This Skill uses three independent roles to complete mathematical modeling. The generated paper is for the user's reference only and is not a work that can be submitted directly. The paper structure and format must be based on the official rules and official template of the target competition for the current edition; past experience must not be used to replace official requirements.
Before Use: Please first read the root directory 使用指南.md, clarify the usage boundaries of this Skill's deliverables, and copy it to the workspace (PROJECT_ROOT):
PowerShell(Windows):
Copy-Item "<SKILL_ROOT>\使用指南.md" "<PROJECT_ROOT>\使用指南.md"
bash(macOS/Linux):
cp "<SKILL_ROOT>/使用指南.md" "<PROJECT_ROOT>/使用指南.md"
<SKILL_ROOT>is the installation directory of this Skill (includingSKILL.mddirectory),<PROJECT_ROOT>is your problem/project directory.
Mandatory Execution Protocol
When the user explicitly names this Skill or the task matches this Skill, strictly execute the following protocol; do not downgrade it to a suggestion:
- In the first progress update, echo: this Skill has been activated,
SKILL_ROOT、PROJECT_ROOT, current stage, target competition and edition, and the role and tool entry points planned to be read. Unconfirmed official rules must be explicitly marked as pending verification. - Before starting each stage, actually read that role's
SKILL.md; when using PDF, Excel, paper search, DOCX, or LaTeX, then actually read the corresponding tool'sSKILL.md. Knowing the file path does not mean it has been executed. - Strictly call the scripts and templates provided by the Skill. When initialization, conversion, compilation, or validation tools already exist, it is forbidden to hand-write alternative implementations to save time.
- Treat any validation warning as incomplete. Only when the current edition's official rules or the user explicitly requires allowing deviation may you record the "rule source, deviation item, reason" and continue; you must not arbitrarily lower length, chart, formula, citation, or compilation quality targets.
- When the environment lacks an engine, search source, renderer, or dependency, report the blockage and continue completing the parts that can still be verified; it is forbidden to silently switch tools, skip validation, or pass off inferior artifacts as completed.
- When the runtime environment supports Subagents and the task will generate or modify fixed deliverables, dispatch independent quality inspection immediately at key nodes according to
references/Subagent调度.md; it is forbidden to wait until the entire workflow ends before dispatching for the first time, and author self-inspection cannot replace independent acceptance. Apart from fixed quality inspection, do not proactively dispatch other Subagents; only execute within the selected scope when the user explicitly enables specific collaboration tasks. - Before delivery, run all completion gates specified for the current role. If any command has not been run, the exit code is nonzero, independent acceptance has not passed, unresolved issues remain, or artifacts change after the gate, you must not claim "completed".
- For a real competition or other task with an explicit deadline, you must read
references/交付与截止时间协议.md, verify the current edition's official deadline and submission constraints, get the complete export pipeline working as early as possible, and maintain a Checkpoint V1 that can be submitted immediately; subsequent optimization must not break that version. - The final reply must list the entry points actually read, key commands actually run, exit codes, Subagent gate status, core quality metrics, and remaining blockers; do not merely say "checked".
Default Subagent Policy and Stage Gates
- Default mode dispatches only the
M1/P1/P2/W1/W2quality inspection Subagent; optional collaboration such as research, prototyping, and parallel experiments is disabled by default, and the workflow is not asked about or paused. - Users can enable one or more optional tasks in natural language and can disable them at any time; categories not explicitly selected remain disabled.
- The main Agent is responsible for scheduling and authoritative artifact writing; the quality-check Subagent must not have participated in writing or revising the artifact under review, and by default is read-only and returns evidence.
- The complete process passes through, in order,
M1modeling final check,P1minimum runnable result,P2programming final check,W1evidence outline, andW2paper final check.P1Only after passing can full computation and formal figure generation proceed,P2after passing, enter paper planning,W1and only after passing does long-form main text and dual-format typesetting begin. FAILThe corresponding role must make corrections based on the returned evidence and re-verify; when the artifact under review undergoes substantive changes, the relevantPASSimmediately becomes invalid.- A single-stage task runs only the corresponding gate. When the environment does not support Subagents, it can only be marked as
BLOCKEDor restricted delivery and must explicitly report "independent acceptance not completed"; the main Agent's self-check must not be described as independently passed or claimed as fully completed.
For detailed default switches, optional collaboration tasks, dispatch inputs, fixed receipts, and parallel boundaries, see references/Subagent调度.md。
root directory contract
SKILL_ROOT: the directory where this file is located, read-only. Role descriptions, algorithm materials, scripts, and templates are all read from here.PROJECT_ROOT: the directory where the user's problem and project are located; all new artifacts can only be written here.- The two root directories must be different; by default, overwriting any file in
SKILL_ROOTis prohibited. - Input attachments are read-only. If a template really needs to be modified, first copy it to
PROJECT_ROOTbefore handling it.
Routing
| User intent | Load entry point | Whether the previous stage is required to be completed |
|---|---|---|
| Complete modeling, complete the full problem | references/roles/建模手/SKILL.md → references/roles/编程手/SKILL.md → references/roles/论文手/SKILL.md |
Execute in order |
| Only do problem analysis and model selection | references/roles/建模手/SKILL.md |
No |
| Only write code, run results, and generate figures | references/roles/编程手/SKILL.md |
Requires the problem and executable model description; if missing, first fill in the necessary analysis |
| Only write or modify the paper | references/roles/论文手/SKILL.md |
Requires the problem, model, real run results, and figures; if missing, fall back to the corresponding stage |
Do not force the complete process in a single-stage task.
