Research writing
Write the strongest paper the supplied evidence can support. The goal is not a plausible paper-shaped document. It is an argument whose facts, claims, and limits survive inspection.
Preserve the author's research choices and voice. Improve the framing when the evidence permits it, but do not silently replace the author's scientific priorities with your own. Venue instructions and an author-provided writing sample outrank the stylistic defaults in this skill; neither can override the evidence rules.
Work in Lattice
Lattice ships this skill disabled by default; the user enables it in Settings → Skills. Do not change skill preferences on the user's behalf to activate this workflow.
- Read the selected text and its surrounding context before editing, and follow the project's main LaTeX file and
\inputor\includestructure rather than assuming the open file is the entire paper. - Preserve existing macros, math environments, labels, citation keys, comments, and unrelated edits. Make targeted edits to the requested files instead of replacing the project with a new manuscript.
- Use the paper library's supplied paths to inspect sources.
.research/papers/<id>/paper.mdis cached paper text;blog.mdis a secondary explanation, not the original paper. Neither a search result nor a bibliography entry proves that a paper supports a claim. - Use Lattice's available literature and bibliography tools for adding, upgrading, or removing references.
Never directly write
.bibfiles or fabricate citation keys; use the keys returned by the tools. - After editing LaTeX, use the available project build tools to check compilation and unresolved citations or cross-references. Report unavailable checks and remaining errors honestly.
- Keep audit notes out of manuscript files unless requested. If editing files, report the changed locations and material evidence gaps instead of duplicating the whole manuscript in chat.
Choose the mode
- Author-draft revision: edit the researcher's own target-manuscript prose while preserving its authorship, argument, and scientific content. Use this when provenance or an authorship-sensitive review matters.
- Packet to paper: formulate a complete paper from results, notes, figures, sources, and other research artifacts.
- Section drafting: write one or more sections while maintaining the whole paper's claim and terminology ledgers.
- Restructure: repair the argument or section order of an existing draft.
- Revision: improve accuracy, scope, clarity, and prose without changing the research record.
- Audit: report unsupported claims, missing evidence, structural failures, and generic model prose before changing text.
For ordinary paper authoring, read these references as their phase becomes relevant rather than loading all of them at once:
references/evidence-and-claims.mdbefore formulating claims.references/paper-architecture.mdbefore outlining or drafting sections.references/revision-and-style.mdbefore drafting or revising prose.references/research-judgment.mdonly when the task also asks for research framing, experiment critique, or a choice among defensible paper stories.
Return the requested artifact
Put the user's requested manuscript, section, revision, outline, or audit first. Keep private evidence ledgers and diagnostic labels out of manuscript prose unless the user asks to see them.
- For drafting, return manuscript-ready prose followed only by unresolved evidence gaps that materially limit it.
- For revision, return the revised text first. Add a concise change record when the user requests one, a change affects a claim, or provenance-sensitive author-draft revision makes generated material or altered certainty relevant.
- For audit, report findings with locations, evidence consequences, and the smallest safe repair. Do not silently rewrite unless the user also asks for a revision.
- For a blocked task, use the evidence-gap format below instead of producing a plausible completion.
Start with an evidence gate
Normalize the supplied material into four categories before writing:
- direct observations and results;
- the author's interpretations;
- claims about prior work, each with a source;
- assumptions, open questions, and missing evidence.
Then build a claim-evidence ledger using the contract in
references/evidence-and-claims.md. A complete paper normally has one to three
umbrella claims. This is a default, not a quota. Each claim needs a precise
scope, supporting evidence, live alternatives, and evidence that would falsify
or weaken it.
Keep two working layers. The private audit layer contains source locators, evidence and claim IDs, statuses, wording ceilings, and unresolved gaps. The prose-facing research record contains the scientific objects, methods, equations, settings, values, citations, negative results, and limits, without the audit vocabulary or a prewritten narrative. Draft from the research record; verify against the audit layer. Do not make a reader follow the evidence-control system used to keep the paper accurate.
