I assume you mean "electronic dictionary", and I assume infinite programming skills, plus access to recording facilities and speakers.
Such a dictionary needs to be paired with a grammar (and, obviously, a grammatical analysis), and an annotated text corpus of some sort. What a dictionary does is distill out the essential unpredictable properties of individual lexical items. The English word "cat" is pretty simple in terms of form – the singular and plural are completely predictable (read the grammar), there are no case forms (read the grammar). "Child" is a bit more complicated so you have to list the plural. Since this dictionary is not exclusively about word forms, you may also assemble idiomatic usages of "cat" including special meanings in compounds, for example "catwalk", or "child" (as in "child node"). Consult the OED for zillions of examples of that type.
In the complete dictionary of Logoori, you would list the (formal) causative of "buy" because it actually means "sell", and is semantically unpredictable compares to the formal causative of "plow" which is "make a person plow". There would be information on transitivity and stativity because "boil" doesn't distinguish "I boiled the water" vs. "The water boiled" or "The water is boiled", and different roots interact with the tense system differently to yield an interpretation – you need to include that kind of information. You need to indicate how the "perfective" stem is formed since that is not entirely predicable from the usual citation forms. It would take years to enumerate all of the forms that would have to be listed in the perfect dictionary, so I'll cut this short.
Semantic information is particularly difficult, since 1 or 2 word English translations are usually very inadequate. Pairing lexical entries with actual usage (from the corpus) is ideal, and why I stipulated electronic and programming skills. Providing all of the desired information for the root -gáraba is TMI (pages and pages for "bean leaf"), but with an electronic resource you can structure the presentation of information, suppressing the dozen corpus examples until it's relevant.
Usually, dictionaries are organized around the language equivalent of "noun, adjective, verb" (the major lexical classes), with "preposition" having an uneasy place (can't you just put them in the grammar? Yes in Bantu usually, no in English). One does not ordinarily list grammatical morphemes in a dictionary, for example "mʊ-. Cl. 1 [sg], plural usu. Cl. 2". But in an electronic dictionary this can also be listed, and arguably should be available (revealed when you click "more").
As a toy example, you could start to sketch a Swahili dictionary, because there are already such dictionaries, grammars, and speakers, and the level of complexity of the language is relatively low (minimal arbitrariness, not as ludicrously prolific in word-formation as some Bantu languages or the languages of the Eskimo family).
The usual problem in dictionary construction, in terms of knowing what to include, is that you do not know what to include. I recently learned that "e" in Turkish spelling is ambiguous between two vowels, so a dictionary needs to indicate whether that vowel is open mid or close mid. If you leave that information out, you have a less-than-perfect dictionary. If you have links to pronunciations by speakers, you at least in principle have supplied the necessary information (which the Turkish government dictionary does). Most dictionaries of Bantu languages are less than perfect because they omit tone. So there is no general and precise answer to the question "what should I include", but the direction of the answer is "Anything that cannot be provided by rule", and "anything needed to prove the analysis. With my proposed toy dictionary, you can discover in one instance what should be included.