The Interview Prep Budget, Each Round Type Scored Drill, Refresh, or Skip

This is the standard I use to decide which interview round types get a job seeker's limited weekly prep hours. It is for anyone with a target list of 10 to 30 employers and more formats to practise than hours to practise them. It scores each round type on four dimensions. The result is a weekly hour split where every round type has a label (Drill, Refresh, or Skip) and a date to re-score.

What does a round-type prep budget decide?

A round-type prep budget decides how many of your weekly prep hours each interview format gets. The basis is what your own target employers actually run, not what the market runs on average. It replaces "practise everything" with a split you can defend and re-check every quarter.

A round type is an interview format that needs its own skill and its own way of practising. This standard scores seven:

• Algorithmic coding: solve a problem from scratch with no assistant, usually against the clock.
• AI-assisted coding: build or fix code with an AI assistant, either your own or one the employer provides.
• System design: design a service or architecture out loud and defend the tradeoffs.
• Case: a structured business or product problem, worked through live.
• Behavioral: questions about past experience, scored against the employer's values or levels.
• Take-home with live defense: build something on your own time, then explain, review, or extend it live with an engineer.
• In-person onsite: any round held in the room, which takes away your own screen, tools, and notes.

Each round type ends with one of three labels:

• Drill: dedicated hours every week, timed practice, and a target for measured improvement.
• Refresh: a fixed maintenance floor. Enough to stay fluent, not enough to grow.
• Skip: zero hours until the next re-score.

You score the employer-by-round pair, not the employer. Employers tend to change one round at a time as a pilot rather than redesign the whole loop. Meta, for example, runs an AI-enabled coding round where AI is allowed inside a tool Meta provides. Its prep guide still asks candidates not to use unauthorized outside assistance, including AI tools like ChatGPT. Score "Meta" as a single stance and you will be wrong about at least one of its rounds.

Score the round, not the company. One employer can ban AI in one room and provide it in the next.

Which four dimensions decide where prep hours go?

Four dimensions decide it:

• how often the round appears across your target set
• whether failing it ends the loop
• which way the format is moving
• how far you are from passing it today

The first three measure the market. The fourth measures you.

The scales are my own choice, not a published standard. I looked for a documented way to count rounds, a minimum number of reports per employer, or a recency window, and found none. Treat my thresholds as a starting point. Using the same thresholds across the whole list matters more than the exact cut points.

Frequency (0 to 3)

Count employers, not reports. For each round type, count the employers in your set that show it in their own published process or in candidate reports dated after their last policy change. Divide by the number of employers you have usable data for.

• 0: no employer in the set
• 1: fewer than a third
• 2: a third to two thirds
• 3: more than two thirds

What it proves: you will meet this round if you keep applying to this list. When it lies: candidate reports lag policy changes by months, so counting every report ever posted overstates the old format. A Canva report from before its June 2025 announcement describes a Computer Science Fundamentals screen that Canva replaced with an AI-assisted coding round.

Gate (0 or 2)

Score 2 if failing the round ends your candidacy before anything else is weighed. That covers a phone screen, an online assessment, or a stage with a stated hard filter. Anthropic's CodeSignal assessment is a clean example: each level stays locked until you pass the one before it. Score 0 otherwise. If you don't know whether a round is a gate and it comes before the onsite, score it 2.

What it proves: a weak result here can't be made up for anywhere else. When it lies: employers rarely label their gates. A behavioral round that is really a hard filter can read like a friendly chat in candidate write-ups.

Direction (-1, 0, +1)

Score +1 if dated evidence shows employers in your set adding the format, -1 if they are replacing it, and 0 if neither. Evidence means a dated employer page, a dated invite, or a dated pattern in reports. A general trend piece doesn't count.

What it proves: next quarter's mix will differ from this quarter's. When it lies: press coverage of internal documents describes plans, not policy. A Google spokesperson still called its AI-assisted round a pilot even as internal documents pointed to a wider rollout.

Gap (0 to 3)

Run one timed mock per round type, under the conditions the employer actually uses, with a peer scoring.

