Agentic Coding

What People Build

A guide to the kinds of projects that tend to work well with AI coding tools, and why.

Some tasks are simply easier to describe and easier to check than others. The projects that tend to work best with AI coding tools share a common structure. The input is identifiable, the transformation can be described in rules or steps, and the output can be checked against something you already know well enough to evaluate.

That pattern explains why the same kinds of projects keep coming up, from messy exports and metadata cleanup to literature screening, static websites, and recurring reports.

Questions at the Outset

Before you start, ask: What is the input? What change am I asking for? How will I tell whether the result is correct? If the answers are still vague, the project needs more definition before you bring a tool into it.

Data Cleaning and Transformation

Data cleaning is often the clearest place to start. Spreadsheets and exports have visible inputs and outputs, and you can usually check the result by comparing a handful of rows against the original.

Clean messy exports

An export with broken dates, stray line breaks, and encoding problems. These tools can standardize the fields, repair the text, and save a clean version.

Merge and deduplicate

Combine records from several sources into one sheet. Match on DOI, title, ID number, or name, and send uncertain duplicates to review.

Reformat for a target system

Take records in one format and convert them into the import structure another system expects: CSV to JSON (a structured data format many platforms use for import), one metadata schema to another, or a local spreadsheet to a platform template.

Generate summary reports

Turn usage statistics, survey responses, transaction logs, or study results into the tables and charts a memo, report, or paper already needs.

Monthly reporting

Monthly reporting works as a starting point because the files recur and the output is already familiar. You know which columns matter and which totals belong in the final table. When the numbers look wrong, you notice.

Inventory First

Before changing anything, ask the tool to inventory the file: row count, columns, missing values, inconsistent formats, obvious oddities. That baseline is your comparison point when you evaluate the cleaned version.

Working with Research Literature

Research literature work fits well when the criteria can be written down clearly. Tasks like deduplication, reformatting, screening, and restructuring exported records all follow rules you can state in a sentence or two, which makes them easy to delegate and easy to verify afterward.

Search strategy builder

Describe a topic and generate platform-specific search strings for PubMed, PsycINFO, Web of Science, or other databases, including subject headings and Boolean syntax.

Citation formatting

Take a mixed list of sources and convert them all to APA, Chicago, or another target style. Clean obvious errors before the list goes into a paper or guide.

Literature screening

Screen exported abstracts against inclusion criteria, classify them by sub-topic, and flag borderline cases for manual review.

Publication tracking

Pull a department's or research group's recent publications from OpenAlex or Scopus, then organize them by topic, method, venue, or co-authorship pattern.

Systematic review support

Systematic review work usually needs to be broken into separate operations like deduplication, screening, tagging, and reporting. If you keep those stages distinct, you can see what the tool is doing at each step and catch errors before they propagate.

Metadata and Records

Metadata work follows the same pattern as data cleaning: local rules applied consistently across many records. If your field mappings, date conventions, and controlled vocabularies are already documented somewhere, the tool can get surprisingly far on its own.

Batch record creation

You have filenames and brief descriptions for a collection. Generate draft records with normalized dates, structured fields, and suggested categories for review.

Schema crosswalks

Map records from one metadata schema (a set of field definitions) to another: Dublin Core to MODS, one CSV layout to another, or local conventions to the structure a repository expects.

Validation and quality checks

Run records through a checker before import and flag missing required fields, invalid dates, broken identifiers, and vocabulary violations.

Dataset documentation

Take a CSV with no documentation and draft a codebook: column names, data types, value ranges, missing data patterns, and variable descriptions for review.

When metadata projects go well

Metadata projects tend to go smoothest when you hand over the exceptions along with the rules. If you know that dates before 1900 use a different format, or that certain collection codes map to non-obvious categories, stating those cases in the prompt saves a round of correction later.

Building Things for the Web

With static sites, interactive visualizations, and small browser-based tools, the output stays visible while you work. You can open the page, compare it to the source data, and see whether the result holds up. When something looks wrong, you can describe the problem concretely ("the third column is cut off," "the dates are sorting alphabetically") and the tool can usually fix it in one pass.

Online exhibits and showcases

Turn a spreadsheet of items into a browsable web exhibit with search, filtering, images, and short descriptive text.

Interactive visualizations

Build charts, maps, timelines, or network diagrams from your data and embed them in a site, guide, or course page.

Dashboards

Visualize trends over time: usage patterns, collection growth, research output, or survey results. Update the dashboard when new data arrives.

Text analysis

Process transcripts, archival materials, or open-ended responses through entity extraction or thematic coding, then visualize the patterns.

Concrete feedback

Web work is more forgiving than server-side code (the logic that runs behind the scenes, invisible to the user). Errors show up the moment you load the page, so the feedback loop is short and grounded in what is visible.

Presentations and Documents

This category works best when the job is structure, formatting, and first-pass assembly. The tool is good at organizing material and producing clean layouts. Argument, emphasis, and the order of ideas still require your judgment, and the transitions will almost always need rewriting.

Slide decks from an outline

Turn a speaking outline into a deck with title slides, section breaks, placeholders, and a presentable visual structure.

Handouts and guides

Turn notes into a formatted handout, a one-page quick reference, or an HTML guide accessible from any browser.

Policy and procedure documents

Take an old policy document and a marked-up revision, then generate a clean redline or a pair of clean and tracked-changes drafts.

Data management plans

Draft a data management plan from a grant application or project description, following the funder's template and flagging anything that still needs input.

Structure First

Treat the first draft as layout and structure. Getting headings, section breaks, and visual hierarchy in place is often the most tedious part of document assembly, and the tool handles it quickly. The prose itself will still need your voice.

Automating Repetitive Workflows

Routine chores are a good fit when you know how to do them but find the repetition tedious. The steps for link checking, format conversion, quality-assurance passes, and recurring reports are already known, and the value is in having them done reliably without spending an afternoon on them.

Link checking

Check all the links on a web page, in a guide, or in a reference list. Flag broken ones and suggest likely replacements.

Meeting prep and follow-up

Pull together agenda items from email threads and shared documents, then format an agenda and draft minutes from your notes after the meeting.

Training documentation

Turn training notes into step-by-step documentation as a PDF or web page other people can use later.

Assessment and evaluation

Compile survey results, rubric scores, or pre/post test data into summary statistics and charts shaped for a report or accreditation document.

The compounding return

Thirty minutes saved on a recurring report or cleanup step adds up over a year. You usually put those hours toward something you had been deferring, like documentation or outreach.

Where This Gets Risky

These tools fit less well when the result is hard to verify or the task depends on judgment you have not written down. Work that touches credentials, patron data, or consequential decisions carries a different kind of risk. A mistake may not be caught until after it has caused harm.

High-stakes decisions

Admissions, hiring, disciplinary decisions, and other consequential judgments need stronger human oversight than this kind of workflow can provide.

Production systems with secrets

Live credentials, patron data, and security-sensitive infrastructure require a higher bar for review and deployment than a quick iterative loop can guarantee.

Tasks with unclear success criteria

If you cannot explain what a good result looks like, you will have a hard time guiding the tool or evaluating what it returns.

Polished prose with substantive claims

Drafting help is useful here, but argument, evidence, emphasis, and final wording still need close review.

Where to Start

Start with work you already know how to do by hand. With a familiar task, you have something concrete to compare against. As you compare, you learn how the tool tends to work. You see which instructions it follows precisely, and which it interprets in ways you did not ask for.

Good First Projects

A messy export, a recurring report, a static page, or a batch of records. Pick something small enough to finish in a sitting, where you already have a clear sense of what the output should look like.

Related guides

Further Reading