Behavioral interviews and the research talk

Before this lesson: nothing is required. The technical lessons give material for stories: lesson 12 (a training run you designed) and lesson 14 (a system you shipped). Time: about 70 minutes, then about 90 minutes to build your own story bank with the notebook.

---

Why this matters #

Behavioral rounds look "soft", but they can stop an offer after a strong technical loop. At senior levels they often decide the level of the offer. Research scientist and research engineer loops add a research talk (a "job talk") or a paper deep dive. There the panel decides whether they want to think with you every day. Researchers lose these rounds in predictable ways: a long background, "we" instead of "I", no numbers, no real failure, and a talk that explains the method but not the decisions behind it. All of this can be fixed with preparation, and the preparation can be reused for every company.

1. Predict first #

2. The picture #

Why behavioral questions exist. They rest on one assumption: past behavior predicts future behavior. "Tell me about a time when…" asks for one real situation from your past. The interviewer writes down evidence for a fixed list of signals. The Tech Interview Handbook lists eight common ones: motivation, proactivity, working in an unstructured environment, perseverance, conflict resolution, empathy, growth, and communication. At the end the interviewer recommends hire or no hire, and a level.

The level comes from the scope of your stories. A story whose impact stays in your own work reads as junior. A story that changed the work of the whole team (usually three or more people) reads as senior. A story that changed the work of several teams reads as staff level.

A story is a data point. STAR is only a format that makes the evidence easy to extract: what was at stake, what you did and why, and what changed because of it. A story without a personal decision gives the interviewer nothing to write down.

The first figure is a 150-second answer as a bar. The split decides what the interviewer hears.

Interactive figure: loading…

Why a research talk exists. In 45–60 minutes the panel checks four things:

1. depth: do you understand why your method works, and when it fails?
2. taste and judgment: did you choose a good problem, and make good decisions under limits?
3. communication: can a mixed audience (domain experts, ML people, managers) follow you?
4. fit: do your skills transfer to their problems?

MIT's CommKit says this about industry talks: the heart of the talk is you, your skills and your fit for the role, and the hiring team is interested in your reasoning, not only your results. Behavioral stories and the talk test the same thing from two sides: your decisions.

3. Definitions and main results #

3.1 STAR, with a time budget #

STAR = Situation, Task, Action, Result. Add a second R for Reflection: what you learned, and where you applied it later. The reflection shows growth, which is one of the scored signals.

At about 150 words per minute, a 2–3 minute answer is 300–450 words. For 2.5 minutes (150 s, about 375 words):

Part	Share	Seconds	Content
Situation	~15%	20–25	context, stakes, the constraint (a deadline, a budget, a user); one number
Task	~10%	~15	what you owned; the goal
Action	50–55%	75–80	2–4 actions: the options you saw, what you decided and why, what you did
Result	~15%	20–25	a measurable outcome, including what was lost or not solved
Reflection	5–10%	10–15	the lesson, and a later case where you applied it

Rules that make a story score.

• Say "I" for your decisions and actions. Say "we" only for the team's result. Credit others by role ("the engineer who owned the data pipeline").
• Give at least one number in the Situation and one in the Result.
• The Action contains a decision with an alternative, not only a list of tasks.
• Prepare bullets, not a script. A memorized script sounds recited and breaks at the first follow-up question.
• Share nothing confidential: no internal numbers, no private data, nothing under a non-disclosure agreement. Use relative numbers ("3× fewer GPU-hours") when the absolute ones are private.

3.2 The six themes and what each one tests #

Theme	Typical prompt	What the interviewer checks
Conflict	"Tell me about a disagreement with a colleague."	You understood the other side; you used data or a small experiment; it ended in agreement or in "disagree and commit"; the relationship survived.
Failure	"Tell me about a time you failed."	A real failure with your share in it; no blame; how you noticed it, how you recovered, what you changed afterwards.
Ambiguity	"Tell me about a project with an unclear goal."	How you reduced the uncertainty: questions, a written proposal, a small experiment; a decision with incomplete information.
Influence without authority	"How did you get people you did not manage to change something?"	You understood their goals; you used evidence, a prototype, allies; the change happened.
Ownership (end to end)	"Tell me about something you owned from start to finish."	Scope; follow-through on the boring parts (tests, documentation, hand-over); unblocking yourself and others.
Hard trade-off	"Tell me about a difficult technical decision."	At least two real options; the criteria and the numbers; the risk you accepted; how it turned out.

