the six boxes during the demo window, talking while you draw. Do not draw in silence for two minutes. Do not skip the draw because you have a slide. The draw is how an FDE and a solutions interviewer believe you. A slide of six boxes is a prop. A hand drawing boxes in order is a person who owns the object. If you are on a video loop, still draw on paper and hold it up, or draw in a shared doc. The artifact is the motion, not the graphic.

Switch altitude at 7:30 out loud: "That is how the slice works. Here is how I would run it as a workstream." Then owners, bars, cost, top risk, ask. If you do not announce the switch, the room thinks you are still in architecture and they will ask architecture questions into the pack window. Announce. You already do this in steering when you change slides. Do it with your voice.

The ask at 11:30 in a rehearsal is a stand-in: "I need a knowledge owner and access." In a real loop the ask might be "happy to send the one-table pack" or "I would start week one on ingest and the forty questions." Still an ask. Then the tease: "I have a 30/60/90: first 30 is listen, golden set, one named owner." Then stop. Stopping is a skill. Count to three in your head. Let them pick up.

If they interrupt at minute three, let them. Pivot to their question, then ask: "Do you want me to finish the walk or stay on this?" That question is leadership. Finishing the speech over their interrupt is not. You practiced interrupts on day 26. Today you will still hate them. Hate them and yield. Yielding is how senior people look. Over-talking is how nervous people look.

Rehearse the switch sentence until it is automatic: "That is how the slice works. Here is how I would run it as a workstream." Without that sentence the twelve minutes is one blob, and blobs overflow. With it, the panel can follow you the way a steering committee follows a tab change. You are not giving a keynote. You are running a meeting that happens to be about you. Meetings have sections. Name them.

1) Why not a bigger model? 2) How do you know it works? 3) What about confidential data? 4) What would you build in week one with an engineer? 5) Tell me a time you killed something. 6) How do you handle a VP who wants the CEO chatbot? 7) What's your Python actually like? 8) Why this seat and not PM / not engineering? Write 4–8 sentences each, in your voice, then speak the hardest four. Do not write essays. Essays become read-alouds. Answers are spoken.

1) Why not a bigger model? "Because the failure is retrieval, stale corpus, or no refuse path, not parameter count. I would measure recall@5 and groundedness on the owner's forty. If those miss, I fix chunking, the floor, and the corpus — not the logo. A bigger model on an empty retrieve still invents. Cost at 10× also gets worse. Model last." If they push a vendor name, you can say you would run the golden set on the already-paid tool first. That is enough.

2) How do you know it works? "Golden set written by the owner, not by me. Recall@5 on retrieve. Groundedness on the answer — citation or refuse. Bar I would not ship under: 0.7 and 0.9 as proposed numbers, not yet measured. Thumbs-up is a tile, not the bar. If I have not run it, week one is the table, not the UI." Labeled hole, if true. Do not invent a 0.82 you did not measure.

3) Confidential data? "Class first. Atlas is internal in this design; HR and credentials are out. Copies: index, logs, vendor. Path: lightest approved enterprise endpoint with a DPA, no consumer tools. ACL on the chunk. If Security says no, sanitized tenant with a death date, or we kill. I will not paste SOP text into a personal account." Then stop. Do not give a law seminar.

4) Week one with an engineer? "Ingest the named space. Forty questions running as a table. I write eval rows and chunking rules. They write pipeline and auth. Week two is UI and ACLs. I can read the script and change constants. I do not write the auth layer." This is the closer from day 23. Do not change it under adrenaline. Consistency is credibility.

5) A time you killed something. Use the messy-program story's hard call, compressed. A real kill from your past, not Atlas. What you wanted, what it cost, what stayed broken. Then one line: "I would use the same muscle on a use case that misses evals." If you have never killed work, do not invent. Say the smallest no you actually owned (a go-live delay, a scope cut) and that a portfolio kill is a gap you are not bluffing. Honesty here is better than a hero story.

6) VP wants the CEO chatbot. Drill A. Leadership no. One owned corpus, forty questions, two-week shadow, retrieve-only if they need a screen. Knowledge owner and access, not a bigger model. Short. You authored this yesterday. Do not rewrite it into a speech today. 7) Python: the eight-sentence stance. Read, change constants, no auth layer, no LeetCode claim. 8) Why this seat: day 22 evidence. Last three years, what you are not now, the default pair if relevant. Do not insult the other seats. They may be on the panel.

