Revision research: what evidence says about effective revision
Last updated: 2026-05-17
Synthesis
Section titled “Synthesis”The composition-studies literature on revision is older and more robust than most computational writing systems acknowledge. The central finding, first articulated by Sommers in 1980 and refined by Faigley & Witte in 1981, is that experienced writers and novice writers revise differently in kind, not just in degree. Novice writers treat revision as cleanup: surface fixes, word swaps, proofreading-shaped activity. Experienced writers treat revision as re-seeing the work: they revisit structure, argument, and meaning across multiple passes, and they treat each pass as having a distinct purpose rather than attempting to fix everything at once.
This has direct implications for an agentic writing system. A skill that promises to “improve the paper” combines what experienced writers do across many discrete passes — and produces the characteristic novice-revision artifact: superficial change without structural improvement, or structural change without local polish. Scriptorium’s one-skill-one-responsibility discipline is the agentic analog of the experienced-writer practice of single-purpose passes.
The further implication is that revision is iterative and
non-linear. The “outline-then-write-then-edit” linear model that
guides most novice writers is not how experienced writers actually
work. Scriptorium should not lock authors into a single ordering of
skill invocations; the orchestrator (manuscript-pipeline, v0.2)
should support re-entry into earlier critique passes after
transformation passes, because that is what working revision looks
like.
Evidence
Section titled “Evidence”Sommers (1980), College Composition and Communication 31(4): 378–388. “Revision Strategies of Student Writers and Experienced Adult Writers.” A landmark empirical study: eight freshman students and seven experienced adult writers (journalists, editors, academics) wrote three compositions, rewrote each two times, suggested revisions for an anonymous-author composition, and were interviewed three times. Key findings:
- Student writers operated with implicit, uncodified revision procedures, predominantly word-level — reaching for the thesaurus, polishing diction, fixing surface errors.
- Experienced writers operated with codified principles. They characterized revision as a recursive activity — “finding what I want to say” rather than “fixing what I already said.”
- Experienced writers made larger, deeper changes throughout the writing process, not at the end. Sommers’s framing: revision is not an editing stage; it is the writing process itself.
Sommers’s contribution is that she demonstrated, empirically, that experts and novices conceptualize the purpose of revision differently. The expert’s revision is closer to what Einsohn (see copyediting-vs-developmental) would call developmental editing applied to one’s own draft.1
Faigley & Witte (1981), College Composition and Communication 32(4):400–414. “Analyzing Revision.” Built on Sommers by introducing a taxonomy of revision changes that has become canonical in composition research:
- Surface changes — formal and meaning-preserving edits (spelling, punctuation, tense, abbreviation, word-form, substitutions, reorderings).
- Meaning changes — additions, deletions, substitutions, and reorderings of content that change what the text says.
Meaning changes further split into microstructural (sentence- or local-paragraph-level meaning shifts) and macrostructural (thesis-level, argument-level, organization-level meaning shifts).2
The Faigley & Witte taxonomy maps neatly onto editing levels (see copyediting-vs-developmental):
| Faigley & Witte | Editing level | Scriptorium category |
|---|---|---|
| Surface | Copyediting | Normalization |
| Microstructural meaning | Line editing | Transformation (narrow) |
| Macrostructural meaning | Developmental | Transformation (broad) |
Faigley & Witte’s contribution to scriptorium specifically: the
taxonomy provides a vocabulary for describing what a transformation
skill did. A semantic diff produced by compression should be
classifiable as predominantly surface or microstructural; one
produced by argumentative-flow should be classifiable as
macrostructural. If a “compression” skill is producing macrostructural
changes, the skill is doing the wrong job.
