If you’re an Australian law academic or student who manages research in citation management software like Zotero, you know the drill. You have a bibliography beautifully formatted according to the Australian Guide to Legal Citation (4th ed), but now you need to import those references into your Zotero library. Maybe you’re new to citation management software or didn’t think you needed it, but the next best journal for your article uses a different citation style
The problem? There is no easy way to get this information from a bibliography into your Zotero library. You end up manually copying case names, legislation titles, and journal articles field-by-field into your reference manager. It’s tedious, error-prone work that somehow feels beneath the dignity of modern legal research.*
I went searching for a better way. I found a handful of web apps promising to convert formatted citations into reference manager formats, like RIS. There’s even a decent one floating around on ChatGPT. But every single tool I tried shared the same fatal flaw: they were built for general citation styles—APA, MLA, Chicago—not for the precise, idiosyncratic world of the Australian Guide to Legal Citation. When I fed them an AGLC-formatted bibliography, they stumbled. Cases became journal articles. Legislation turned into journal articles. The output was a mess and would not be much better than manual entry.
That’s when I realised the problem wasn’t the AI; it was the training data. These tools were taught to recognise patterns in general citations, but AGLC4 CSL style in Zotero requires certain information to be added in specific fields for it to be rendered correctly. A citation like Donoghue v Stevenson [1932] AC 562 isn’t just a string of text—it’s a formula: [Case Name] [Year] [Volume] Law Report [First Page]. If your AI doesn’t understand that grammar, it can’t parse it correctly.
Teaching Kimi to Speak Legal Citation
This insight coincided with my first serious exploration of Kimi, specifically the K2 model with its new ‘OK Computer’ functionality. I’d dabbled before, but this was the first time I entrusted an entire project to Kimi’s ecosystem. My goal was straightforward: create a web app that could take an AGLC-formatted bibliography and convert it into a clean RIS file for seamless Zotero import.
The key difference? I didn’t just ask Kimi to ‘build a citation converter.’ I fed it AGLC4 samples—actual cases, legislation, journal articles, and books formatted precisely to render AGLC correct citations. This was my attempt at an exact field mapping that Zotero’s RIS format expects for legal references.
In effect, I taught Kimi the semantics of AGLC, not just the syntax. The OK Computer interface made this iterative training process surprisingly intuitive. I could paste AGLC samples, test outputs, refine instructions, and rebuild the application logic in a single, flowing conversation. The interface feels familiar—much like chatting with other LLM chatbots—but I noticed something refreshing: the free tier didn’t gasp for breath when my context window grew or when I hit what would have been rate limits elsewhere.
The Free Tier That Actually Lets You Build
Here’s where I need to give credit where it’s due. This entire project—the initial concept, the parsing logic, the web interface design, the debugging of edge cases, the deployment—was accomplished on Kimi’s Free Tier. All of it.
I could be mistaken but had I used ChatGPT’s free tier, I fear I would have hit message caps or context limits before achieving something production-ready. Claude’s free tier would have similarly throttled the rapid iteration cycles this kind of project demands. Kimi didn’t. I could maintain the full context of our conversation, refer back to previous decisions, and continue refining without that subtle pressure of ‘you’re running out of free requests.’
Is Kimi K2 a wholesale replacement for Claude or ChatGPT? It’s too early for me to make that call. For general writing tasks or creative brainstorming, the differences are probably marginal. But for this specific use case—building a nuanced, browser tool with some iterative training—the cost-performance ratio is genuinely impressive. When you’re paying nothing but getting the context window and patience to teach an AI a specialist skill, that’s pretty good value.
What the App Actually Does
The web app itself is simple by design. You paste your AGLC bibliography into the input field, click ‘Convert to RIS’. Then ‘copy to clipboard’ in the web app and in Zotero simply ‘import from clipboard’. Behind that simplicity is a parsing engine that’s learned the specific patterns of Australian legal citation:
- Cases: It correctly identifies case names and maps year, reporter, starting page etc to the proper RIS field so Zotero renders them with AGLC precision.
- Legislation: It recognises jurisdiction, year and legislation or bill name, formatting them exactly as the style guide demands.
- Journal Articles: It handles author name formatting, journal title, and volume/issue.
- Books: It parses author, title, publisher, and year into their respective fields, preserving the content for italicisation of publication titles.
The result is a RIS file that, when imported into Zotero with the AGLC CSL style installed, produces references that look like they were hand-crafted by a pedantic law librarian – well, almost.
A Work in Progress
This is a starting point, not a finished product. The current version handles the most common AGLC item types—cases, legislation, journal articles, and books. But AGLC4 is comprehensive, covering everything from treaties to Hansard to social media posts. I need your help to expand it.
Try it out. Convert your bibliographies. Import them into Zotero. If a particular citation type isn’t rendering correctly—or if you encounter an edge case I haven’t anticipated—send me the example. I’ll retrain the model, update the application, and push improvements live. This is how we build tools that actually serve the Australian legal research community: through collaboration and iteration.
The era of general-purpose AI is exciting, but sometimes the real magic happens when you take a capable model and teach it to master something specific. For me, Kimi K2 proved to be the right tool at the right price point (free!) to solve a niche problem properly. Whether Kimi can displace the incumbents for everyday tasks remains to be seen, but for specialised projects like this? The results speak for themselves.
Let me know if you find this useful or if you discover a particular item type that isn’t rendering correctly. Send me an example, and I’ll update the application.
* The fine print: You will still need to create new footnote references using Zotero, which will help keep your dignity as modern legal researcher in check! I recommend keeping it as AGLC initially, just to ensure that all the content has rendered correctly before converting the citation style to Blue Book or whatever house style is required.