Other frequent prompts: "Tell me about yourself", your most impactful project, a tight deadline, learning something new fast, mentoring someone, and "why this company?". The six themes cover most of them.

3.3 Building 8–10 stories from a research career #

A research career has more stories than it seems; they hide inside projects. A method:

1. List the periods and projects. Each degree, each paper, each large training run, each collaboration, teaching and supervision, tools nobody asked for, a rejected paper, a project that was stopped.
2. For each one, list the moments. A decision, a disagreement, a mistake, a surprising result, a deadline, a moment when you changed someone's mind or they changed yours.
3. Write a one-line headline for each moment, with a number. For example: "chose the smaller model to fit the GPU budget; lost 2 points of accuracy; got 1.5 back with distillation".
4. Fill a story × theme matrix. Keep 8–10 stories so that every theme has at least two stories, and no story serves more than three themes. Two stories per theme let you pick the better fit for the exact question. They also give you a fresh story when two interviewers in the same loop ask about the same theme.
5. Write bullets for each STAR part, 300–450 words in total. Record yourself once, listen, cut the Situation, and make the Action more concrete.
6. Prepare the follow-ups for each story. Interviewers dig with: "What exactly did you do, not the team?", "What would you do differently?", "What did the other person think?", "How did you measure it?", "What if X had happened?".

Interactive figure: loading…

Where to look in your own work (questions to answer from memory):

• A long training run that failed or had to be restarted: what did you own, and what had you assumed? (failure, ownership)
• Two designs and compute for only one: how did you decide? (trade-off)
• A time you convinced collaborators, a supervisor or another group to adopt a metric, a constraint or a tool. (influence, conflict)
• A project that started with a vague question. (ambiguity)
• A student or junior colleague you helped through a failed experiment. (influence, growth)

"Tell me about yourself" (60–90 s): who you are now (role, area, one result); the path in two sentences; why this role is the next step. Keep only what matters for this job.

3.4 The research talk: structure and timing #

The 45-minute plan. Industry talks are often a 60-minute slot: about 40–45 minutes of talk and 15–20 minutes of questions. Ask the recruiter for the exact format and the audience.

Minutes	Part	Content
0–3	Opening	The problem in one sentence a non-expert understands; why it matters; your main result in one sentence.
3–9	Background and gap	Only what the audience needs: what is known, what is missing, why it is hard.
9–17	The key idea	The simplest version first (one picture), then the real version.
17–29	Main results	2–3 results with their evidence: baselines, ablations, a failure case.
29–36	Design decisions	The 3–4 choices that mattered, their alternatives, and what did not work.
36–40	A second line of work	Optional: a short example with a common thread (a method, a skill).
40–44	Limits and next steps	What fails today; what you would do next; the link to their problems.
44–45	Summary	Three take-aways.

Slides. One message per slide, written as the slide title in a full sentence ("The graph model halves the error at gauges", not "Results"). Every plot has labeled axes, units and a baseline. Show an equation only with its meaning in words next to it. For a mixed audience, give each section in layers: first the message that everyone can follow (the "why" and one picture), then the detail for the experts, then the message again. As a rule of thumb, one slide per 1–2 minutes. Put deep details on backup slides after the summary.

The 5-minute version (about 600 words at 140 words per minute, with some slack):

Time	Part
0:00–0:45	the problem and why it matters, in plain words, and the main result in one sentence
0:45–2:00	the idea, in one picture
2:00–3:30	one result, with one number and its baseline
3:30–4:15	the limits, and why it matters for this team
4:15–5:00	slack (questions, or running over)

The 60-second version keeps one sentence from each of the first four rows: the problem in plain words, what you built, one result, and why it matters for this job (about 150 words).

Interactive figure: loading…

3.5 Defending every design choice #

Before the talk, write a decision table with 10–15 rows, one for each choice that a careful listener could question. The table below belongs to an invented example talk: "A fast surrogate model for river flood forecasts". All names and numbers are made up to show the form.