Stop when any missing item would change a central conclusion, comparison, causal interpretation, or citation. Return this instead of filling the gap:
Status: BLOCKED
Available evidence:
- ...
Evidence gaps:
| Gap | Claim or section blocked | Why inference is unsafe | Minimum input needed |
Safe work possible now:
- evidence ledger entries that are already supported
- provisional claim candidates marked untested
- a non-narrative outline with explicit gaps
If the gaps are local and non-blocking, continue with a working draft and mark
them as [EVIDENCE GAP: ...]. Never make the prose sound final while its
central evidence is provisional.
Preserve provenance in author-draft revision
An author-provided style sample and an author-written target draft are different inputs. Another researcher's published paper may inform general writing judgment, but it cannot supply authorship for a new manuscript. Use this mode only when the researcher supplies prose they wrote for the target work and has authorized its revision.
Before editing:
- freeze the original draft and record its hash;
- extract its claims, comparisons, numbers, units, citations, quotations, equations, tables, named entities, uncertainty, and limitations;
- identify requested changes and passages that are already correct;
- mark missing sections or arguments that would require new prose.
When the original and revised drafts are available as files, run
uv run --no-project python <skill-directory>/scripts/audit_revision.py ORIGINAL REVISED after the prose pass, resolving the script relative to this skill's installed directory.
If the Python runner is unavailable, perform the comparison manually and report that the script was not run; do not install dependencies without permission.
Treat its protected-span differences as review signals, not automatic errors.
A PASS only means the script found no differences in the spans it recognizes; it can miss unit changes, reassigned citations, and some LaTeX math.
The bidirectional semantic audit below remains required.
Edit the smallest unit that solves a real scientific or reader problem. Preserve the author's section and paragraph architecture, terminology, and sentence choices unless one of them causes that problem. Do not rewrite a correct paragraph merely to regularize style. Protect evidence-bearing spans during a prose pass and restore them exactly unless the evidence audit authorizes a change.
Keep a revision ledger with each changed span classified as correction,
clarification, restructure, or generated. If a missing passage must be
written from the research record, disclose it as generated instead of blending
it into the author-written draft. When provenance is the acceptance criterion,
return the gap and ask the author to draft it before editing.
Finish with two independent walks: map every original claim and qualification into the revision, then map every revised factual statement back to the original draft or research record. Report dropped claims, new claims, altered certainty, and generated spans. Writing in the statistical style of human papers does not change generated prose into human-authored prose.
Formulate the paper
1. Establish the reader and paper type
Identify the expert reader, venue constraints, contribution type, and the decision or understanding the paper should change. Distinguish method papers, empirical findings, theory, systems, datasets, replications, negative results, surveys, and position papers. Do not force all of them into one narrative.
2. Find the defensible story
Look for the actual relation among problem, evidence, and contribution. A useful story may turn on an observed contradiction, an under-constrained problem, a testable hypothesis, a unifying representation, a failure taxonomy, or a boundary exposed by negative evidence. Use a tension only when the research record contains it. Do not manufacture drama around an ordinary improvement.
When two materially different stories are defensible, show their claims, evidence, and tradeoffs and ask the author to choose. In an explicitly autonomous run, choose the story with the strongest evidence coverage and record the rejected alternative in the working notes.
3. Build the argument before prose
Create:
- the claim-evidence ledger;
- a claim dependency graph;
- a terminology and notation ledger;
- a prose-facing research record that removes IDs, statuses, wording ceilings, and drafting advice while preserving every conclusion-changing fact;
- an outline in which every section answers a reader question;
- a figure and table plan in which every item supports a claim;
- a working abstract used to expose missing logic, not as final copy.
Each outline unit must name its claim IDs, evidence IDs, strongest alternative, and boundary. Delete a section that serves no claim or reader need.
Privately distinguish evidence by its role: explanatory evidence makes the scientific object intelligible; decisive evidence most directly changes the central claim; corroborating evidence tests recurrence or breadth; boundary evidence changes how a result must be read. These roles are not manuscript labels. Develop explanatory and decisive evidence, combine corroborating tests by the question they answer, and place a boundary beside the inference it changes.
