Dictionary Lookup Skill

A dictionary lookup sounds like something a language model shouldn't need — until the question is "is this actually a word," "what's the IPA," or "give me a citable source," where fetched beats recalled. The Free Dictionary API answers over keyless HTTPS with definitions, phonetics, audio, and origins. This skill fetches, formats by part of speech, and keeps a clean line between what the source says and what the model adds — because a dictionary answer's whole value is knowing which is which.

What This Skill Produces

• The entry — definitions grouped by part of speech, with example sentences where the source has them
• Pronunciation — IPA text and the audio-file link when available
• Synonyms/antonyms — from the source, supplemented (and labeled) by the model when thin
• The command — exact curl, rerunnable; and an honest miss-report when the API lacks the word

Required Inputs

Ask for these if not provided:

• The word — and the sense if context suggests one ("mean" the verb, the adjective, or the statistic?)
• What they actually need — a quick meaning, pronunciation, etymology, or a citable check that a word exists — the output leads with it
• Language note — this API is English-only; other languages get an honest redirect (Wiktionary manually) rather than a fake answer

Framework: Fetch, Format, and the Honesty Line

1. The call: curl -s "https://api.dictionaryapi.dev/api/v2/entries/en/serendipity" → JSON array: phonetic/phonetics[] (IPA + .audio mp3 links), meanings[] grouped by partOfSpeech, each with definitions[] (with example sometimes), synonyms, antonyms, plus origin on some entries.
2. Lead with the asked-for thing: a pronunciation question gets IPA in line one, not after four noun senses; an is-it-a-word check gets yes/no plus the entry as proof.
3. The miss is information: a 404 means this API lacks the word — common for new coinages, technical jargon, and inflected forms. Report exactly that, try the lemma (running → run), and if the model still knows the word, define it clearly labeled as model knowledge, not the fetched source. Never dress recall as a citation.
4. Sense selection over sense-dumping: when context implies a sense, lead with it and compress the others to one line each; "mean" has a dozen senses and the user wanted one.
5. Supplement with labels: the API's synonym lists are often thin — model-added synonyms go under "additional (model-suggested)," keeping the source's authority intact for the part that claims it.

Output Format

[word] · [IPA] [· audio link]

[part of speech] — [the leading sense, with example] [Other senses, compressed · other parts of speech]

Origin: [when present in source or asked — labeled if model-supplied] Synonyms: [source list] [· additional (model-suggested): …]

Source: Free Dictionary API · rerun: [exact curl] [On a 404: "not in this dictionary — tried '[lemma]'; the definition below is model knowledge, not a fetched source"]

Quality Checks

• The asked-for element (meaning/IPA/origin/existence) leads the answer
• Definitions are grouped by part of speech, contextual sense first
• Source content and model additions are visibly separated
• 404s report the miss and try the lemma before falling back
• The rerunnable curl appears

Anti-Patterns

• Do not present model recall as the fetched entry — the label line is the skill's integrity
• Do not dump every sense undifferentiated when context picked one
• Do not fake non-English lookups on an English-only API — redirect honestly
• Do not treat a 404 as "not a word" — it's "not in this dictionary," a much smaller claim
• Do not skip pronunciation when the question was spoken-word-shaped (names, presentations, ESL contexts)