You can ask most of the big AI models to create files (which could be a document, spreadsheet or slide deck) but you can also use it to write code to create a website or run a simulation. There is another page here with prompts for simulations or roleplaying (which can require the Ai to write code). Here is an example of how to create a searchable data base that combines vibe coding and agentic use.
Women Composers of Solo Piano Music
I wanted to combine vibe coding and agentic AI in a hard task. I also have a long interest in playing piano music by women composers (and helping my students find this repertoire). The internet has made much more rare sheet music available, but of course you have to know for whom to look. I wanted to see if the various AI models could create a large searchable database. I also wanted to see if it could duplicate or exceed what was in the landmark 2002 A Guide to Piano Music by Women Composers, Volume I, by Pamela Youngdahl Dees (which includes 150 composers). I deliberately did not suggest the website Piano Music She Wrote because I did not want to use that important scholarly work (and it is behind a paywall) but I also wanted to know if an AI web crawl could find the 2500 works that took months and months of scholarly research to catalogue.
- You can compare the two best outputs here
- Claude https://claude.ai/public/artifacts/6cca62d4-410c-496a-bbe1-9d862bbeb207
- Codex 5.2 https://josebowen.github.io/PianoWomen/. Note, when you click on a composer the info is displayed at the top of the page–it works better for searching.)
- The results from more models and techniques are further below.
Lessons Learned
- Managing a big AI project is very much like managing a group of humans. Who is best for which task? When are the check-ins? Who is willing to tell the truth?
- The prompting was secondary and the project went MUCH better when I provided only one step at a time.
- It took a LOT of credits and I exceeded my weekly Claude allowance ($20 tier PLUS extra $20 in add-on credits) in no time.
- It was better to start with a specialized bibliography tool (Undermind) then have Claude Opus expand that and create a good protocol for the work and organize the database. Then I needed to switch to a faster model (Claude Sonnet or Gemini Thinking) to search archives around the world and double check links. Then back to a smarter model to for checking and formatting.
- All models needed repeated iteration, but eventually this made things worse (even worst with the smartest models that got more frustrated). Switching between models (ChatGPT to Claude) to check each others work, was better.
- All of this took a lot of monitoring. I thought I would apply what I had learned to a second database (Jewish composers of solo piano in the 19th century) and I now had a strategy but it still takes a lot of MANAGEMENT. Still, these are huge projects that have taken humans lifetimes, so it takes me a a few hours to get a customized database, that is still amazing.
- Codex and Cowork only work while my laptop is on so I used them only for the agentic work (ordering sheet music).
The Prompt
- Note the emphasis on workflow in this prompt. Steps 1-3 are just vibe-coding while steps 4 and 5 are agentic.
- STEP 1
- Start by doing an exhaustive search of published academic work for resources about 19th century women composers (born from 1800 to 1900) who wrote classical piano music. Look for books, articles, archives (including national and specialized archives), encyclopedias, indexes and websites that list women who published classical compositions or that list published compositions by women or index sheet music.
- Examples include A Guide to Piano Music by Women Composers, vol 1: Composers born before 1900 by Pamela Youngdahl Dees. Be aware that sources are often incomplete so find multiple sources.
- STEP 2
- Your job is to create a searchable data base of 19th century female composers of piano music to include any women who published verifiable sheet music, born between 1800 and 1900. This should be a global index but include all European countries. Your task is to make a comprehensive list. Do that my searching Wikipedia, the IMSLP website, rare piano scores and the sources discovered in step 1.
- Then make a nicely-formatted list that is organized by country and within that a chronological list of women composers born in that country.
- Include the key and duration of each piece.
- Include a link to sheet music or archives or published editions where they exist.
- List only piano compositions (in opus or chronological) and only list women composers who wrote solo piano compositions. Use this format:
- COUNTRY
- Last name, First name (birth and death dates)
- Brief biographical information
- Piano composition opus 1 (duration, key, and links to sheet music)
- Piano composition opus 2 (duration, key, and links to sheet music)
- STEP 3
- Create the code (and html file) for a website that provides searchable access to this data base and is well-formatted and easy to use. An excellent model is that of the Swedish Musical Heritage https://www.swedishmusicalheritage.com Search this and other national archives as resourses but also create a new website that has these features but is exhaustive for women piano composers of the 19th century.
- Searchable elements should include country, composer name, title, key and piece duration.
- STEP 4
- Create and deploy a public website where I can share this easy-to-use searchable database.
- Ask me any clarifying questions before beginning.
- STEP 5 (not all models were agentic)
- Look through my pdf library and identify pieces I am missing that are not also available as pdfs downloads from IMSLP and find archival sources on WorldCat and then order them for me from SMU Interlibrary Loans.
