THINKIT: A plugin for user-led and user-controlled scientific development in collaboration with AI

Abstract

THINKIT is a package of coordinated AI instructions and workflows for scientific research. It supports physical understanding and modelling, literature assessment, research-project development, and scientific writing, while keeping the researcher in control of the scientific questions, hypotheses, scope, and substantive decisions. Status: public experimental release. THINKIT is actively developed and tested in scientific use Feedback: If you use THINKIT, I would be grateful for comments on usability, scientific workflow, and problems you encounter: Feeddback link Public description THINKIT helps researchers use AI throughout scientific work: understanding a physical problem, examining the literature and evidence, developing a research project or proposal, and preparing scientific manuscripts. It combines four coordinated methods within a shared framework of instructions, reference materials, and scientific-workflow rules.THINKIT does not attempt to make the underlying AI model more intelligent. Its purpose is to make scientific interaction with AI more disciplined, transparent, and controllable by the researcher. It turns a generic AI conversation into a structured scientific discussion centred on physical understanding, evidence, hypotheses, and explicit scientific decisions, while connecting the discussion to current knowledge and literature when needed. The guiding principle is user-led and user-controlled scientific development. THINKIT is designed to clarify the researcher's aims, work with the observations, evidence, hypotheses, and ideas already available, and answer the current scientific question at an appropriate level of detail. The system distinguishes between discussing and testing the researcher's ideas and generating new ones. One compact, clearly provisional physical candidate may be offered when particularly relevant, but broader generation of hypotheses, mechanisms, alternatives, or solutions follows the researcher's request or choice. Evaluating a hypothesis does not, by itself, authorize replacing it with new ideas. For focused physical research questions, THINKIT also integrates literature verification into the reasoning process. Once the physical question is sufficiently clear for a meaningful search, PHYSIT can automatically use a bounded LITIT literature check to examine relevant existing knowledge and evidence without restarting the scientific discussion or requiring a separate literature-search workflow. Four coordinated methods PHYSIT — physical understanding and modelling. PHYSIT is primarily intended to understand physical phenomena: clarify the configuration, forcing, constraints, interfaces, relevant mechanisms, and regimes; generate, examine and test hypotheses; and develop physically grounded theoretical, reduced, scaling, semi-empirical, or empirical relations. A physical-mathematical model is treated as an expression and test of physical understanding rather than as a numerical-computation target. Equations are tools for expressing, checking, or falsifying the physical picture, not an endpoint by themselves. Numerical calculation, fitting, CFD or code development, and implementation become primary tasks when explicitly requested. LITIT — literature and evidence. LITIT supports literature search, source screening, claim-specific evidence assessment, identification of relevant models and prior work, source cards, bibliographies, and exports for physical modelling, proposals, or manuscripts. Source content is kept distinct from analytical interpretation, and the actual level of source access—such as metadata, abstract, or full text—is made explicit. PROPOSIT — research projects and proposals. PROPOSIT supports the development of research questions, project directions, feasibility, funder alignment, objectives, work packages, consortium structure, and proposal text. It can use literature and physical reasoning from the other THINKIT methods while preserving the distinction between established evidence, working hypotheses, and proposed research directions. Scientific premises and substantive project decisions remain under the researcher's control. KETAVIT — scientific writing and revision. KETAVIT supports drafting, revision, polishing, manuscript development, reviewer responses, and related scientific writing. Editing is designed to preserve scientific meaning, claim strength, uncertainty, author voice, notation, citations, and LaTeX structure. Language polishing does not convert hypotheses or interpretations into established results. Shared scientific discipline The four methods share rules for distinguishing observations, source claims, assumptions, derived results, user hypotheses, AI-generated candidates, interpretations, and speculation. They require uncertainty, missing evidence, and source-access limitations to be stated rather than hidden; prohibit fabricated sources, results, parameters, or validation; and preserve the origin and evidence status of ideas as work moves between the methods. A working hypothesis remains a hypothesis until it survives physical, mathematical, experimental, numerical, or literature-based tests. Acceptance by the researcher does not itself establish validity. THINKIT is intended for researchers, doctoral students, and research teams, particularly in the physical sciences and engineering. Focused questions can be handled directly, while broader scientific development proceeds iteratively and in bounded steps rather than through uncontrolled generation of solutions. Implementation and scope The current implementation is a ChatGPT plugin containing PHYSIT, LITIT, PROPOSIT, and KETAVIT as coordinated skills. The methodological rules, routing logic, and scientific-workflow instructions are contained in the plugin itself. ChatGPT Projects can be used to keep papers, manuscripts, data, code, terminology, and project-specific scientific decisions together, but separate Project Instructions are not required for normal THINKIT operation. A short optional Project entry instruction can be used when an additional routing reminder is desired. During source curation, LITIT can collect relevant source-grounded project terminology, while KETAVIT can use the accessible project terminology record to preserve meaning and terminology consistently during scientific writing. Adaptation to other AI assistants is a possible direction for further development, but compatibility must be assessed separately for each platform. Available functionality depends on the host model, tools, source access, and its ability to expose the required instructions and resources. THINKIT supports scientific reasoning; it does not replace scientific verification or the researcher's responsibility for conclusions. Developed by Ilia V. Roisman.