The dict-of-dicts shape is the simplest trie that works: every node is a dict, keys are next characters, values are child dicts. A sentinel key ('__end__' here) marks 'a word ends at this node', which is what distinguishes a stored word from a mere prefix. setdefault is the one-liner for 'get or create the next node'. Insertion and lookup are both O(length) and the storage cost is one dict entry per character of input. This is the right starting trie for autocomplete prototypes, spell checkers, and the LeetCode 'Implement Trie' family.

The class-based form swaps the sentinel-key trick for an explicit is_word boolean per node, which reads more clearly and survives serialization libraries that mangle dict keys with leading underscores. __slots__ cuts the per-node memory footprint in half, which matters for million-word dictionaries (English has roughly 200K words; biology has millions of taxon names). The _descend helper is the shared walk used by both search and starts_with; deduplicating it is what keeps the public methods two lines each.

Autocomplete is the trie's killer feature: descend to the prefix node, then DFS the subtree to enumerate every stored word under it. Sorting children at each step gives lexicographic order; iterating node.children directly gives insertion order. The limit parameter and the early-exit checks inside walk keep the function bounded for huge dictionaries, since you almost never want all matches in a UI. Real-world autocomplete adds frequency weights to each node, then uses a heap during the walk to return the K most-frequent suggestions instead of the alphabetically-first K.