Results
- Claude Sonnet 4.5 (Free) https://claude.ai/public/artifacts/5abe9296-3961-4d00-b445-8b02594d43a3
- Free Claude divided the work into tranches and eventually found 30 composers (after several “keep going” prompts) but it triple verified (part of my system prompt). It found more composers, but needed paid credits to add them.
- Claude Opus 4.6 ($20/month Paid) + Sonnet for grunt work. https://claude.ai/public/artifacts/6cca62d4-410c-496a-bbe1-9d862bbeb207
- Even with paid Claude, I ran out of credits periodically and had to wait hours to have it expand beyond the initial 67 composer (but eventually to 253 composers—the highest total). It included a good bibliography of sources created short bios and found figures not in any of the standard references including Black composers Catalina Berroa and Estelle Ricketts (where scores are very hard to find), but it also routinely fabricated pieces when the link was wrong.
- It admitted that with the available time it could not be exhaustive, but it made up some pieces and created some links that did not work. Follow-ups corrected much of this It seemed very upset about the Western bias and said many versions of “Europe still dominates (~60%) — and that ratio is structural, not a research failure. IMSLP digitizes public-domain scores, which overwhelmingly come from European archives. Women composers from Africa, most of Asia, and the Middle East either composed in non-piano traditions, remain under copyright (died after 1953), or had their manuscripts never digitized.” (Some of that might be my underlying system prompt to avoid Western bias, but it was absent from the other models.)
- In the end, however, it overthought the problem and things got worse as I tried to expand the data set. Opus suggested that I use Sonnet to scrub the data and have it check every link. I did that, and indeed that worked. It cleaned up the data and found 330 composers. Checking these links I found no errors (but I did not check the over 1000 entries.)
- OpenAI Codex 5.2 Thinking (Free for this month) https://josebowen.github.io/PianoWomen/
- Codex sits on my laptop and created files in a folder on my laptop, but that also meant it stopped when I put my computer to sleep. It ran as an html file on my computer but I had to load the files to github to share them). It worked for a hours and hours on this and found 230 composers, including Cecilia Arizti (although it only found two works, but included correct links to the Library of Congress records.) . Most of the pieces and links are real and work, although the formatting is less good than Claude. Codex also required some follow-ups:
- 1. If a woman composer listed did not write piano music , then eliminate her from the list. This is a data base of only women composers who wrote solo piano music (and were born between 1800-1900). For example, Lili’uokalani (1838–1917) did not write music.
- 2. Consolidate the list of countries and use only the most common name used today. For example, consolidate the Austri-Hungary, Austrian Empire etc into just Austria. Remove the word “Kingdon” from the country list and consolidate Denmark and the Kingdon of Denmark lists into just Denmark. Use common modern names for countries.
- Kimi 2.5 Agent (Free). https://2e6vvvbcy5vhi.ok.kimi.link
- The Kimi found only 27 composers (but I was out of credits after one follow-up). Kimi asked me a few clarifying questions and was quickly finished (10 minutes?) Some of the pieces are false with links to the right website but to non-existent pieces. No keys or durations. Still, as an agent, it deployed the website on its own and all I had to do was click the link.
- Gemini Pro (Free). https://gemini.google.com/share/26c109d103df
- I started with Gemini which needed a few follow-up directions, but it was fast and it deployed a decent website of 67 composers and 295 pieces. There are no links to sheet music, but it did find and correctly identify works by Cecilia Arizti, a Cuban woman who is not listed is any of the standard sources above and whose sheet music is only available in an OOP book and not online.
- Undermind.aihttps://app.undermind.ai/report/80a0cbfae57617da3362dca0aba219444780375d1b0ba4c3f3099e55b40383ef
- Undermind is just an academic literature review tool so I could only have it create a bibliography. It does not search for websites but most of the books and articles it found were not used by the other Ai tools.
Jewish Composers of Piano Music
- This is how AI creates work. After the first experiment I thought I could do the same for Jewish composers. I used a similar prompt (and without the agentic steps of downloading the actual sheet music and making ILL requests.) This was more complicated as all of the AI tools pointed out (Jewish ancestry? what are Jewish themes?) I will return to the other models when my credits revive but even with gemini, I had to stop asking it to add more composers (credits gone) and so lots of missing names. A big database is possible but it takes lots of tokens.
- Here is what ChatGPT 5.2 produced after a few tries: https://josebowen.github.io/JewishPianoChatGPT/. Most of the links work and the links to complete works lists are accurate and helpful. But as often happens the list gets smaller at some point as the smarter model overthinks accuracy and I had to ask it to restore names, but this is good start.
- I then had Claude Sonnet 4.6 check the expand the version ChatGPT had produced. It added composers and expanded links and bibliographies. https://claude.ai/public/artifacts/12596f9e-a154-42f9-a333-becff665d4d7
- Free Gemini worried less so more names but most of the links don’t work https://josebowen.github.io/JewishPiano/
- Free Grok would not produce downloadable files and just printed all of the code in the body of the chat, but it was never quite right.