TokenMix Research Lab · 2026-07-03

Research AI Input Pack: What to Prepare Before Using GPT, Claude, Gemini, or DeepSeek

Most academic AI failures do not start with the model.

They start with the input.

Researchers often ask a model to do everything in one giant chat:

• summarize these papers,
• write the related work,
• polish this section,
• check citations,
• reply to reviewers,
• debug the reproduction code,
• explain this MATLAB error,
• and generate a workflow diagram.

That can work for small tasks. It breaks down when the task has many moving parts. The model may write fluent sentences, but miss the research boundary, overstate evidence, invent links between papers, or ignore important details in code, tables, and reviewer comments.

Before choosing GPT, Claude, Gemini, DeepSeek, Kimi, Qwen, or GLM, build a research AI input pack.

An input pack is a structured bundle of context that tells the model:

• what the project is,
• what evidence exists,
• what stage the work is in,
• what should not be changed,
• what output format you need,
• and what risks must be checked.

This article gives you a reusable template.

You Can Use AI to Build the Input Pack First

When researchers hear “input pack,” they often assume it means more manual preparation.

That is not the point.

The purpose of an input pack is to turn scattered context into a structure that models can understand. AI can help you build that structure. In fact, for many research workflows, the first AI task should not be “write the paper.” It should be “organize my materials into an input pack.”

A reliable sequence is:

Step 1: Rough input
Paste what you already have: title, abstract, advisor notes, reading notes, PDF summaries, lab notes, error logs, code snippets, reviewer comments.

Step 2: Ask AI to build the input pack
Ask the model to extract, classify, normalize, and ask missing questions. Do not ask for final prose yet.

Step 3: Human verification
You verify the extracted fields, correct mistakes, add missing context, and only then move to writing, revision, reproduction, or rebuttal.

In other words, the input pack is not a purely manual step before AI. It can be the first step of the AI workflow.

Copyable prompt: turn scattered material into an input pack

Use this prompt when your context is messy:

I will give you scattered research materials.

Do not write the paper yet. Do not produce the final conclusion.

Your task is to organize the materials into a Research AI Input Pack.

Return the following structure:
1. Project card: topic, field, research object, core question, scope, current stage, target output;
2. Materials inventory: what materials I provided and what each material can be used for;
3. Key evidence: what can be treated as evidence and what is only a hypothesis or idea;
4. Missing information: what is still needed before writing / analysis / reproduction;
5. Risk notes: where the model may invent citations, overgeneralize, or misread experimental results;
6. Suggested next step: literature map, paper outline, citation check, reviewer response, or reproduction debugging.

Rules:
- Do not invent information I did not provide.
- Mark uncertainty explicitly.
- If a field is missing, ask me a question instead of guessing.
- Make the output structured so I can copy it into the next prompt.

Scattered materials:
[paste materials]

Tell the model what not to do

If you only say “organize this,” the model may turn your notes into polished but premature prose.

Add constraints like these:

Only organize the materials. Do not write final paragraphs.
Only extract information I provided. Do not add non-existent references.
List missing fields first. Do not assume you already know them.
Separate evidence, speculation, and items that need confirmation.

Different raw materials produce different input packs

For example, if you have ten PDFs and messy reading notes, do not immediately ask AI to write related work. First ask it to generate a literature list and topic clusters. Then use that structured output in a second step to draft the related work. This is usually much more reliable than asking for a final section in one shot.

The Rule: Prepare Context Before Choosing a Model

Model choice still matters.

Claude is often strong for long documents and academic prose. Gemini is useful for PDFs, charts, screenshots, and multimodal materials. GPT is strong for planning, outlines, and structured writing. DeepSeek and Kimi can be useful for code, reproduction analysis, and engineering tasks. Qwen and GLM can help with Chinese academic expression, localization, and visual reasoning.

But no model can reliably infer missing context.

If you only provide a paper title and ask for a literature review, the output will be generic. If you paste reviewer comments without the revised manuscript, the response may sound polite but fail to answer the actual criticism. If you paste MATLAB code without the expected output, the model may fix syntax while missing the mathematical goal.

Use this sequence instead:

1. Define the task.
2. Build the input pack.
3. Choose the workflow or Skill.
4. Choose the model.
5. Run a second-model check for high-risk outputs.

Quick Overview: 8 Input Packs

The rest of this article gives copyable templates for each pack.

1. Project Card

Use a project card before any serious research task. It prevents the model from guessing your field, scope, and target output.

Project Card

Topic:
Field / discipline:
Research object:
Core research question:
Keywords:
Scope:
Out of scope:
Current stage:
Target output:
Target venue or audience:
Deadline:

Constraints:
- Do not invent citations.
- Separate evidence from speculation.
- Mark uncertain claims clearly.
- Ask if key context is missing.

When to use it

Use the project card when you are:

• entering a new field,
• defining a research question,
• asking for a literature map,
• planning a paper outline,
• or deciding which model or Skill to use.

