A curated list gives graphs three problems but your target loop is graph-heavy — what do you do?

## What a category weight in a curated list means Curated lists are compressed by construction. Their authors sampled a large corpus and picked a few representatives per category, weighting roughly by how often each category appeared across a general population of interview loops. That weighting is a reasonable prior and it is not your posterior. If you know the loop you are preparing for leans heavily on one category — because of the domain, the team, published interview accounts, or a recruiter's own description of the rounds — the list's general-purpose weights are the wrong weights for you. So the skill is reading the list as a map with an explicit scale, noticing where the scale is coarse for your purposes, and refining exactly there. ## Coverage is not one thing: categories contain sub-patterns The reason three problems can genuinely cover one category and badly under-cover another is that categories differ enormously in internal variety. A narrow category may have a single mechanism with small variations; two problems really do span it. A broad one — graphs is the standard example, dynamic programming another — is a family of related but distinct mechanisms. For graphs the sub-patterns include, at minimum: exploring depth-first versus level-by-level and what each gives you, connected components and reachability, shortest paths in unweighted graphs versus weighted ones with non-negative weights, detecting cycles, ordering nodes so dependencies come first, merging groups incrementally, and the recurring trick that the graph is *implicit* — nodes are states and edges are legal moves, with no edge list anywhere in the statement. A three-problem sample cannot touch all of that. It will typically hit traversal, maybe components, and leave the rest blank. So the audit is not "is the category present" but "which sub-patterns did these three problems actually instantiate". ## Running the audit The procedure is mechanical and takes under an hour: 1. **Write down the sub-patterns** of the category from your own knowledge, not from the list. If you cannot enumerate them, that is itself the finding — you do not yet know the category well enough to judge its coverage, and the first supplement is learning its structure. 2. **Map each list problem onto a sub-pattern.** Some will map to two; some sub-patterns will get nothing. 3. **Rank the uncovered ones by likelihood for this loop.** Not all gaps deserve equal funding. Dependency ordering and shortest paths come up far more than exotic sub-patterns. 4. **Supplement two to three problems per gap**, easy before medium, and keep them together as a block so the within-category comparison that builds the pattern actually happens. 5. **Re-run your recall check on the whole category afterwards**, including the original three, since the goal is a category you can enter cold — not a longer list of solved rows. ## The constraints that keep the supplement honest Supplementing is a reallocation, and the time comes from somewhere. Three rules bound it: **Never fund depth from a blank category.** A thin category still gives you a starting move on an unseen problem; a blank one gives you nothing. If some category on the map is still at zero, it outranks the seventh graph problem, even for a graph-heavy loop — because loops are not homogeneous and one round will not be about the specialty. **Cap the depth.** The goal is being able to start confidently on an unseen variant, not exhaustive mastery. There is always another sub-case. Past roughly the point where you can name the sub-pattern and sketch the approach, further problems in the same sub-pattern buy very little. **Prefer breadth across sub-patterns over depth within one.** When time is short, one problem in each of four uncovered sub-patterns beats four problems in one of them, for the same reason breadth-first beats count-maximizing at the category level: the failure that ends a round is having no move, not having a slightly rusty move. ## When advising someone else The same audit is the most useful thing you can do for a person who is preparing and feels stuck. They usually arrive with a list and a completion percentage. Ask which sub-patterns of their heaviest category they can name, and the gap becomes visible in a couple of minutes — to them, not just to you, which is what makes them able to run the audit again themselves later. The most common mistake to talk them out of is switching lists. A larger list feels like the answer to "my list is thin here", and it usually is not: larger lists add breadth across categories they have already covered while giving the thin category a few more problems chosen by a general-purpose prior rather than by their specific gap. Targeted supplementation, chosen against an explicit sub-pattern map, is cheaper and lands where the deficit actually is.