Working Report
Write a self-contained report of what the study establishes and what remains open. Make it suitable to share with an advisor or collaborator. Combine concise prose, equations, data plots, and explanatory figures so the reader can follow the argument with little reconstruction. Let the explanation and evidence determine its length.
Set the Scope and Read the Sources
Default to English and LaTeX. Follow the user's requested language and format, including Markdown. Determine the research question, current stage, report scope, available materials, and target file from the conversation and workspace. Ask about missing information that would change the report. Keep edits in the requested report and its supporting files; a separate publication manuscript needs its own editing request.
Read the relevant report, project instructions, current code and tests, experimental records, results, and literature. Use implementation and tests as evidence for software behavior, raw data or validated analyses for empirical findings, and primary literature for borrowed claims. Establish novelty against verified prior work and performance against appropriate measurements. Keep the source paths or identifiers needed to recheck important statements.
Write from available evidence. Run additional analyses or experiments within the user's requested scope. Describe pending research by its purpose and current status.
Separate the Report From Agent Notes
The report and its supplementary information (SI) contain the science: motivation, necessary foundations, formulation, methods, setup, results, interpretation, limitations, hypotheses, and planned investigations.
Keep execution history in one agent-facing Markdown file. Reuse a suitable existing file; otherwise use
beside the report. Record significant attempts, errors, fixes, decisions, unresolved execution issues, evidence paths, and commands or settings needed to resume work. Keep a short current-state summary and dated execution entries. Keep this file outside the reader-facing deliverables.
Include scientifically meaningful negative results and rejected hypotheses in the report. State the method, assumptions, and conditions each finding concerns. If an error invalidates evidence, revise the report to state what the remaining valid evidence supports. Put the debugging and correction history in the agent notes.
Structure the Report
Choose sections and their order to fit the study. Explain why the problem matters, what is modeled or measured, how the method follows from the formulation, what was done, what the evidence shows, and which questions remain.
Plan the figures alongside the section outline. Identify the comparisons, relationships, mechanisms, and unfamiliar concepts that become easier to understand when shown. Give each visual a clear question to answer and place it near the explanation or result that uses it. Keep a point in prose when a short sentence conveys it more directly. Choose visual coverage by reader need, not by a figure quota.
Use concise, specific noun phrases for section headings at every level. Name the scientific topic, method, quantity, or comparison, and keep headings at the same level grammatically parallel. Put explanatory questions, full-sentence claims, and qualifications in the opening paragraph. Keep headings short enough to scan in the table of contents while retaining the terms that distinguish each section.
Write connected paragraphs with one main point each. Explain why a concept, equation, figure, or experiment is needed before using it. State the reasoning that connects successive steps. Use one term for each concept and one meaning for each symbol within its scope.
Explain the scientific mechanism and why each methodological choice fits the problem. Distinguish implementation facts, mathematical deductions, assumptions, approximations, experimental observations, empirical choices, and interpretations. State the conditions under which each conclusion holds.
Explain Concepts in Context
Assume secondary-school mathematics and science unless the user gives a different baseline. Define unfamiliar concepts and abbreviations at first use. Connect each new concept to quantities, observations, or relationships already introduced. Pair a concise definition with a simple explanatory figure when geometry, dependencies, transformations, or a concrete example make its meaning easier to grasp. Aim for enough understanding to follow and assess the study.
When
is available, read it and use its prerequisite analysis and first-principles method. Use the known reader baseline and report objective. Apply the teaching method while composing the complete report.
Write natural, precise academic prose throughout the report and SI. Use specific claims, clear sentences, and transitions that follow the actual argument. In Chinese reports, use established disciplinary terminology and verify uncertain translations in authoritative sources. Retain the English term with a concise Chinese explanation when no established Chinese rendering exists. Remove generic praise, inflated language, empty emphasis, stock transitions, and repetitive summaries. Integrate explanations where the argument needs them. Reader exercises and comprehension checks belong only in explicitly requested conversation.
For each formula or theorem, explain what it says, when it applies, how it is used here, and what the result means. Aim for understanding and correct use. Include proofs or detailed derivations when the user requests them or when they support the study's own theoretical contribution.
Explain simple tools in the main text. For a complex tool, give enough explanation in the main text to understand and use it in this study. Put the fuller explanation in a dedicated SI section and link to it at the point of use. Write the SI for the same reader baseline. Explain the tool's meaning, conditions, and application so the reader can understand it without searching for missing foundations. Use subsections when helpful.
