This longitudinal study examined how the phonemic-orthographic context affects the spelling of the schwa (/ɨ/) by Portuguese beginning spellers at two time points in the first school grade. The schwa is phonetically unstable and phonologically ill-defined, has an unpredictable realization, is frequently deleted at the syllable's end, and is often spelt as <e>, a very high frequency grapheme with numerous phonological renditions. In addition to cognitive and other alphabetic tasks, 41 first graders were asked to spell 40 consistent words of medium-low frequency: 5 CV.CV (consonant, vowel. consonant, vowel) with well-articulated vowels; 10 C/ɨ/C.VC, the first vowel being a schwa, thereby creating potential phonological consonantal clusters, half legal (/fɨliʃ/, /fliʃ/), half illegal (/pɨdal/, /pdal/); 10 CV.C/ɨ/, the last vowel being a schwa, potentially creating phonological monosyllables half with a legal coda (/mɔlɨ/, /mɔl/) and half with an illegal coda (/n'avɨ/), (/nav/); in addition, the children spelt 15 CVC ending with /l/, /ɾ/ and /ʃ/, the only legal Portuguese codas. Participants were also asked to spell equivalent pseudowords at a second point in time. Our results show that children were sensitive to allowable letter patterns from the Time 1 assessment point. Although alphabetic spelling was not entirely mastered, children used <e> more in first syllables than at the end of the word, and more in illegal than in legal phonological consonantal clusters, although the pattern of significant differences did change over time. The results were similar for pseudowords. Also, children used <e> more at the CV.C/ɨ/ words whose last C was /l/, than in monosyllabic CVC words ending with /l/. This was not observed with pseudowords, where the grapheme <e> was used with a similar frequency in the two types of items. Overall, these results show that children's acquisition of this kind of context-conditioned orthographic knowledge occurs simultaneously with alphabetic letter-sound learning and depend largely on intuitive statistical learning reflecting the regularities of the written code to which they are exposed.