Structuring research workflows

Structuring Research Workflows

Overview

Turn computer vision research material into two source-grounded text deliverables: a detailed workflow and a copy-ready prompt for drawing the same workflow as a polished academic figure. Support the full vision research spectrum rather than assuming a particular CNN, Transformer, 2D/3D representation, or learning paradigm. Keep both deliverables structurally identical so every visual module can be traced back to the written method.

Required References

Read both files on every invocation:

• references/workflow-template.md for extracting and writing the detailed workflow.
• references/diagram-prompt.md for constructing the complete diagram-generation prompt.

Output Contract

Return exactly these two top-level sections in the user's language:

1. 详细科研流程
2. 完整流程图生成提示词

Place the complete diagram prompt in one fenced text block so it can be pasted directly into ChatGPT or another image-generation system.

Generate text only. Do not call imagegen, create an image, or add a third deliverable unless the user explicitly makes a separate request for an actual image.

Workflow

1. Inspect all supplied material: prose, papers, equations, code, configs, logs, tables, and reference figures.
2. Establish the source of truth. Prefer executable code and active configuration for an implementation diagram; prefer the paper for a paper-method diagram. State material conflicts instead of silently merging them.
3. Build a method graph containing inputs, preprocessing, ordered core stages, branches, joins, objectives, outputs, evaluation, and training-only or inference-only paths.
4. Record exact names, shapes, units, formulas, parameters, and directionality where supported. Mark absent values as 未提供 or not specified; never invent them.
5. Write 详细科研流程 using references/workflow-template.md.
6. Convert the same method graph into 完整流程图生成提示词 using references/diagram-prompt.md.
7. Cross-check one-to-one coverage: every written stage appears in the diagram prompt, every visual module is supported by the written workflow, and every branch reconnects correctly.

Computer Vision Adaptation

Choose visual primitives from the actual vision task instead of forcing every method into a 3D CNN or generic tensor pipeline:

Vision research content	Preferred visual primitives
Classification, detection, segmentation, or keypoint tasks	images, labels, boxes, masks, heatmaps, feature maps, pyramids, prediction heads
Restoration, enhancement, super-resolution, or inverse imaging	degraded/clean image pairs, degradation models, multi-scale features, residual paths, reconstructed outputs
Video understanding, action recognition, or tracking	frame sequences, temporal axes, clips, trajectories, memory states, track identities
Depth, pose, neural rendering, point cloud, or 3D vision	cameras, rays, depth maps, point clouds, voxels, meshes, projections, coordinate frames
Vision Transformers and representation learning	patches, tokens, embeddings, attention or state-space blocks, multi-scale representations
Self-supervised, semi-supervised, weakly supervised, or domain-adaptive learning	multi-view augmentations, labeled/unlabeled paths, pseudo-labels, consistency links, source/target domains
Vision-language and multimodal learning	image encoders, text tokens, shared embeddings, cross-modal alignment, prompts, retrieval or generation outputs
Generative vision and diffusion models	latent variables, noise levels, denoising trajectories, conditions, samplers, generated images
Remote sensing, medical imaging, and spectral imaging	modality-specific images, image volumes, spectral cubes, acquisition channels, annotations, clinical or geographic outputs
Compression, acceleration, and deployment	teacher/student paths when present, pruning or quantization blocks, hardware targets, latency, memory, throughput
Classical vision and geometry	keypoints, descriptors, matches, transforms, optimization loops, geometric constraints

Use task-specific visual symbols only when supported by the source. Do not add a CNN, Transformer, attention module, distillation path, or 3D tensor merely because it is common in computer vision.

Accuracy Rules

• Preserve formulas, dimensions, units, stage order, and train/eval distinctions exactly.
• Separate main flow from auxiliary branches with distinct line styles.
• Explain ambiguous or conflicting evidence before the two deliverables when clarification is essential; otherwise list the assumption inside 详细科研流程.
• Do not claim a method is faithful to a paper when it is an adaptation.
• Do not add fashionable modules, metrics, datasets, or experimental details absent from the source.
• Do not use this skill as a generic workflow formatter for research outside computer vision.

Example Invocation

Use $structuring-research-workflows to analyze the attached computer vision paper and implementation config. Treat the active config as the source of truth, explain any paper-versus-code differences, then return the detailed workflow and the complete publication-style diagram prompt only.

Visual Rules That Must Survive Every Prompt

• Use a clean white academic canvas and a single restrained title line.
• Never place a blue banner, colored ribbon, badge, star decoration, slogan, or subtitle strip beneath the title.
• Permit only thin hairline rules beside the title when useful.
• Use restrained semantic colors, fine borders, coherent arrows, and task-appropriate computer vision objects.
• Reserve at least 6% safe margin on every side and at least 10% at the bottom.
• Keep all content inside the central 88% of the canvas; shrink or reflow modules instead of cropping.
• Require a final text, topology, and cropping checklist in the diagram prompt.

Final Check

Before responding, verify:

• There are exactly two top-level output sections.
• The detailed workflow is source-grounded and complete.
• The diagram prompt is self-contained and copy-ready.
• The title area contains no colored banner or promotional decoration.
• The prompt adapts across broad computer vision tasks without defaulting to one architecture family.
• No actual image-generation tool was invoked.