Fasaha (فصاحة) means eloquence and correctness of expression in Arabic — text that reads as if it was written by a native speaker, not translated. This skill fixes AI-generated or AI-translated Arabic so it reaches that bar.

Relationship to style/voice skills: this skill fixes structural/linguistic correctness only. For register and tone on top (MSA vs Darija, formality, terminology), fasaha now ships a self-contained profile at — generic, editable by any agent, updated as the user states preferences. Run first (make the Arabic correct and native-sounding), then apply the matching register profile second. If the user has a separate skill defining their personal tone/voice, that may be used instead of the built-in profile (or layered after it) — the built-in file is the default when no separate voice skill exists. Either way they compose — don't skip one because the other ran.

Grounded in: the QALB (Qatar Arabic Language Bank) annotation guidelines — the closest thing to a published standard for correcting AI/MT-generated Arabic — plus patterns specific to LLM translation failures that predate/postdate QALB's 2013 scope (see Section 6). Primary source: Zaghouani, Habash & Mohit, QALB Guidelines v0.90, CMU-Qatar/Columbia, 2013 — http://nlp.qatar.cmu.edu/qalb/QALB-guidelines_0.90.pdf. Full bibliography with links in .

Don't skip straight to a "polished" single-pass translation — the calque problem (Section 2) is much easier to catch by explicitly checking sentence boundaries against the source than by trying to write well the first time.

The single most common and most visible error. Any English word that isn't a proper noun, product name, or standard technical identifier must be translated into Arabic.

The most damaging error because it's invisible sentence-by-sentence but obvious paragraph-by-paragraph. Symptom: every Arabic sentence mirrors one English sentence, in the same order, with the same boundaries. Grammatically correct, but it reads as translated rather than composed.

Fix by restructuring, not by translating harder:

Generic-but-wrong translations that are grammatically valid but semantically off, especially for technical or domain terms.

Standard MSA grammatical correctness — gender/number agreement, broken plurals, missing obligatory prepositions, definite article consistency, redundant/missing words.

Mechanical but highly visible if wrong — these are the errors a native reader notices instantly.

AI output can slip into dialect, especially when the prompt, source material, or context contains dialectal Arabic. QALB Section 5.6 classifies dialectal words into six categories — knowing which category a word falls into determines whether it needs correcting to MSA at all.

When to flag instead of translate: if a passage is more than roughly a third dialectal, or uses a dialect the reviewer isn't confident in, don't attempt word-by-word MSA correction — flag it. Partial, uncertain correction of heavy dialect usually makes the result worse, not better. Full classification table and more examples: .

QALB's guidelines (2013) were built against older statistical/rule-based MT and human learner text — they predate LLM translation, so their error distribution assumes MT output that's often choppy or ungrammatical. LLM translations are usually grammatically clean but calqued (Section 2) and prone to Latin leakage in specifically technical/marketing vocabulary (Section 1) — errors QALB's annotators rarely saw. Don't assume QALB's category weights (spelling-heavy, from 2013-era MT) match what you'll actually find in modern LLM-generated Arabic — in practice, Sections 1-3 above catch more real issues in LLM output than Sections 4-5. Dialect leakage (Section 6) is more likely to come from dialectal source material or prompts than from the LLM itself, since modern LLMs default to MSA unless steered otherwise.