4. Draft evidence-bearing sections first
Draft Methods and Results before polishing the Introduction. This forces the paper's promises to match the work that was actually done. For each result, state the question, setup, comparison, observation, interpretation, and limit. Tell the reader which value, line, point, or qualitative case carries the conclusion. Report losses, null results, and metric disagreements beside the wins they qualify.
Draft each paragraph around one scientific pressure point: the observation,
failure, comparison, derivation, or decision the reader must understand next.
Give unequal findings unequal space, and do not force neighboring concepts into
matched clauses or a miniature-paper template. Use
references/revision-and-style.md for the paragraph and sentence passes.
5. Draft the remaining sections
Use references/paper-architecture.md for section contracts. Related Work
compares assumptions, mechanisms, or evidence, rather than summarizing papers
one at a time. Discussion distinguishes observation, interpretation, and
speculation. Limitations state which claim each limitation narrows and what
evidence would resolve it.
6. Run an argument-hierarchy pass
Before polishing, write down the paper's central answer, the decisive test that supports it, and the single boundary that most changes how it should be read. Use those three items to revise the whole draft:
- let the Introduction connect the concrete problem or question to the central answer or contribution and its decisive evidence; the exact order follows the paper type, but the full inventory of controls and secondary analyses belongs in Methods or Results;
- map every occurrence of each major claim boundary. Preserve the first necessary local qualification and one section-appropriate synthesis; keep any later occurrence only when it adds a distinct consequence, alternative, or test;
- after reporting a result, explain only the next non-obvious inference; do not paraphrase a measurement and then enumerate every claim it excludes;
- classify every experiment as a central test, mechanism probe, robustness check, or practical boundary. Report all of them, but reserve repeated cross-section emphasis for the central test and the most claim-changing boundary;
- organize Discussion around the inferences that several results jointly support, not around another walk through the Results section.
- after provenance is verified, write about the study rather than the packet: state which experiment or comparator exists or is missing, and mention how evidence was supplied only when its provenance affects validity or reproducibility.
This pass changes rhetorical prominence, never the research record. It must not remove mixed or negative evidence, detach a necessary qualifier from its claim, or make a section falsely self-contained by repeating the entire audit trail.
7. Rewrite the final abstract and title
After the evidence-bearing sections stabilize, rewrite the abstract from scratch. State the achievement, the precise problem or difficulty, the approach, the strongest evidence including a number when one is central, and the relevant boundary. Set the title last. It should identify the subject and real distinction without claiming more than the paper proves.
8. Run global audits
Audit in this order:
- provenance of every fact, number, quotation, and citation;
- support and scope of every empirical or theoretical claim;
- separation of observation, interpretation, and speculation;
- agreement among title, abstract, introduction, results, and conclusion;
- terminology, notation, metrics, populations, and comparison directions;
- figure, table, caption, and cross-reference integrity;
- reader comprehension and prose;
- venue and formatting requirements.
Correctness audits precede style. A style pass may not change a number, citation, claim polarity, causal strength, population, or uncertainty without returning to the evidence ledger.
Non-negotiable invariants
- Do not invent facts, results, citations, quotations, methods, experiments, or a discovery history.
- Every central claim must be falsifiable, scoped, and linked to evidence.
- Untested claims remain questions or hypotheses, never abstract or conclusion findings.
- Causal language requires causal evidence. Association is not intervention.
- Negative, mixed, and failed results cannot disappear for narrative ease.
- Prior work must be represented from sources, not titles, metadata, or memory alone.
- Technical terms do not change merely to avoid repetition.
- Figures and tables are evidence, not decoration; the prose also states the decisive comparison.
- Missing evidence stays visible. Fluency is never a substitute for support.
- Style rules are heuristics. Scientific meaning, author voice, and venue conventions take priority over punctuation, sentence-length, or vocabulary preferences.