Sound like a delivery lead. Short sentences. Numbers. Names of owners. "I don't know" plus the next experiment. No "leverage synergies." You already know how you sound when a program is on fire. That voice is the hireable one. The interview voice people invent is slower, fancier, and less true. Do not invent it. Steal from your last SEV or your last steering fight.

Short sentences are not stupidity. They are how a room under time can follow you. "Users are the 40 Atlas PMs. Corpus is one wiki space. v1 refuses when retrieve is empty. Eval is forty owner-written questions. I need Maya Chen named this week." Five sentences. Five things the room can attack. A compound paragraph with three caveats gives them nothing to hold. You are not writing a paper. You are running a meeting.

Numbers: frequency, headcount, bars, dollars, days. If you do not have a number, say the method and the date you would have it. "I do not have 10× cost yet; week one I would compute tokens in and out on the forty and multiply." That is a number-shaped sentence. "Cost will be reasonable" is not. Owners: even in a design, name the role (knowledge owner, Security, you, the engineer). "We" without a name is how work dies. You have preached this. Preach it about yourself in the loop.

"I don't know" is a pass when it is followed by the next experiment and the partner. "I don't know how their ACL maps into the index. I would sit with the wiki admin in week one and if we cannot map it, we do not ship." That is hireable. "I don't know" alone is a shrug. "I don't know" plus a lecture on how you would theoretically research is a blog. Experiment and partner. Then stop.

Cut filler on the recording. "Sort of," "I think," "kind of," "at the end of the day," "going forward," "unlock value," "journey," "space," "lean in" as a verb. You will hate hearing yourself. Good. Cut until it sounds like a status call. Status-call you is the person they can put in front of a customer. Keynote-you is the person they put in a brand video and then ignore. You want the first job.

If English is not your first language, or if you are soft-spoken, do not compensate with jargon. Compensate with structure: spine, clocks, first lines of the eight. Structure reads as confidence in every accent. Jargon reads as covering. You already have structure. Use it. If you speak too fast under adrenaline, plant your feet and breathe at the switch (7:30). The switch is a built-in breath. Take it.

Numbers without owners sound like a dashboard nobody runs. "Recall@5 of 0.7" is a bar. "Recall@5 of 0.7, knowledge owner writes the items, I run the table every Friday, we freeze prompt changes if we miss two weeks" is an operating system. Prefer the second in Q&A. The first is a claim. The second is a job. You are applying to do a job.

Run 1 is for truth. Stand up. Full twelve minutes. Timer. Notes only for where you ramble, not for rewriting the design. You will hate Run 1. Everyone hates Run 1. The point is the ramble map: minute 2 autobiography, minute 5 vendor tour, minute 9 second example, minute 11 no ask. Circle those. That map is the cut list.

Cut 20 percent. If Run 1 was 15 minutes, 20 percent is three minutes. Remove whole beats, not adjectives only — adjectives will not give you three minutes. The second example goes. The origin story goes. The platform sentence goes. The joke goes. What remains should still include user, refuse, eval, kill, honesty, ask. If a cut would remove one of those, cut somewhere else. Those six are the scoring.

Run 2 is for muscle. Same standing, same timer, cut version. If Run 2 is still long, cut once more, but do not do a Run 3 today. Two runs plus eight answers is the session. Three runs is how you lose the eight answers, and the eight answers are half the loop. A slightly long Run 2 with eight answers is better than a perfect 12:00 with no Q&A prep.

After Run 2, write the script as artifact #5: the outline plus the sentences you actually said that worked, not the sentences you wished you said. Include the honesty sentence, the switch sentence, the ask, and the three keeper lines from day 27. This is the speaking copy. Put it in the pack. Tick the walkthrough artifact. If you only have the outline, you do not have artifact 5. The script is what you said, cleaned, not what this academy wrote.

If you have a human, Run 2 is the one you do to them. Ask them to interrupt once and to ask two of the eight. That is closer to a loop than a perfect monologue. If they say you sounded unlike yourself, believe them and cut the fancy. If they say they could not tell the seat, your intro is still wrong — fix the intro, not the six boxes.