Fixed deliverables
Modeler
Deliver only:
题目分析报告.md术语表格.md
Programmer
Deliver:
- Runnable Python (
.py) or MATLAB (.m) code; both languages may be delivered at the same time. - Code run result tables, such as
.csv, the problem-required.xlsx。 - At least 3 candidate figures for each category of raw data figures, model run process figures, and model final result figures, totaling at least 9; it must also cover all subproblems of the problem, with at least 1 figure in each of the three categories for each subproblem. File names use formats such as
raw_q1_*、process_q1_*、result_q1_*, with no upper limit. SVG, PNG, and grayscale previews of the same content count as 1 logical figure. - Must include at least 1 overall modeling flowchart, covering input, subproblem dependencies, model, solving, verification, and output; when a subproblem has an independent branch, loop, or a solution chain of three or more steps, add the corresponding subproblem flowchart. File names use
flow_overall_model.*、flow_qN_model.*, counted separately from the three categories of candidate figures, and do not replace result figures. results/复现清单.json, including random seed, SHA-256 of input files, runtime and dependency versions, key parameters, and a unique reproduction command.
Paper Writer
By default, deliver a Word paper; generate a LaTeX paper only when the user explicitly requests it.
- Word:
完整论文.docx; when converting from LaTeX, also retain the hash-bound.conversion.json。 - LaTeX (optional):
完整论文-LaTeX/source project, the完整论文.pdfactually compiled from it, and the hash-bound.build.json。
When the user explicitly requests LaTeX, generate both formats simultaneously, keeping the main text content, data, figures, tables, and conclusions consistent. Internal inspection records are not additional deliverables.
Regardless of the target competition, the paper by default contains at least 8 formal figures; when the current official rules or the user's explicit requirements conflict with this, their requirements prevail and the basis is recorded.
Staged Feedback Loop
- When any Subagent gate returns
FAIL, return to the corresponding role based on the files, commands, numerical values, or source evidence in the receipt; after correction, rerun the affected gates, and the main Agent must not verbally override them. - When the Programmer finds that formulas are incomplete, constraints conflict, or the model is not implementable, stop guessing, list the evidence in the conversation, and return to the Modeler to correct the two Markdown deliverables.
- When the Paper Writer finds that key conclusions lack support from real results, figures, tables, or literature, return to the Programmer or Modeler to fill the gap; fabrication is prohibited.
- After correction, continue from the blocked stage and do not repeat stages that have already passed.
Modeling Constraints
- First fully understand the problem, attachments, objectives, constraints, and evaluation criteria, then form conclusions and a modeling plan.
- Each sub-problem may use at most two independent model systems. In physics problems, expansions of the same governing equation or the same physical mechanism under different approximation orders, boundary conditions, or precision levels count as one model family and must not be mechanically split into multiple models; only when the governing mechanism, state variables, or solution objectives are substantially different may they count as separate models. Auxiliary sensitivity analysis, error checks, and visualization do not count as independent models.
- Avoid directly applying overly common simple models to pass them off as innovation; innovation must come from the problem structure, data processing, constraint design, algorithm improvement, or validation method, and its basis must be explained.
- Data judgment criteria are determined by the problem, official rules, domain literature, or data analysis results.
- Do not forcibly delete one of two models merely because they reach similar conclusions; whether to retain it is determined by the research purpose and evidentiary value.
Progressive Loading
First read the current stage's SKILL.md, then read the required references according to the "when to load" table therein; loading all materials at once is prohibited.
| Current Task | Additional Reading |
|---|---|
| Selecting a model or looking up an algorithm | references/算法索引.md, then read one or a few related assets/*.md |
| Searching for papers | tools/paper_search/SKILL.md |
| Reading the problem PDF | tools/pdf/SKILL.md |
| Processing Excel | tools/xlsx/SKILL.md |
| Drawing figures | tools/figure/SKILL.md |
| Generating a Word paper | tools/docx/SKILL.md |
| Generating a LaTeX paper | tools/latex/SKILL.md |
| Converting a LaTeX paper to Word | tools/docx/SKILL.md |
| Dispatching Subagents or stage quality checks | references/Subagent调度.md |
| Real competition, deadline, or final submission | references/交付与截止时间协议.md |
For complete navigation, see references/README.md。
Completion Determination
- First satisfy the "Mandatory Execution Protocol" and provide reviewable gate results.
- When claiming full completion, all independent Subagent gates involved in the current task must be
PASS, and no substantive change has occurred in the post-pass output without re-verification; when the Subagent capability is lacking, only reportBLOCKEDor limited delivery. - All computational conclusions come from actual run results.
- Formulas, tables, and figures are consistent with the code results.
- References can be traced via OpenAlex, AnySearch, or the original publication page.
- The paper has been configured according to the official rules of the target competition for the current edition, the length target has been confirmed, the number of formulas/non-empty figures and the figure coverage of all subproblems have been checked, figure numbering and in-text citations are continuous, and references and in-text citations correspond bidirectionally. The Word branch must also pass native OMML, DOCX structure, conversion warning gating, and rendered page count checks; the LaTeX branch must also pass environment diagnostics, real compilation, logs, authoritative resource-source-PDF hash binding, body and appendix page counts, blank pages, page size, font embedding, and image DPI checks.
- All artifacts are located in
PROJECT_ROOT,SKILL_ROOThas not been rewritten.