Pinned-Paragraph Rewriting
A long-form editor that lets writers pin the passages they like while the rest of the page iterates toward a complete draft.
When drafting long-form work, many writers first arrive at a barely complete draft, then get stuck between local edits and rewriting the whole piece. This editor uses a diffusion language model to process the entire page at once: after users set the topic, tone, audience, and length, the page converges through several iterations instead of generating slowly from left to right.
When users reach an opening, transition, or conclusion they like, they pin that paragraph directly. The next generation round rewrites only the unlocked areas, with change intensity at paragraph boundaries showing where substantial revision is still happening. If they want to preserve an argument but try a different structure, they can lock the sentence meaning so the model changes only organization and wording.
The editor retains a full-page snapshot and paragraph-level diff for every round. If the structure in an intermediate version is better, users can retrieve one of its paragraphs and continue iterating the rest. Writers can therefore see an article evolve from scattered ideas into a coherent structure, rather than confronting an untraceable long draft after generation ends.
The first version focuses on drafting, rewriting, and structural exploration from material the writer already has. It does not research facts on the writer’s behalf or present unverified content as reliable source material. It is for writers who want to assess the whole piece as they go, turning a full rewrite into a series of reversible choices with individual sections frozen in place.
Why now
On August 20, the DiffusionGemma Technical Report entered discussion on Hacker News. S2 Its parallel refinement of 256-token text blocks creates a new opening for iterative, full-page rewriting; as of August 21, the post ranked 14th with 131 points and 34 comments. S1S2
Target user
Long-form writers who already have source material but are stuck on structural rewrites. This often happens when a newsletter, essay, proposal, or chapter draft has just taken shape: they have passages worth keeping, but neither want to patch the piece paragraph by paragraph nor risk scrapping it entirely. Paragraph pinning turns global exploration into small, reversible choices, especially for writers comparing alternative structures.
Minimal entry point
The generation layer can use DiffusionGemma’s official open weights directly, with block diffusion inference as its foundation. S1 The document layer uses a structured paragraph tree that stores a text lock, meaning lock, and rewrite state for each paragraph. Precisely locked spans are excluded from noising. Meaning locks are first converted into claim constraints, then checked for consistency after each round. The interface shows the current full page, change intensity, and an iteration timeline. Snapshots are stored incrementally by paragraph hash, with single-paragraph retrieval from earlier rounds. The first release limits document length and concurrent rounds to test whether writers use locking repeatedly.
Punching above its weight
Initial users can come from long-form writing communities, independent newsletter writers, and AI-writing forums. Promotional material should show one messy draft converging over successive rounds rather than list features. Let writers upload an old draft and publicly demonstrate how the rest changes after the opening is pinned. Shareable round-comparison pages can turn real revision processes into distribution. A demo released around the current DiffusionGemma discussion could attract developer-minded writers willing to test a new interaction model. S2
Competitors & gaps
- SudowriteGoogle
- Sudowrite already covers ideation, planning, writing, and editing. Its Story Bible retains characters, worldbuilding, and outlines, while Rewrite adjusts existing text to instructions. S3 That workflow works well for moving from source material to chapters and can constrain later generation. Its core interaction, however, is still to select text and ask the model to rewrite it. Public documentation does not show a way to pin any paragraph while the rest of the page converges over multiple rounds, nor does it show change intensity for areas still undergoing major revision. The opportunity is structural exploration, not another rewrite button. Pinning, unpinning, and retrieving an earlier paragraph must be frictionless, or users will return to familiar local rewrites.
- NovelcrafterGoogle
- Novelcrafter already includes scene content, summaries, Codex entries, and prompts in its revision history. Users can view and restore any earlier version. S4 These capabilities suit long-form material management and provide a safety net for ambitious edits. Its restoration unit is primarily a field or scene version. Public materials do not show a visual merge that pulls a single paragraph back across versions, nor do they show the text’s round-by-round evolution during generation. A paragraph-pinning product could turn version management into real-time writing control. The key is not saving more copies, but letting writers specify what cannot move before each round. If that control still requires complicated prompting, existing revision history may already be sufficient.
How it makes money
Monthly subscription with a fixed allowance of full-page iterations. Additional credits are sold per generation round, avoiding opaque pricing for complex model costs.
The case against
Freezing arbitrary paragraphs changes the model’s original sampling path. The more lock points there are, the more likely paragraph boundaries are to break in reference, tone, or argument. S1 Meaning locks are harder still: the system must first extract the claims that cannot change, then determine whether the rewrite remains faithful. A mistaken judgment can make writers think an idea was preserved when its meaning has actually drifted. Multi-round full-page generation also increases GPU memory, queueing, and snapshot-storage costs. If change intensity flickers too much, it can disrupt reading. Losing one critical qualifier could be enough for users to stop submitting important long-form work.