Define each symbol and its object type at first use. State what every index labels and its range. Distinguish physical and dimensionless quantities, sets, operators, and random variables. Default to italic scalars, bold lowercase vectors, and bold uppercase matrices and higher-order tensors; explain departures at first use. Use upright type for standard group, operator, and function names and descriptive subscripts. In LaTeX, define recurring mathematical commands in one shared preamble.
Give units and the shapes and axis meanings of non-scalar quantities. Check dimensional consistency in equations, losses, and normalizations. Specify summation indices and ranges, integration variables and domains, and the measures or probability distributions used in integrals and expectations. State relevant normalization, cutoffs, smoothing, numerical regularization, and boundary or initial conditions. Retain every coefficient, exponent, and constraint needed to reproduce the method. For implemented methods, match mathematical objects to their tensor layouts and explain any mapping between them. State the group and transformation law when making symmetry claims, and distinguish invariance from equivariance.
Use the user's questions to identify missing definitions or reasoning steps. During a requested revision, repair those gaps in the report and update affected passages. Keep each explanation proportional to its role in the study.
Make Research Status Visible
Present scientific findings and open research questions in the main text. Report verified findings within their established scope in ordinary prose. Use these inline flags for unresolved content:
| Flag | Meaning and required context |
|---|
| [Hypothesis] | A proposed explanation or claim awaiting a discriminating test. State its basis and what evidence would bear on it. |
| [Preliminary] | A finding supported by current evidence whose reliability or claimed scope still requires a specific check. State the evidence and the pending check. |
| [TODO] | Research work still to perform, such as a derivation, comparison, measurement, or analysis. State what it should establish. |
Use a bold, bracketed prefix at the start of the passage, table entry, or figure statement it qualifies. In LaTeX, use
\ReportStatus{Hypothesis}
,
\ReportStatus{Preliminary}
, or
from the shared style file; in Markdown, use
,
, or
. Keep the following prose in normal type and explain the labels once if needed. Distinguish completed work from planned work, and state uncertainty beside the claim it affects. Keep execution fixes and agent reminders in the agent notes.
Present Evidence and Concepts Visually
For each result, state what it supports and what remains open. Give the setup, comparison, metric, units, uncertainty, and assumptions needed to assess it. Use plots to expose trends, distributions, differences, convergence, and uncertainty that prose or a list of numbers obscures. Use tables for exact values or compact repeated comparisons. Generate quantitative visuals from traceable data or stated calculations, retain relevant variability, and make axis scales and comparison conditions explicit.
Use explanatory figures to show the objects and relationships needed to understand the science: a configuration, a mechanism, a method's information flow, or the action of a mathematical operation. Choose the simplest accurate representation and label the quantities and relationships in the report's terminology. Distinguish schematic illustrations and hypothetical examples from measured or computed results. Show assumptions or omitted dimensions when they affect interpretation; an illustrative arrow alone does not establish causation.
Use figures to show comparisons and relationships, and use the nearby prose to explain their meaning, evidence, and limits. Remove passages that merely repeat every label or data point. Write self-contained captions that identify the panels, symbols, conditions, and main observation; make each table understandable from its title, body, and notes. Keep labels short, layouts uncluttered, and explanatory detail in the caption or surrounding text.
Draw publication-ready figures in the clean, precise style of Nature and PRL articles. Apply one visual style to data plots and concept diagrams throughout the report and SI, including figures made with different tools. Reuse shared style settings for fonts, type sizes, line weights, spacing, panel labels, and color assignments. Give the same quantity or category the same visual meaning across figures. Use clear layouts, crisp lines, restrained color, and adequate white space; use shapes, line styles, or markers as well as color where distinctions matter.
Default to Seaborn's
palette for discrete data series and conceptual categories. Use a sequential palette for continuous magnitudes and a diverging palette for values around a meaningful midpoint. Use the existing plotting environment and install
there when needed. Preserve the specified fonts and layout when applying its palettes or themes.
Use Times New Roman for figure text and compatible fonts for mathematical symbols. Keep all text inside figures in English, including axes, legends, and annotations; use Chinese captions in Chinese reports. Size labels, legends, and annotations to look approximately as large as the body text after insertion at the intended width. Check exported figures at that size and verify their actual fonts. Keep editable sources. Export report figures as PDF and prefer those files in the report; keep PNGs or other formats as useful previews or auxiliary outputs. Preserve vector text and lines where supported and retain adequate resolution for embedded raster content.