Choice	Alternatives	Why this one (evidence)	Cost or risk	What would change my mind
a graph network on the river mesh	a CNN on a regular grid; a Fourier neural operator	the mesh follows the river; 20% lower error than a CNN of the same size on the test basins	slower data loading; fewer reference implementations	a grid model that matches it once it gets the river-bed depth as input
pretrain on simulations, fine-tune on gauge data	train on gauge data only	gauges cover 2% of the nodes; pretraining halves the error at the gauges	the simulator's bias carries over	the gain disappears with five more years of gauge data
a loss weight of 5 on flood peaks	plain mean squared error; a quantile loss	peak error falls from 30 cm to 12 cm; the mean error rises by 1 cm	one more weight to tune; bias at low water	users care more about low water (shipping) than about floods

A row without evidence is a weak spot. Run the ablation, find the reference, or say openly that it was a judgment call, and why.

The usual panel questions: "Why not X?", "Does it generalize to Y?", "What is the weakest part?", "How would it scale with 10× the data or compute?", "What would you do next with six months?", "How would this help our product or our science?".

3.6 Answering questions: a routine #

1. Listen to the end, then pause for one or two seconds.
2. Clarify if the question is unclear: "Do you mean the training cost or the inference cost?"
3. Answer first, in one sentence. Then give the reason or the evidence.
4. If you do not know, say so. Then reason aloud: what you would expect, and how you would check it (an experiment, a reference). This is better than a guess, and it shows how you think.
5. For a hostile or very long question, accept the valid part, answer the core, and offer to continue after the talk.
6. Use backup slides for the deep questions you expect.
7. Close the loop: "Does that answer your question?"

3.7 The paper deep dive #

Some loops ask you to present a paper: your own, or a recent one the panel picks (20–30 minutes). It uses the same tools, plus a critical part:

1. The problem and its assumptions.
2. Earlier approaches and the limit this paper attacks.
3. The key idea in one sentence and one picture; then the method in enough detail to implement it.
4. The evidence: do the baselines and ablations support the claims?
5. Your assessment: strengths, weak points, missing experiments, compute cost.
6. Next steps: what you would try, linked to the team's problems.

Expect to derive one key equation and to answer "why this loss and not another?". Keshav's three-pass method helps to read the paper: first the title, abstract, headings and conclusions; then the figures and the main argument; then the details, as if you had to re-implement it.

Answers to section 1. (1) About 30–40 s. Situation and Task together take about 25% of the answer, so most of the time goes to the Action, where the evidence is. (2) Minute 3. State the main result in the opening; the rest of the talk earns it. A mixed audience that waits 20 minutes for the point stops listening.

4. Worked example: one story and its timing #

The story below is invented, to show the form. Use your own facts in your stories.

Prompt: "Tell me about a hard technical trade-off you made under a deadline." Theme: hard trade-off. The same event also serves for ownership, and for failure if you center it on the leak in your own pipeline.

Part	Bullets (what the speaker says)	Spoken words (target)
Situation	"In the second year of my PhD I trained a model that segments crop fields in satellite images. Two weeks before a conference deadline, our test score was 0.91 IoU, 6 points above the best baseline."	~55
Task	"I was the first author. I owned the data pipeline and all experiments."	~25
Action	(1) "During the error analysis I saw test tiles that looked almost the same as training tiles. I wrote a near-duplicate check and found that 18% of the test tiles had a near copy in the training set, because neighbouring tiles overlapped." (2) "I saw two options: submit with a footnote, or rebuild the split by region and rerun everything, which needed about 9 of the 14 days left." (3) "I chose to rebuild. I told my supervisor the same day, with a one-page note: the overlap numbers, the plan, and the risk of missing the deadline." (4) "To save time I reran the main model and the two strongest baselines first, and moved the ablations to the appendix."	~200
Result	"We submitted on time. The honest score was 0.84, still 4 points above the best baseline, because the baselines dropped too. The paper was accepted. We lost two ablations from the main text."	~55
Reflection	"Since then I build the split by group before the first experiment, and I run a near-duplicate check on every new data set. In my next project it found an overlap on the first day."	~30
Total		~365 words ≈ 2 min 26 s

Check of the shape, step by step.

1. Total: $55 + 25 + 200 + 55 + 30 = 365$ words. At 150 words per minute this is $365/150 = 2.43$ minutes $= 146$ s.
2. Action share: $200/365 = 55\%$, inside the 50–55% target.
3. Situation + Task: $80/365 = 22\%$, which is $0.22 \cdot 146 \approx 32$ s. This matches predict 1.
4. The trade-off has two options, the criteria (time left, honesty of the result), the numbers (18%, 9 of 14 days), the accepted risk, and a result that includes what was lost.
5. The reflection names a later case where the lesson was applied.