Multi-pass discipline in expert practice. The follow-on composition-studies literature consistently observes that experienced writers separate concerns across passes. The argument for separation is not aesthetic; it is cognitive. Trying to fix structure, argument, prose, and citation at once produces shallow attention to each. Writing-process research summarized in Hayes & Flower (1980) and extended in Hayes (2012) frames revision as a planning-translation- reviewing cycle that re-enters itself rather than a terminal proofread.3
The outline-then-write vs. draft-then-outline debate. Practitioner writing on academic composition (notably G. Whitesides’ essay on writing papers, Advanced Materials 2004) advocates an outline-first discipline: write the outline, get reviewer agreement on the outline, then write the prose. This pushes macrostructural work earlier — the argument is settled before any sentence has the inertia of being already-written. Whitesides’s specific recommendation: the outline should be on the order of one page per figure, with each section’s purpose explicit before any paragraph is composed.4
The counter-position — draft-then-outline — is associated with writers who prefer to discover their argument by writing it. Both work; the failure mode that scriptorium must guard against is neither — drafting without an outline, then editing without a re-outlining pass. That produces local fluency over global incoherence, and it is the failure mode LLMs are most prone to amplify because they are good at local fluency.
How this informs scriptorium
Section titled “How this informs scriptorium”Passes have purposes. Each scriptorium skill targets one of Faigley & Witte’s categories. Authors invoking multiple skills should understand the sequence as a multi-pass revision in the expert-writer sense, not as a single “improve this” instruction.
The orchestrator (v0.2) must support re-entry. Sommers and the
follow-on literature show that expert revision is recursive: a
developmental change at iteration 3 may require returning to a
copyediting concern. manuscript-pipeline should not enforce a
strict linear sequence; it should sequence intelligently and allow
re-entry. The state file’s phase field (draft, review, revision,
submission) tracks where the manuscript is overall, but does not
constrain skill ordering within a phase.
Surface vs. meaning diffs in skill output. When a transformation
skill modifies prose, its semantic-diff output should classify the
change as surface, microstructural, or macrostructural. This makes
the conservative-edit posture inspectable: a compression skill that
declares its changes are surface-and-microstructural-only can be
verified against its own output.
The “Remaining Weaknesses” section is the expert-revision instinct. Experienced writers know what they did not fix and can articulate it. Skills should do the same — not because the LLM has expert judgment, but because the discipline of articulating what was not addressed is the same discipline that distinguishes expert from novice revision.
Don’t bundle. A single skill that does compression and argumentative-flow and citation-cleanup is, in revision-research terms, a novice-revision skill: it tries to do everything at once and is unlikely to do any one thing deeply. The scriptorium pattern explicitly prevents this bundling, and the rationale is the four decades of composition-studies evidence on what bundling produces.
Open questions / weak evidence
Section titled “Open questions / weak evidence”- Most revision research is on student writing or general English composition. Scientific writing — with its rigid section structure, high citation density, and quantitative claims — has been less studied as a revision context.
- Whether multi-pass discipline transfers to AI-assisted writing the same way it transfers to unassisted writing is empirically open. Plausible but not established.
- Sommers and Faigley & Witte are from a pre-word-processor era. Some of their observations about how writers physically revise (longhand, retyping) are dated; the conceptual taxonomy is not.
References
Section titled “References”Footnotes
Section titled “Footnotes”-
Sommers N. Revision strategies of student writers and experienced adult writers. College Composition and Communication. 1980;31(4):378–388. doi:10.2307/356588 (also indexed via the NCTE archive: doi:10.58680/ccc198015930). ↩
-
Faigley L, Witte S. Analyzing revision. College Composition and Communication. 1981;32(4):400–414. doi:10.2307/356602 (NCTE archive: doi:10.58680/ccc198115887). ↩
-
Hayes JR, Flower LS. Identifying the organization of writing processes. In: Gregg LW, Steinberg ER, eds. Cognitive Processes in Writing. Erlbaum; 1980:3–30. Updated framework in Hayes JR. Modeling and remodeling writing. Written Communication. 2012;29(3):369–388. doi:10.1177/0741088312451260. ↩
-
Whitesides GM. Whitesides’ group: Writing a paper. Advanced Materials. 2004;16(15):1375–1377. doi:10.1002/adma.200400767. ↩