Label panels with bold, upright lowercase letters
,
,
, and so on, without parentheses. Place each label at the upper-left of its panel with consistent size and spacing, clear of plotted content and axis labels. Order panels from left to right and top to bottom, and describe them in the same order in the caption.
Add verified references to the bibliography and cite them beside the claims they support. Check the bibliographic metadata and read the source to confirm its support. Reuse suitable existing entries. Use
when the writing task requires help understanding a paper. Keep unverified reference candidates in the agent notes.
Maintain a Coherent Current Version
Make each version of the report and SI a standalone account of the current study. Place new results where they belong in the argument. Replace outdated statements. Search the main text, SI, captions, tables, and equations for affected uses; update terminology, symbols, definitions, assumptions, methods, numbers, summaries, and conclusions together. Adjust headings, transitions, and cross-references to keep the report consistent from beginning to end.
When a question or task is resolved, replace its flagged passage with the supported result or method. Remove outdated flags, caveats, and pending-work statements everywhere they appear. Keep only genuinely unresolved research questions open.
Order the report by scientific reasoning and experimental conditions. Keep the history of discovery, writing, debugging, and revision in the agent notes. Describe physical time evolution, experimental sequences, and algorithmic iterations when they are part of the science being studied.
Preserve source artifacts and valid user-authored content. Keep edits within the agreed scope. Record the evidence behind changed conclusions in the agent notes.
Format and Verify
Select the LaTeX template by report language:
for English and
for Chinese. English defaults to
with
;
is available for an explicitly requested arXiv layout. Chinese uses
with
and XeLaTeX. Read
references/latex-style.md
before preparing LaTeX and copy the selected template's supporting files. Follow the user's explicit format requirements and preserve compatible project organization. Apply the selected contents, heading, caption, and table styles consistently throughout the report and SI. Keep contents entries aligned by level, continuation lines aligned with their titles, and page numbers aligned at the right margin.
Use
for division in units throughout the report and SI, including equations, figure labels and captions, and table headings, cells, and notes. Keep denominator exponents positive and group multiple denominator factors in parentheses, as in
,
, and
.
For a new report, keep the selected entry point named
at the project root. Keep metadata and chapter inputs there; put chapter content in separate
files under
. Split large chapters into subfiles or subdirectories as needed. Create
for compilation and formatter output, and copy the finished PDF to
beside
. Include
in every report directory, including directories that are not Git repositories, to exclude temporary files while keeping report sources and the final PDF available for version control.
Format LaTeX with
, enabling
and line-break modification. Keep each complete prose sentence on one source line, preserve paragraph breaks, and indent structured environments consistently. Disable fixed-width wrapping. Check that formatting preserves the content and LaTeX semantics.
Compile LaTeX with the available project toolchain and inspect the PDF. Check scientific consistency, references and citations, equations and symbols, figures and tables, and readability from the stated baseline. Review text-heavy passages for a comparison or relationship that a figure would make clearer, then revise the figure and prose together. Inspect every figure at its final insertion size for readable labels, correct fonts, consistent styling, overlap, clipping, and an accurate correspondence with its caption and evidence. Fix build and presentation errors within scope. If compilation is unavailable, deliver the source and state that the rendered output is unverified.
Review Independently and Deliver
By default, ask a fresh, independent subagent to review the report and SI before delivery. Give it the research goal, reader baseline, source files, rendered output, and supporting evidence. Let it assess the work directly from those materials.
Ask the reviewer to check:
- Section hierarchy and order, continuity of the argument, and consistency between the main text and SI.
- Technical correctness, assumptions, methods, results, citations, and whether the evidence supports each conclusion.
- The current scientific contribution, supporting evidence, originality relative to verified prior work, and remaining validation. Identify the further research needed for a future publication while preserving the working report's explanatory depth.
- Natural academic language, precise terminology, clear explanations, and unnecessary or formulaic wording.
- Visual coverage of important data and difficult concepts, scientific accuracy of diagrams, and consistent, readable figure styling at the final page size.
Request specific, prioritized findings with locations and reasons. Check each finding against the evidence and make supported corrections throughout the affected parts. Rebuild and inspect changed output, then have the reviewer recheck substantive fixes. Continue until material issues that can be resolved within the writing task are addressed. Carry remaining research needs into the delivered report as limitations or TODOs. If subagents are unavailable, perform a self-review and state that independent review remains unperformed.
Deliver the report, any SI, and rendered PDFs when available. Briefly state what changed, what was checked, and what remains unresolved. Share or submit the report only when the user requests it.