AskBox

Lalitha A R

AI · Research · Tooling · Design / UX · Data & Analytics · Software Engineering · Product Management · Marketing

Anonymous asked:

My background is quite different from many researchers here. I have spent over 30 years in industrial biotechnology, focusing on fermentation, bioprocess development, manufacturing, and process scale-up. I'm now exploring how AI can support scientific decision-making in industrial R&D. One thing I've found particularly valuable is using Claude not only for literature review, but also for: Comparing experimental results across multiple studies Identifying conflicting conclusions in the literature Designing experiments and DOE plans Reviewing SOPs and suggesting improvements Brainstorming hypotheses before moving to laboratory validation I'm still learning the best ways to integrate Claude Science into an industrial research workflow, so I appreciate posts like yours. I'm curious—have you found any prompting techniques or workflows that significantly improve the quality of scientific reasoning, beyond standard literature summarization?

Thank you :>

I haven’t tested Claude Science extensively. Incase of general Claude, here are some things that work really well,

- If you are indeed using it to find literatures and approaches, tell it to maintain something I call the papertable.md, basically let it maintain 1 file with a table of title, authors, doi/link, core theme/topic and another file where it lays out exactly why it thinks that is relevant and it rating each paper out of 100 in relevancy. This is really helpful when we want to audit or even check by ourself.

- Incase of it reading/running analysis on existing data, let it create a python (or whatever the language of your choice is) script that prints out the numbers and stats it claims the data has. Usually, when I know I will use a number - it will create a folder isfv/ (immutable scripts for verification) where it creates a copy of the data it runs the experiment from, the script, along with a README.md that outlines the core stats, purpose etc. At times, the script may print correct but still the number Claude says will be off or it will assume usual case and round it off (can mess up loads of calcs if you want more decimals and stuff)

- If you are using Claude code to read pdfs etc, let there be a source/ folder where you dump all the pdfs, and let it first read and fetch all core numbers/data/facts it needs from each pdf to a paper_files/where each paper is title of the paper.md per paper. Spawn another agent after this process is complete which verifies every number and stat. This helps alot because one in saving the tokens beyond one time whenever you need to refer a paper, another in decreasing hallucinations by a lot.

- Be careful of the context it’s pulling the paper from. We once had this situation where there were two stats: 6/10 and 6/16 and both as a number for exact same context/problem framing evth. Claude chose one over another, but then someone in our team flagged that there are 2 numbers and turns out, the one Claude pulled in was - yes in the same subject, but not relevant to the problem at all cuz that paper’s original goal was something else.

- Claude sometimes dismisses numbers/fact because it isn’t ‘scientific’ enough, but often what counts as valid in academic research and industrial research differs alot. Make sure to trust your judgement above what it says and decide for yourself. Also Claude sometimes default to writing for the Reviewer 2 and preempt everything, and adds quite a bit of jargon for the sake of it (inherited from academia a bit yeah), instead of just treating it as a record of scientific activity. Make sure to let it know of the tone and utility/goal of the work, it helps better.

If I remember more, I will update here.

Anonymous asked:

When discussing once you told me to go for primary sources when collecting data, aka I have to look if a statistic/claim is empirical to a paper or if they are citing some other paper as the source of that claim. But I come across really good awareness organization's websites, they will have verified before using a statistic right? Why can we not cite them?

We have a lot of respect for these organisations and the work they do. The reason they are not citable here is not about credibility — it is about purpose and audience. Awareness organisations write for the public. Their goal is to persuadeView more

Anonymous asked:

I am working on research where I am reading a lot of different country's law on a particular topic. But I am struck - how will I know whether what I find is the most recent updated law? I tried using the most recent once as available on the site but turns out there was an amendment just few weeks back that wasn't updated in the page but was available in the news and a court case, I almost missed it.

First, go to the official regulatory body's website directly — not a third party that hosts the document. If we are talking of Indian Food safety, it is fssai.gov.in, for the EU it is eur-lex.europa.eu, for FSANZ it is foodstandards.gov.auView more

Anonymous asked:

We’re considering moving away from Webflow for something more AI-friendly. Any recommendations? Our biggest pain point today is the MCP experience: Slow, Browser-dependent, Token-hungry. What I’m looking for is pretty simple: Have an excellent building and editing experience with AI. For me or any person of the team. But I don’t want to spend any time thinking about Core Web Vitals, technical SEO, or infrastructure. Curious to hear what you’re using and what you’d recommend 🙏

A mix of custom and switching to Quarto for documentation/static sites? But that wouldn't do without thinking of technical stuff, yes. Maybe try to map out core stuff you keep altering, any patterns? If you notice any patterns, it can be aView more