Andrew Stellman regularly uses AI for writing, but not because he believes it writes well. On the contrary, he has made the model’s consistent poor prose a deliberate part of his workflow. Instead of starting from an outline he now brainstorms and dumps ideas into an AI, collects notes and experiments, and only later asks the model to draft an article. The AI-generated draft is almost always terrible — and that is useful: the obviously machine-produced paragraphs act as placeholders he can rearrange, then replace with human writing.
What is ‘Yoda-speak’?
Stellman identifies a recurring oddity he calls “Yoda-speak,” where generated sentences have awkward, sometimes reversed word order or fragmented structure — reminiscent of the Star Wars character Yoda’s syntax. He cites an example from his current manuscript: “The eight pond sentences fill the left region, every word clickable.” That phrasing is unusual and hard to imagine in ordinary speech. Claude Opus also used the term “Yoda-speak” when Stellman asked it to help decipher similarly awkward sentences.
Another example, from a later chapter outline about the attention mechanism, reads: “What’s left is to let the two heads’ work combine, so run the combined eight numbers through one more small table, eight in and eight out, square, called the output projection.” The sentence conveys something recognizable, but its comma-laden, shuffled structure makes it difficult to parse.
One word at a time: autoregression
Stellman reminds readers that large language models (LLMs) are next-token predictors: at each step the model predicts the next token, appends it to the response, and proceeds. Once a word is generated it becomes part of the context; the model generally cannot go back and reorder previously produced tokens. If the model emits a noun early and only later brings up descriptive details, those details often end up tacked onto the end of the sentence, producing the kind of awkward phrasing we see in Yoda-speak.
He also cites a 2024 paper by Wilson Wu, John X. Morris, and Lionel Levine, titled “Do language models plan ahead for future tokens?” which finds evidence that models sometimes prepare information useful later, but even so, anticipation is not the same as being able to revise already emitted text.
The “better” second answer phenomenon
Stellman notes another common behavior: after producing a sentence the model often immediately supplies a reformulation and indicates the later version is better. He hypothesizes that this is related — the model cannot change what it already wrote, so it writes another sentence that it considers superior. Self-correction and Yoda-speak may be two faces of the same autoregressive limitation.
An experiment with Astra and Luna
To test whether prompt form influences these outputs, Stellman ran a small experiment using OpenAI’s Astra as an orchestrator, which spun up multiple Luna sub-agents. Each sub-agent received nearly identical prompts with small tweaks. Astra asked the sub-agents to flag grammatical constructions similar to Yoda-speak and to iterate the prompt to increase the effect.
In the first simple test, a Luna agent received facts about a bicycle written as full sentences (color, narrow frame, silver bell, worn leather seat; it leans beside the garage; Rina rides it to work and locks it). Luna produced a normal, natural paragraph. Reordering those same full-sentence facts still produced tidy output.
Next, Stellman switched to a made-up device, the “feed throat,” with facts given as full sentences. Again, Luna generated a clean relative-clause formulation: “The feed throat is a tapered steel slot, a hand’s width across, through which a clay tile slides.”
Then Astra imposed two rules: the output had to be a single sentence and it had to begin with a fixed opening phrase that ended with the noun “slot” (for example, “Slide the clay tile through one more steel slot”). When the three descriptive facts were provided as full sentences, Luna still tended to form a relative clause after the noun: “Slide the clay tile through one more steel slot, which is a hand’s width across, tapered, and called the feed throat.”
Crucially, when the same facts were supplied as short fragments rather than complete sentences ("a hand’s width across tapered called the feed throat"), the model appended those fragments in order after the noun: “Slide the clay tile through one more steel slot, a hand’s width across, tapered, called the feed throat.” The only change between runs was the shape of the facts in the prompt.
A similar pattern appeared with another example: “Rest the cracked bowl inside one more padded cradle, two palms wide, crescent-shaped, called the handling saddle.” Once the model had committed to the noun and the descriptors arrived as loose pieces, it tacked them on in sequence.
Interpretation and limits
Stellman concludes that the model copies the shape of its input: attention is sensitive not only to content but to how facts are expressed. If facts come as full sentences, the model tends to integrate them into relative clauses; if they come as fragments, the model reproduces fragments appended after the noun. Because the model is autoregressive it cannot go back to insert details before already-generated tokens.
He acknowledges limits: his view is external — he cannot see internal token weightings or exact attention distributions, nor can he prove that the model didn’t already “have in mind” the descriptors before emitting the noun. Still, the observed outputs align with a hypothesis that interaction between attention (what the model focuses on in the prompt) and autoregression (inability to revise emitted tokens) produces Yoda-like phrasing and the common “here’s a better version” behavior.
Why this matters
Understanding these mechanisms helps explain why machine-generated prose sometimes looks oddly ordered and how prompt structure can influence output form. For writers and developers, awareness of the phenomenon suggests practical strategies: craft prompts with consistent sentence structure when you want integrated phrasing, or expect fragmentary appending if descriptors are supplied as loose fragments. Stellman’s pragmatic takeaway is twofold: you can use the model’s bad writing as a drafting aid, and you can influence generation by shaping the prompt's form.
Short summary
Yoda-speak may arise when attention attends to the prompt’s format and the autoregressive generation process prevents the model from reordering tokens already produced. As a result, descriptors that arrive too late or in fragmentary form are often appended after the noun, producing unusual word order; the model’s habit of offering later alternatives can be a separate symptom of the same limitation.
