Skip to contents

Getting help from an AI assistant

Source: vignettes/articles/manual-09-ai-assistants.Rmd
manual-09-ai-assistants.Rmd

Manual

An AI assistant (Claude, ChatGPT, or a locally hosted model) can help you plan an analysis, draft R code, fix errors, and explain a step in MRMhub QUANT, whether or not you write much R. It cannot run your data or confirm the numbers; that stays with you.

Use it in three steps

  1. Give it the package reference first. An assistant has read little about a recently published, specialised package, so from memory it invents function names that sound right but do not exist. MRMhub publishes its whole interface as one file: https://slinghub.github.io/MRMhub/quant/llms.txt. Give the assistant that link and tell it to use only functions that appear there, or paste the file’s contents if it cannot browse the web. For regular use, add the file once to a Claude Project or a ChatGPT custom GPT. Ask it to keep the mexp <- f(mexp, ...) step pattern in the recommended pipeline order, and to use current terminology in prose (interference source/target, precursor and product ions, not parent/daughter) and the plain name MRMhub.
  2. Ask for what you need, adapting one of the prompts below.
  3. Run it and check the numbers yourself. Run generated code on data you understand first, the bundled demo (system.file("extdata", "MRMhub_demo.tsv", package = "mrmhub")) is ideal; confirm every function it used appears in the function reference; and cross-check the results against QC plots and metrics, not just that the script ran.

Prompts to copy

Ground the assistant (step 1), then adapt one of these.

Explain a step

In one short paragraph, explain what correct_drift_loess() does and why drift
and batch corrections in mrmhub are fitted on QC samples, not study samples.

Draft a workflow script

Write an R script using mrmhub that imports my data, normalises by internal
standard, corrects signal drift, quantifies against calibration curves,
computes QC metrics, filters failing features, and writes an Excel report.
Keep the mexp <- f(mexp, ...) pattern for every step, in recommended order.

Explain and fix an error (paste the line and the exact message)

Running: mexp <- normalize_by_istd(mexp)
I get: <paste the exact error message>
What does it mean, and how do I fix it?

Plan the analysis

My samples are plasma extracts with stable-isotope-labelled internal standards
and a 7-point external calibration curve. Which mrmhub steps apply, in what
order, and where do QC samples come in?

Prefer not to write R at all? build_workflow() generates a correct, runnable Quarto workflow from your data and metadata; handing that to an assistant to explain or adjust is safer than asking it to write a pipeline from scratch (see Build a workflow without code).

Limits and risks

An assistant produces plausible code that can be wrong. It cannot run your data, cannot guarantee the code is correct, and does not know MRMhub’s functions reliably unless you show it the reference (step 1). Treat its output as a first draft to verify, never as a result.

Do not paste study data into a cloud assistant. Sample intensities, subject metadata, and unpublished results sent to a hosted service (Claude, ChatGPT) leave your control and may be retained. Share only code and the function map, llms.txt, and reference pages, never the measurements. If the data must stay on the machine, use a locally hosted model (through Ollama or LM Studio), which keeps everything on-device but depends even more on being shown the reference.

Next steps