• 0: passed with time to spare
• 1: passed, but only just
• 2: partial solution or shaky structure
• 3: failed

What it proves: where extra hours buy the most improvement. When it lies: practising in your own editor with your own assistant tells you nothing about a round where the employer supplies the tool or bans tools entirely.

Add the first three dimensions together to get a relevance score: Relevance = Frequency + Gate + Direction, which runs from -1 to 6. Gap stays separate because it is the only dimension you control.

Where does the evidence for each score come from?

The evidence comes from five sources, and they rank differently. The employer's own published page beats the invite, the invite beats the recruiter, and all three beat press reports and candidate write-ups. Record which tier each stance came from, because a stance is only as reliable as its source.

Here is what this looks like for five employers, ordered by how authoritative the source is:

Use this table as a template for your own rows, not as current truth. Every stance has a date because every stance can change, and several already have. What each row teaches:

• Canva says on its own engineering blog that it expects backend, machine learning, and frontend candidates to use tools like Copilot, Cursor, and Claude in technical interviews. That is the strongest kind of evidence. Canva also says it still assesses computer science fundamentals through that round.
• Anthropic bans AI by default on its own site, with one dated exception: an engineering article allows AI for the performance-engineering take-home. Earlier job postings said "Please do not use AI assistants during the application process," and that text was later replaced. That is one employer with three dated stances.
• Meta allows AI only inside its own interface, in one round of about 60 minutes. Candidates pick from several models, including GPT-5, Claude Sonnet 4.5, and Gemini 2.5 Pro. The assistant can't edit your files, so you write or paste every line yourself.
• Amazon's stance comes from internal guidance reported by Business Insider. It warns that breaking the rule "may result in disqualification from the recruitment process." Amazon's public material does not publish one universal rule for candidates.
• Google's round is a pilot "code comprehension" round with Gemini as the approved assistant. Press reports place it in junior to mid-level roles on select US teams.

I found no employer that publishes an "allowed with disclosure" policy. Treat that category as unproven, and don't put any employer in it without its own written words.

When a stage isn't covered, ask. The published method is simple. Read the invite and assessment instructions for "AI," "generative," or "assistance." If they say nothing, email the recruiter. If you still can't get a clear yes, treat it as a no.

Hi [Recruiter name],

Thanks for scheduling my [role] interviews. Before I prepare, I want to make sure I follow your rules on AI tools for each stage:

1. [Online assessment / phone screen]: are AI assistants allowed, and if so, only a tool you provide?
2. [Coding round(s)]: same question.
3. [Take-home, if any]: may I use AI while building it, and will there be a live review afterward?

If any stage allows AI, could you tell me which tool or environment you'll use so I can practise in it?

Thanks,
[Your name]

How many candidate reports are enough?

No one has published a minimum, so this is my choice. I use at least three reports per employer, all dated after its last known policy change. An employer with fewer gets marked thin and counts at half weight in the frequency tally. Reports older than the employer's last policy change are thrown out, however many there are.

Which way are interview formats moving?

Four movements show up in the evidence:

• AI-assisted coding rounds appearing at some employers
• AI bans and in-person rounds coming back at others
• system design reaching below senior level
• take-homes that end in a live defense

Score Direction for each round type from evidence inside your own target set. The sections below tell you what to look for.

AI-assisted coding is replacing single rounds, not whole loops

Canva piloted "AI-Assisted Coding" in place of its Computer Science Fundamentals screen. A secondary source reports that Meta replaced one onsite coding round with a 60-minute AI-assisted session. Google is reported to be piloting a similar round. In each case one round was swapped out and the rest of the loop stayed. Score AI-assisted coding +1 if any employer in your set has a dated swap. Score algorithmic coding -1 only at the employers where the swap actually removed it.

Is LeetCode still worth it? Yes, though not always at Drill. Canva still assesses computer science fundamentals through its AI round. When the rubric shifts toward checking the assistant's output, algorithm gaps still sink candidates. Wherever AI-assisted rounds appear, algorithmic coding stays at Refresh at minimum.