Example prompt

Here is my project card.

[paste project card]

Please help me turn this into:
1. a clearer research question,
2. 3 possible paper angles,
3. a literature search strategy,
4. and a list of missing context I should prepare.

Do not write the paper yet.

2. Literature List

Do not only upload PDFs. Build a small literature list first, even if it is rough.

Literature List

Paper 1
- Title:
- DOI / arXiv:
- Year:
- Field:
- Method:
- Dataset / material:
- Main finding:
- Limitation:
- Why it matters:
- Which section it may support:

Paper 2
- Title:
- DOI / arXiv:
- Year:
- Field:
- Method:
- Dataset / material:
- Main finding:
- Limitation:
- Why it matters:
- Which section it may support:

When to use it

Use this pack for:

• literature review,
• related work,
• systematic review,
• research gap mapping,
• and reading-list prioritization.

Recommended workflow

1. Use GPT to refine the research question.
2. Use Claude to synthesize long paper notes.
3. Use Gemini for PDFs, figures, tables, and screenshots.
4. Ask for a literature map, not a paragraph draft first.

Example:

Use the literature list below.

Task:
Build a literature map for my related work section.

Return:
1. topic clusters,
2. method clusters,
3. agreements and disagreements,
4. missing evidence,
5. which papers support each subsection,
6. risks of overclaiming.

Literature list:
[paste list]

3. Manuscript Draft

When asking AI to revise a paper, do not paste a whole draft and say "polish this."

Separate the manuscript into sections and define the revision target.

Manuscript Draft Pack

Paper title:
Paper type:
Target venue or style:
Target audience:

Abstract:
[paste]

Introduction:
[paste]

Methods:
[paste]

Results:
[paste]

Discussion:
[paste]

Current task:
- outline only / revision / abstract / conclusion / style polishing

Revision target:
- clarity
- logic
- claim-evidence alignment
- academic tone
- shorter length
- stronger contribution statement

Do not change:
- terminology:
- equations:
- variable names:
- citations:
- numerical results:

Good prompt

Use this manuscript draft pack.

Task:
Revise the Introduction for logic and claim-evidence alignment.

Return:
1. diagnosis,
2. revised version,
3. change log,
4. remaining risks,
5. claims that need stronger citations.

Do not invent citations.
Do not change technical terms unless you explain why.

4. Citation Table

Citation checking is not the same as proofreading.

The model needs to know which claim each citation is supposed to support.

Citation Table

| Claim in manuscript | Citation | What the source actually supports | Support strength | Risk |
|---|---|---|---|---|
| [claim 1] | [Author, Year] | [evidence] | strong / medium / weak | [risk] |
| [claim 2] | [Author, Year] | [evidence] | strong / medium / weak | [risk] |

When to use it

Use this pack before:

• submission,
• rebuttal,
• major revision,
• or converting a literature review into a tighter related work section.

Good prompt

Check the citation table below.

Return:
1. unsupported claims,
2. overgeneralized claims,
3. citations that may be misplaced,
4. claims that need additional evidence,
5. safer rewritten versions.

Do not add fake references.

5. Reviewer Package

Reviewer response is one of the easiest places to misuse AI.

If you only paste reviewer comments, the model will produce a polite answer. It may not check whether the manuscript was actually changed.

Prepare this pack instead:

Reviewer Package

Target journal / conference:
Decision type:
Revision deadline:

Reviewer comment 1:
[paste]

Current response draft:
[paste]

Revised manuscript excerpt:
[paste]

What we changed:
[paste]

What we did not change:
[paste]

Reason:
[paste]

Tone requirement:
- respectful
- specific
- concise
- evidence-based

Good prompt

Use this reviewer package.

Check:
1. whether each reviewer concern is answered,
2. whether the response points to a real manuscript change,
3. whether the tone is respectful,
4. whether the response overpromises,
5. whether any unresolved issue should be admitted.

Return a table:
Concern / response quality / manuscript evidence / risk / suggested revision.

6. Reproduction Package

For paper reproduction, the problem is often not "explain this paper."

The problem is a narrow gap:

• Which dataset split was used?
• Which checkpoint matches Table 2?
• Why does the README default differ from the paper?
• Which preprocessing step changes the metric?

Prepare a reproduction package:

Reproduction Package

Paper:
Repository:
Branch / commit:
README section:
Config file:
Environment:
Dataset:
Metric:
Expected result:
Actual result:
Error message:

Known conflict:
- Paper says:
- README says:
- Config says:
- Code default says:

Question:
[one narrow reproduction question]

Good prompt

Use this reproduction package.

Question:
Which dataset split and preprocessing settings most likely correspond to Table 2?

Return:
1. evidence from the paper,
2. evidence from README/config/code,
3. conflict table,
4. most likely setting,
5. uncertainty,
6. next experiment to run.

For this workflow, a useful model pairing is:

• Claude for long paper and README context,
• DeepSeek for config and metric reasoning,
• Kimi for longer repository reading.