Weak and strong lines.

• Weak: "We found some issues with the data and fixed them." Strong: "I wrote a near-duplicate check, found 18% overlap, and rebuilt the split by region."
• Weak: "The deadline was too short." (blame) Strong: "I had not checked for overlap when I built the split; now I do it before the first experiment."

Follow-ups to prepare for this story: "Why not submit with a footnote?" "How did your supervisor react?" "How do you know the new split is clean?" "What would you do with one more month?"

5. In code #

A small tool for your own story bank: does every theme have at least two stories, does any story carry too many themes, and does each draft fit the time budget? The stories are placeholders; replace them and the word counts with your own.

THEMES = ["conflict", "failure", "ambiguity", "influence", "ownership", "trade-off"]
WPM = 150                                   # speaking rate, words per minute

# story -> (themes it can answer, words per part: S, T, A, R, Reflection)
bank = {
    "S1": ({"failure", "ownership"},   (50, 20, 190, 50, 30)),   # a training run that crashed
    "S2": ({"trade-off", "ambiguity"}, (55, 25, 200, 55, 30)),   # two designs, compute for one
    "S3": ({"influence", "conflict"},  (60, 20, 180, 60, 25)),   # another team adopted a metric
    "S4": ({"conflict"},               (80, 30, 140, 50, 20)),   # disagreed about a paper's claim
    "S5": ({"ownership", "trade-off"}, (45, 20, 210, 60, 30)),   # a data leak before a deadline
    "S6": ({"ambiguity"},              (50, 25, 170, 45, 25)),   # a project with a vague goal
    "S7": ({"influence"},              (40, 20, 160, 40, 30)),   # helped a junior colleague
    "S8": ({"ownership", "influence"}, (50, 20, 260, 60, 40)),   # a tool nobody asked for
}

for theme in THEMES:
    stories = [s for s, (themes, _) in bank.items() if theme in themes]
    flag = "   <-- needs a second story" if len(stories) < 2 else ""
    print(f"{theme:<10} {len(stories)} {' '.join(stories)}{flag}")

for s, (themes, parts) in bank.items():
    words = sum(parts)
    seconds = 60 * words / WPM
    flags = []
    if len(themes) > 3:
        flags.append("too many themes")
    if not 120 <= seconds <= 180:
        flags.append("outside 2-3 min")
    if parts[2] / words < 0.45:
        flags.append("Action below 45%")
    if (parts[0] + parts[1]) / words > 0.30:
        flags.append("Situation + Task above 30%")
    print(f"{s} {words:4d} words {seconds:4.0f} s  Action {parts[2] / words:4.0%}  {', '.join(flags)}")

Output:

conflict   2 S3 S4
failure    1 S1   <-- needs a second story
ambiguity  2 S2 S6
influence  3 S3 S7 S8
ownership  3 S1 S5 S8
trade-off  2 S2 S5
S1  340 words  136 s  Action  56%
S2  365 words  146 s  Action  55%
S3  345 words  138 s  Action  52%
S4  320 words  128 s  Action  44%  Action below 45%, Situation + Task above 30%
S5  365 words  146 s  Action  58%
S6  315 words  126 s  Action  54%
S7  290 words  116 s  Action  55%  outside 2-3 min
S8  430 words  172 s  Action  60%

The bank needs a second failure story. S4 spends too long on context, S7 is a little short, and S8 is close to the 3-minute limit. The notebook turns these checks into exercises, with a talk planner and a timing check for a recorded answer.

6. Where it shows up #

• Every interview loop. A behavioral or hiring-manager round, and behavioral questions inside technical rounds ("tell me about a project where…").
• Research scientist and research engineer loops. A research talk or a paper deep dive.
• ML for science. The audience is always mixed: domain scientists, ML researchers and engineers. Explaining a generative model to climate scientists, or a physical constraint to ML engineers, is daily work, and the talk shows that you can do it.
• After the interview. The same stories and the 60-second version serve in networking, recruiter calls, grant and project proposals.