AI bans and in-person rounds are coming back at other employers

Pressure from interview fraud is pushing the other way. Google's CEO said "we'll introduce at least one round of in-person interviews for people just to make sure the fundamentals are there." Computerworld names Google, Cisco, and McKinsey among employers bringing rounds back in person.

Checkr also found that 60% of managers say AI has made virtual interviews harder to trust, up from 51% the year before. Fraud pressure and AI acceptance push toward the same place. In-person rounds and provided-tool rounds both take away the candidate's own hidden tools. If your set shows either one, the in-person onsite row needs practice without your editor: talk the problem through out loud, on paper or a whiteboard.

System design is reaching below senior level, unevenly

Practitioner guides agree on the shape. Most entry-level software engineering roles have no system design interview, though there are plenty of exceptions. The round gets more common at mid-level and is normal at senior. interviewing.io says mid-level and senior candidates will get at least a few system design questions. It adds that some employers put a system design component into the technical phone screen, which is more likely at a startup than at a FAANG company.

No public survey I found breaks round mix down by company size and level, so that is not established. A widely repeated claim that 68% of candidates failed at system design has no primary source I could find. Leave it out of your scoring.

That figure tells you entry-level engineers at large employers claim the skill. It does not tell you their interviews tested it. I don't have a clean senior comparison from the same pull, so I make no claim about how the share changes with level. For a system design interview at mid-level, score Frequency from your own set's reports. If startups dominate your list, check whether system design shows up in the phone screen, which would also make it a gate.

Take-homes increasingly end in a live defense

Are take-home assignments still used? Yes, in the examples here, but more of them end in a live session. Anthropic cut its performance take-home's time limit "from 4 hours to 2 hours." OpenAI has some teams build a piece of software alone, then review and extend it with an engineer. When a live defense follows, that follow-up is functionally the real test. Score the take-home row on the defense. The gap that matters is how well you can explain and extend your own code, not how polished it looks.

How do the scores become Drill, Refresh, or Skip?

Relevance tells you whether a round matters, and Gap tells you whether hours will move it. Drill where both are high. Refresh where only one is. Skip only where relevance is low and the round is not a gate.

The labeling rules, using my thresholds:

1. If Gate = 2, the round can't be labeled Skip.
2. Skip if Relevance is 1 or less and Gate = 0.
3. Drill if Relevance is 3 or more and Gap is 2 or more.
4. Refresh in every other case.

Splitting the hours

1. Start from your real weekly budget. For reference, the Tech Interview Handbook plan suggests about 11 hours a week for roughly three months, about 140 hours in total.
2. Give each Refresh round a floor of one hour a week. That number is my default.
3. Divide the remaining hours across Drill rounds in proportion to Relevance times Gap.
4. If the Refresh floors use up the whole budget, too many rows are Refresh. Trim the target list, or move the lowest-relevance non-gate rows to Skip.

Published estimates give you something to sanity-check against. Most are vendor or blog estimates, not measured outcomes:

Plans also disagree on when to start system design. One says to start at the halfway point, because 2 to 3 hours a week beats 15 hours of cramming. Another puts it in weeks 9 and 10. My view is that the score decides. If system design scores Drill, start now at a steady 2 to 3 hours a week rather than saving it for a late block.

Worked example (made-up numbers, 11 hours a week, mid-level backend): behavioral scores Relevance 5, Gap 1, so it's Refresh. AI-assisted coding scores Relevance 2, Gap 2, so it's Refresh. In-person onsite scores Relevance 3, Gap 1, so it's Refresh. Case scores Relevance 0, so it's Skip. Algorithmic coding scores Relevance 4, Gap 2 (weight 8), and system design scores Relevance 3, Gap 3 (weight 9), so both are Drill. Three Refresh floors take 3 hours. The other 8 split roughly 4 and 4 between the two Drill rows.

The budget assumes your prep hours really go to prep. If rewriting your resume and cover letter for every application eats into the same week, Refolk tailors both to each posting from your own history. That hands those hours back to the Drill rows.

Step-by-step: scoring your target list

The procedure takes about a week of elapsed time. Most of that is waiting on recruiters and booking mocks, and the actual work is a few hours. Do it once in full, then repeat steps 2, 4, and 7 at each re-score.