Do not film a polished video unless a loop asked for one. Polished video is a different artifact and a time thief. Today is live. Live is messy. Live is the job. A camera adds a performance layer you do not need unless the employer uses async video (some do; then one take, sitting, same script, no extra slides). Default is standing in a room saying it.

After minute 12, they own the floor. Your job is to answer the question they asked, in the altitude they are in, with the same identity you used in the walkthrough. If they ask a technical question, do not flee into delivery slogans. If they ask a political question, do not flee into six boxes. Match. Then end with a number or an owner or an experiment. Then stop. Stopping after an answer is how you look senior. Adding a second answer they did not ask is how you look scared.

If you do not understand the question, say so in one line: "I want to answer the thing you care about — is this about evals or about the legal path?" Clarifying is allowed. Pretending you understood and answering a nearby question is how panels write "slippery" in their notes. You are not slippery. You are a delivery lead. Ask the clarifying question the way you would in steering.

If three people ask at once, pick one, name them, and come back: "I will take the eval question first, then the Python one." That is running a room. If you try to fuse three answers, you will fuse badly. Panels are a small steering committee. You know how to chair those. Chair this one from the candidate side. They will feel it.

Hostile questions are still the likely eight with more heat. Do not match heat. Repeat the bar, the clock, the ask. The slightly hostile CFO is why you prepared cost at 10×. Give them the band and the cap. If they want a precision you do not have, label the hole and the week-one method. Then wait. Heat that does not get a dance partner often sits down. You have seen this. Trust it.

When they ask something off the map ("what do you think of vendor Y's new agent"), do not fake a product review. "I have not run Y. I would put it on our forty questions in a death-dated tenant and look at groundedness and $ at 10×. I do not have an opinion from a keynote." That answer is the whole month in two sentences. It works on almost any vendor-flavor question. Keep it.

End of loop: they will ask if you have questions. Use the three from day 26. Then stop asking. Five extra questions at the end of a long panel make you look anxious. Two or three that score the job make you look like a buyer. You are a buyer. You have a pack. You have a seat. Act like it. Tomorrow you will write what you would do if they say yes — and what you will do if they do not, yet.

Artifact #5 is the script after Run 2, plus the eight answers. Put them in the pack. Tick the box. The pack is now speakable. That was the assignment of days 23–29. Day 30 is a different object: a plan for the seat and a plan for the search. Do not spend tonight rewriting week 2 because Run 1 felt bad. Run 1 always feels bad. The month is not a feeling.

If Run 2 revealed a real hole in the design (no refuse path, no non-goals), you may patch that one hole in the wiki page. One. Then stop. Patching is not reopening. Reopening is reading day 14 again and changing nouns. Frozen scope. You would not change a go-live scope the night before because a rehearsal found a font issue. Distinguish font issues from missing kill criteria. Only the second gets a patch.

Sleep is part of the method. A walkthrough in a real loop will happen on a day you also have a job. You need to know you can do this tired. Tonight, tired, after two runs, you still wrote eight answers. That is the proof. A perfect Saturday-morning run is a weaker proof. Do not schedule a third run at 10pm to chase the Saturday feeling.

If you have a loop booked this week, use this script, do not invent a second pack for them unless their posting is the backup seat. Backup seat: you may lead with the pack instead of the demo, or change the intro's seat word. You do not rebuild Atlas. One slice, two altitudes, one pair of seats. That constraint is how you stay sharp. Five custom packs is how you stay busy and unprepared.

Send the "can I walk this to you for 12 minutes" message if you did not last night. A real mouth in front of you, even a friend who does not know AI, will tell you when you sounded unlike yourself. That note is worth more than another pass on the six boxes. If they cannot follow the problem sentence, the problem sentence is still a slogan. Fix that sentence only.

Tomorrow you write 90 days in the seat and a job OS with WIP limits. You also write what you can prove tonight — a close of the month that is evidence, not a pep talk. Bring the pack. You will inventory it. You will not rewrite it. The workstream that is your search starts the day after the month, not during the last night of it. Protect day 30 from assembly-creep.