I've been conducting a systematic evaluation of Sudowrite's "Expand" feature for several weeks now, treating it as a CI/CD pipeline for narrative generation. A consistent, critical failure mode has emerged: the engine's propensity for premature narrative termination. It doesn't simply add detail; it often attempts to architect a full story arc conclusion within a few sentences, effectively collapsing the dramatic tension you're trying to build. This isn't a minor bug—it's a fundamental architectural flaw in its scene-modeling logic.
Consider this input prompt I used to test the boundary conditions:
> *Kaelen stood at the edge of the phosphorescent fungus grove, the map clutched in his hand. The symbols didn't match. He could hear the distant drip of water, and something else—a low, rhythmic scraping.*
A desirable expansion would deepen sensory details, perhaps explore his internal dilemma about the map, or build unease around the scraping sound. Instead, Sudowrite frequently produces output like:
> *...the scraping grew louder, resolving into the form of a massive, rock-hided creature. Kaelen realized the map was a decoy. With a surge of adrenaline, he turned and fled the caves, vowing never to return. The mystery of the fungus grove remained unsolved, a lesson in trust.*
This is a catastrophic narrative compression. It has:
* Introduced and resolved a threat in one paragraph.
* Assigned a motive (decoy) and a character decision (fleeing) without development.
* Imposed a thematic closure ("a lesson in trust") that wasn't in the prompt.
From an infra perspective, this suggests the AI's context window management is prioritizing "narrative completeness" over "scaffolding." It's running a `terraform apply` with a pre-baked, generic `output.tf` that defines an ending, rather than letting the user iteratively build the state.
My current workflow mitigation involves a combination of:
* **Strict Prompt Engineering:** I now preface expands with imperative constraints. E.g., "Expand this scene **only** by adding sensory detail and internal monologue. Do **not** introduce new plot points, resolve tension, or end the scene."
* **Incremental Chunking:** Treating the "Expand" function like a poorly documented API. I apply it to a single sentence at a time, not a full paragraph, to reduce its latitude for narrative hijacking.
* **Rollback and Fork:** Using version history extensively. If an expansion goes off the rails, I revert and try a different, more constrained prompt, forking the narrative state.
The core requirement for a tool like this is to act as a stateless generator of *potentialities*, not a stateful orchestrator of resolutions. Sudowrite's current model seems trained to "finish the thought" in a way that aligns with common story beats, which is precisely what a writer using expand does *not* want. We need a toggle or a fundamental retraining: "Expand Mode" vs. "Conclude Mode." Without this, the feature's utility in building complex, nuanced scenes is severely compromised.
Has anyone else developed a more robust pipeline or systematic constraints to work around this? I'm considering feeding it outputs with explicit YAML-style directives, as if it were a configuration management tool.
--from the trenches
infrastructure is code