7. Common mistakes #

• A long Situation. One minute of context leaves no time for the Action.
• "We" everywhere. The interviewer cannot tell what you did.
• No numbers. "It worked better" gives no evidence; "from 12 h to 3 h per run" does.
• A fake failure ("I work too hard"), or a failure without your own share in it.
• Blaming colleagues, supervisors or the cluster.
• One story for every question, or a story with the wrong theme for the question asked.
• A memorized script that breaks at the first follow-up.
• No reflection, or a generic one ("communicate more").
• Confidential details.
• In the talk: 15 minutes of background; the main result only at the end; equations without their meaning; no design decisions; running over time; defensive answers; no link to the company's problems.

8. Check yourself #

1. A 3-minute answer at 150 words per minute: how many words, and how many seconds for an Action share of 55%?
   Answer

   450 words; $0.55 \times 180 = 99$ s for the Action.
2. Rewrite "We improved the model and the paper got accepted" so that it shows ownership.
   Answer

   A sentence of this form, with your facts: "I ran two ablations, found that the input resolution limited the model, proposed a multi-scale encoder to my supervisor with the plot, and implemented it; the error fell from 0.30 to 0.22, and the paper was accepted." It names your actions, your decision with its evidence, and a measurable result.
3. Why should every theme have at least two stories?
   Answer

   The exact question varies (a conflict with a peer or with a manager), so you pick the better fit; several interviewers in one loop may ask the same theme; and "tell me about another time" is a common follow-up.
4. What separates a failure story from a hard-trade-off story?
   Answer

   A failure story centers on a mistake or a bad outcome that was partly yours: how you noticed it, recovered, and changed your behavior. A trade-off story centers on a decision between options under constraints, with criteria, numbers and an accepted risk; the outcome can be good. One event can give both, told with a different center.
5. In a 45-minute talk, where does the main result first appear, and what goes into the last minutes?
   Answer

   In the opening, within the first 3 minutes, in one sentence. The last minutes: the limits (what fails today), next steps, the link to the hiring team's problems, and a summary with three take-aways. Deep details go to backup slides.
6. A panelist asks about the invented flood talk: "Why a graph network instead of a CNN on a grid?" Give the structure of a good answer.
   Answer

   Answer first in one sentence: "Because the river mesh is irregular, and on our test basins the graph model had 20% lower error than a CNN of the same size." Then the evidence: the comparison and its setup. Then the cost: slower data loading, fewer reference implementations. Then what would change your mind: a grid model that matches it once it gets the river-bed depth. In a real talk every number must be your own.

Recall cardsAnswer in your head first, then reveal. Add the cards to your review deck to see them again just before you would forget.

1. In a 2.5-minute STAR answer, about how much time goes to the Action, and how much to Situation plus Task?

   Show answer

2. What does the second R in STAR + R stand for, and which signal does it show?

   Show answer

3. What sets the level that a behavioral round supports?

   Show answer

4. Name the first three of the six behavioral themes.

   Show answer

5. Name the last three of the six behavioral themes.

   Show answer

6. What coverage rule should a story bank meet?

   Show answer

7. By which minute of a 45-minute research talk should the audience hear the main result?

   Show answer

8. What does a 5-minute research talk keep?

   Show answer

9. Name the five columns of a decision table for a research talk.

   Show answer

10. What do you do when a panelist asks a question you cannot answer?

    Show answer

Add these cards to my reviewOpen the review →

9. Interview angle #

Recognize the format and the theme from the wording.

The interviewer says	Theme or format	Bring
"a disagreement", "pushed back", "a colleague who…"	conflict	the other view stated fairly, data, the end state, the relationship
"failed", "a mistake", "went wrong"	failure	your share, how you noticed, recovery, what you changed
"unclear", "no clear requirements", "open-ended"	ambiguity	questions, a written proposal, a small experiment, a decision
"convince", "people you did not manage", "adopt"	influence without authority	their goals, evidence or prototype, allies, the change
"from start to finish", "owned", "went beyond your role"	ownership	scope, the boring parts, unblocking others
"difficult decision", "trade-off", "two options"	hard trade-off	options, criteria, numbers, the accepted risk
"present your work", "job talk"	research talk	main result first, one picture, decisions, limits, fit
"walk us through this paper"	paper deep dive	claim, method, evidence, assessment, next steps

The routine to say aloud for any behavioral question: pick the story (pause two seconds; that is fine), give the Situation in two or three sentences with one number, say what you owned, spend most of the time on what you decided and did, give the result with a number and what was lost, and end with one sentence of reflection. Then stop and wait for the follow-up.