7. MATLAB / Simulation Package

For MATLAB tasks, describe the mathematical goal before the code.

MATLAB / Simulation Package

Goal:
System / model:
Given equations:
Variables:
Assumptions:
Current MATLAB code:
Error message:
Expected output:
Actual output:

What I need:
- symbolic derivation
- transfer function
- state-space form
- MATLAB function draft
- debug explanation
- validation checks

Good prompt

I am working on a MATLAB simulation.

Use the package below.

Please:
1. restate the mathematical goal,
2. define variables,
3. derive the transfer function,
4. convert it to state-space form,
5. point out assumptions,
6. draft MATLAB code,
7. list checks I should run in MATLAB.

Package:
[paste]

Do not ask the model to "fix MATLAB" without equations, expected output, and error details.

8. Figure Package

Academic image generation should not start with "make it look nice."

Start with purpose, structure, and constraints.

Figure Package

Figure purpose:
Audience:
Canvas:
Style:
Required elements:
Required text labels:
Flow direction:
Color constraints:
Forbidden elements:
What must be accurate:
What can be illustrative:
Export use:
- paper draft
- lab meeting
- slide deck
- blog / teaching card

Good prompt

Create a clean academic workflow diagram.

Canvas: 16:9 slide.
Style: white background, minimal colors, readable labels.

Elements:
1. data collection
2. preprocessing
3. model training
4. evaluation
5. error analysis

Use arrows to show direction.
Keep text labels short.
Do not add logos, QR codes, decorative text, or extra claims.

For research figures, treat AI-generated images as drafts. Redraw final paper figures in vector tools if precision matters.

A Practical Multi-Model Workflow

Here is a stable workflow for academic work:

Step 1: Project card
Use GPT to clarify the task and output structure.

Step 2: Literature list
Use Claude for synthesis and Gemini for PDFs / charts.

Step 3: Draft or revision
Use Claude for long sections and GPT for structure.

Step 4: Citation and claim check
Use a citation table. Ask a second model to challenge weak claims.

Step 5: Reviewer response
Use Claude Opus or another strong reasoning model for tone, strategy, and evidence.

Step 6: Code / reproduction / MATLAB
Use DeepSeek or Kimi when code, config, or repository context matters.

Step 7: Final check
Ask the model to list remaining uncertainty instead of pretending the answer is final.

Cost and Subscription Note

Researchers often use AI in bursts.

During a deadline week, you may use long-context models, coding models, PDF understanding, and image generation in the same day. During data collection or experiments, you may barely use them.

That is why many research workflows are better served by a multi-model, pay-as-you-go setup instead of monthly subscriptions to every platform.

With a unified multi-model API gateway such as TokenMix, a lab or individual researcher can route different steps to GPT, Claude, Gemini, DeepSeek, Qwen, Kimi, GLM, and image models from one place. The important part is not only saving money. It is having one workflow where model choice follows the task.

Related TokenMix resources:

• For the broader model division framework, read Academic Research AI Workflow.
• To check current model availability, use the model catalog.
• To estimate cost before a research sprint, use the pricing page.
• To connect the workflow to code, see the OpenAI-compatible API guide.

Common Mistakes

Mistake 1: Asking for a full paper too early

Ask for a research map, outline, or claim table first. Drafting should come after evidence is organized.

Mistake 2: Uploading PDFs without a goal

Tell the model what to extract: methods, datasets, limitations, debates, or evidence gaps.

Mistake 3: Treating polishing as revision

Polishing improves sentences. Revision checks logic, contribution, evidence, and structure.

Mistake 4: Asking one model to be the final judge

For high-risk claims, use a second model to challenge the answer.

Mistake 5: Letting AI invent missing context

When context is missing, the model should ask or mark uncertainty. It should not fill gaps with confident prose.

FAQ

Do I need all 8 input packs every time?

No. Use the pack that matches the task. A literature review may only need the project card and literature list. A reviewer response needs the reviewer package and relevant manuscript excerpts.

Should I use one model or several models?

Use one model for simple tasks. Use several models when the task crosses boundaries: long text, PDFs, code, reviewer response, reproduction, or diagrams.

Can AI write the final paper?

AI can help with outlines, revision, synthesis, and wording. The research claim, evidence, citations, experiments, and final responsibility remain yours.

What is the best first step?

Build the project card. It is short, but it prevents many downstream mistakes.

Final Template: Minimal Research AI Input Pack

If you only want one compact template, use this:

Research AI Input Pack

Topic:
Field:
Research object:
Core question:
Current stage:
Target output:
Audience / venue:
Materials provided:
Key evidence:
Known limitations:
Do not cover:
Do not change:
Output format:
Deadline:

Please:
1. state what you can answer from the given context,
2. state what is missing,
3. separate evidence from speculation,
4. mark uncertain claims,
5. do not invent citations or results.

Once this pack is ready, choosing a model becomes much easier.

The model is not the workflow.

The input pack is what makes the workflow reliable.