Step 1 is where most lists go wrong. They are wishlists, not lists of places that would interview you. Refolk scores how well you actually fit each posting, which is a quick way to cut a long wishlist down to the 10 to 30 employers worth scoring at all.

Round type: [e.g. System design]
Target set: [N] employers, [level], [function]   Thin employers (half weight): [list]

Frequency (0-3): [employers showing it] / [employers with usable data] = [share] -> [score]
Gate (0 or 2): [which employers gate on it, and the evidence]
Direction (-1/0/+1): [dated evidence of add or replace, with source tier]
Relevance = F + G + D = [ ]

Gap (0-3): mock on [date], conditions [own tool / provided tool / no tool / in person], scored by [peer] -> [score]

Label: [Drill / Refresh / Skip]
Weekly hours: [ ]   Weight if Drill (Relevance x Gap): [ ]
Re-score by: [date]   Re-score now if: [employer + policy page to watch]

How does this scoring go wrong?

It usually goes wrong by trusting a source or a stage beyond what it actually covers. The most common failures are treating press reports as policy, treating a provided tool as permission to use your own, and labeling a gate Skip because it looked rare. Each failure below comes with the false positive it produces and the check that catches it.

1. Treating press reports as policy. Google's and Amazon's stances come from leaked internal documents. False positive: you mark "AI allowed" for all of Google when the pilot covers select US teams only. Check: does the employer's own page or your invite say so? If not, the row stays at its tier and doesn't count as permission.
2. Treating a provided tool as permission for your own. Meta allows AI only inside its own tool, and that assistant can't edit your files. False positive: you Drill with your own assistant and score a Gap of 0 that doesn't hold in the real round. Check: re-run the mock under the employer's tool conditions.
3. Assuming one policy covers every stage. Anthropic bans AI by default but exempts one take-home. At OpenAI, permission to use AI at home doesn't necessarily carry over to the interview. False positive: one stance applied to the whole loop. Check: score each stage separately, which is why the unit is employer by round.
4. Stale counts. A Canva report from before June 2025 describes a round that no longer exists. False positive: high Frequency for a retired format. Check: throw out every report dated before the employer's last policy change.
5. Scoring Skip on a gate. Skip behavioral at a company where it is a hard filter and you fail the whole loop, however strong your coding is. False positive: low relevance from a small employer count. Check: flag gates before applying any gap weights, and apply the rule that a gate is never Skip.
6. Reading take-homes as standalone. Where a take-home has a live defense, the follow-up is functionally the real test. False positive: hours spent polishing code nobody grades on polish. Check: budget time to rehearse the defense, explaining and extending your code live.
7. Inflated adoption stats. Claims like "about 25% of US employers now permit AI assistance" come with no stated method. False positive: a Direction of +1 built on a number nobody can trace. Check: throw out any figure without a named sample, and score Direction only from dated evidence inside your own set.

There is a quieter failure too: letting an unproven category in. No employer publishes an "allowed with disclosure" policy that I could find. If you score one there, you are guessing. Mark it silent and send the recruiter question.

When should you re-score the budget?

Re-score quarterly by default, and right away when any employer in your set changes a written policy. Formats are changing faster than candidate reports pile up, so a split that was right at the start of a search can be wrong halfway through it.

The quarterly interval is my own derivation, not a published standard. Across the five employers in the stance table, there are nine dated policy or format events in about 19 months, roughly one every two months. Anthropic alone logged three: an application-stage ban, a revised policy, and a take-home exception. With a target set of 10 to 30 employers, you should expect at least one change per quarter somewhere in it.

At each quarterly re-score, also re-run one timed mock for every Drill row. If the Gap drops to 1 or 0, the row moves to Refresh and its hours go to the next-heaviest Drill row. That is the point of the system: hours follow the weakest round that matters, not the round you most enjoy practising.

The output is one sheet: seven rows, each with a label, an hour count, and a date. If someone asks why you are skipping case interviews or drilling system design, the sheet answers them. When the date comes round, run the procedure again.