Formats you will meet.

• Behavioral round (30–45 min): 4–6 questions, each with 2–3 follow-ups. The follow-ups test whether the story is yours and true.
• Hiring-manager round: motivation, career plan, the most impactful project, "why us".
• Research talk (45–60 min with questions) and paper deep dive (30–45 min).
• "Tell me about yourself" opens almost every round: 60–90 seconds.

Red flags that interviewers write down: no clear personal role; no real failure; blaming; no numbers; answers longer than about 4 minutes; confidential details; a talk that cannot answer "why this choice?".

Typical questions and short model answers (the shape of a strong answer; fill it with your facts).

1. "Tell me about a time you disagreed with a senior colleague about a technical decision." — State the stakes and the decision in 30 seconds. State the other person's view fairly, and why it was reasonable. Say what you did to resolve it: data, a small experiment, a written comparison, a meeting with one clear question. Give the end: agreement, or "disagree and commit" and how you then supported the decision. Close with a measurable result and what you learned about disagreeing well.
2. "Tell me about a failure." — Choose a real failure where your share is clear. Say how you noticed it, what you did in the first hours or days, and what it cost. Most of the answer is the recovery and the change in how you work, with a later example where the change helped.
3. "Tell me about yourself." — "I am a [role] working on [area]; my main result is [one result with a number]. I came here through [two sentences]. I want this role because [one link between your skills and their problem]." Sixty to ninety seconds.
4. "Give a 5-minute overview of your research to a panel with a domain scientist, an ML researcher and an engineering manager." — The problem in plain words and the main result in the first 45 seconds; one picture of the idea; one result with its baseline; why it matters for the team; every term explained or left out; finished within 5 minutes.
5. "What is the weakest part of your work?" — Name it in one sentence, give the evidence (where it fails, by how much), say what you tried, and say what you would try next. Do not hide it; the panel already sees it.

Typical follow-ups. "What exactly did you do?" "What would you do differently?" "What did the other person think?" "How did you measure it?" "Tell me about another time." After a talk: "Why not X?", "Does it generalize to Y?", "How does it scale with 10× the data?".

10. Practice #

The notebook for this lesson runs in Colab. It has three small tools as exercises, each with a check cell and a folded solution:

1. coverage(bank): from a stories × themes table (pandas), return the themes with fewer than two stories and the stories with more than three themes.
2. plan_talk(outline, total_s): scale a 45-minute outline to 5 minutes. Drop the optional blocks, scale the rest, round to 15-second steps, and keep the exact total.
3. star_timing(marks, end_s): from the time stamps of a recorded answer, compute the seconds and the share of each STAR part, and flag the problems.

Timed practice. Build your own story bank (about 90 minutes): steps 1–4 of section 3.3, then check it with exercise 1. Write bullets for each story, record each story once, and redo the ones above 3 minutes or with an Action share below half. Then answer the three hardest follow-ups of each story aloud. For the talk: write the 5-minute version first, then the 45-minute version and the decision table (10–15 rows), and rehearse twice with a recording and a timer.

Say it aloud drill. Ask a friend to read you one prompt from each of the six themes, in random order, and to interrupt with one follow-up. Answer each in 2–3 minutes with a timer. Then give your 60-second "tell me about yourself" twice in a row, without notes.

11. Watch and read #

How to SpeakMIT OpenCourseWare · Patrick Winston · 64 min

How to start a talk with a promise, repeat the main idea, use slides and props, and end with your contributions.

How to Give a Great Research TalkMicrosoft Research · Simon Peyton Jones · 59 min

One key idea per talk, motivate with an example first, and do not try to present the whole paper.

• Tech Interview Handbook, Behavioral interviews: the question types, STAR, and how to prepare stories from your projects.
• Tech Interview Handbook, Behavioral interview rubrics: the eight focus areas and how scope sets the level.
• Tech Interview Handbook, Self introduction: the "tell me about yourself" answer.
• MIT Communication Lab CommKit, Industry job talks: audience, structure, academic versus industry talks, preparing for questions.
• S. Keshav, How to Read a Paper (ACM SIGCOMM Computer Communication Review, 2007): the three-pass method, useful for the paper deep dive.
• MIT OpenCourseWare, How to Speak: the course page of Winston's lecture.