- The Blog

Hosting on Cloudflare's R2

Unknown — Mon, 01 Dec 2025 00:00:00 +0000

I’ve been building an internal knowledge base at work recently, and wanted a quick-and-dirty means of hosting so that everyone across the organisation could access.

Normally, I’d simply whack it on a public Github repo and use Cloudflare Pages to build and host from there. However, some parts of the site are moderately sensitive - or at least information that I’d rather not have slurped up by crawlers et al.

The perfect opportunity, then, to test out something I’d seen mentioned on a HN post a while ago: hosting the site directly from an R2 bucket.

Domains and databases🔗</a></h2>
Setting up the bucket is straightforward:

In the Cloudflare admin panel, go to `Storage & Databases -> R2 Object Storage</code>.</li>`
`Here, create a new bucket - with a name that’s easy to identify with the site you’re hosting.</li>`
`Upload single file - try “hello world!” in test.html</code> - for testing later.</li>`
`Navigate to the Settings</code> tab and add a Custom Domain</code>. This should (obviously) be one that you hold with Cloudflare, and the DNS record update should be handled automatically.</li> </ol> The bucket is now serving to anyone visiting your site.example.com</code> domain. Try it out by visiting https://site.example.com/test.html</code> - you should see your “hello world!” from step 3.`Adding content🔗</a></h2> In my case, I had a folder of static HTML pages built on my local machine. The easiest way to dump these into the bucket is to click-and-drag into the upload pane on the Cloudflare panel. This is fine for a one-off, but will get tedious with regular updates. Rclone</code> is the best tool I’ve found for syncing content to a bucket. Download and install rclone</code> if it isn’t already on your system.</li> Back on the main R2 Object Storage</code> page of the Cloudflare Admin Panel, click the {} Manage</kbd> button under the Account Details heading.</li> Create User API token</code>, with Object read & write</code> permissions. For added security, limit it to specifically the bucket for your site.</li> Save down the credentials somewhere, ready to enter into rclone</code>.</li> In your terminal, type rclone config</code>, and follow the instructions to link your R2 instance with rclone</code> - name it r2</code> for convenience.</li> Test out the connection by running rclone ls r2:name-of-bucket/</code>.</li> If it’s been configured correctly, you should see your test file test.html</code> from earlier listed.</li> Now, to sync your assets to the bucket, run rclone sync ./assets/* r2:name-of-bucket/</code>.</li> Run rclone ls r2:name-of-bucket/</code> again to check everything’s uploaded successfully.</li> </ol> Fixing index.html</code>🔗</a></h2> We can now navigate to site.example.com</code> and see our page in all its glory… ERROR 404 Object not found This object does not exist or is not publicly accessible at this URL. Check the URL of the object that you're looking for or contact the owner to enable Public access. </code></pre> Ah. A reminder that R2 is not really designed to host whole sites: whereas most servers will default to index.html</code> if a client requests the root of a directory, R2 simply follows our request to the letter, attempting to serve the page /</code>, which obviously doesn’t exist. Fortunately, there’s a quick fix for this, using Transform Rules. We’re going to create two rules: one to fix the site root, and another for the habit that some static site generators have of defaulting their links to /section/subsection/</code>, without an explicit .html</code> page. From the main Cloudflare Admin panel, select the domain you’re serving from (e.g. example.com</code>)</li> In the sidebar, select Rules -> Overview</code>, then click the Create rule</code> button</li> Rule one: Name the first rule something like Index.html for site root</code></li> Select Custom filter expression</li> In the When incoming requests match… section, set: | Field | Operator | Value | | :—: | :—: | :—: | | URI Path | Ends with | / |</li> Under Then…, select Rewrite to.. and Dynamic from the dropdown.</li> Enter concat(http.request.uri.path, "/index.html")</code> as the expression, and hit Save</code></li> </ol> </li> Rule two: Name this second rule something like Index.html for subdirectories</code></li> Select Custom filter expression</li> In the When incoming requests match… section, set: | Field | Operator | Value | | :—: | :—: | :—: | | URI Path | Does not contain | . |</li> Click ‘And’ to add another condition: | Field | Operator | Value | | :—: | :—: | :—: | | URI Path | Does not end with | / |</li> Under Then…, select Rewrite to.. and Dynamic from the dropdown.</li> Enter concat(http.request.uri.path, "index.html")</code> as the expression, and hit Save</code></li> </ol> </li> You can limit these rules to just your specific site (rather than all that are hosted on your domain), you can add an ‘And’ rule of: | Field | Operator | Value | | :—: | :—: | :—: | | Hostname | Equals | site.example.com</code> |</li> </ol> Note that the second rule has a very slight difference in the rewrite: we omit the leading /</code> from index.html</code>. Wrapping up and costing🔗</a></h2> Give all those settings a minute or two to propogate across the network, then try again visiting your site. It should now present you with your index.html</code> page, just like a real server. It is worth mentioning that this sort of hosting isn’t totally free. Generally, you can serve up to 10 million requests per month before Cloudflare will start charging you - you can play with their calculator here</a>. Frankly, if you’re hitting those numbers, you can probably afford to set things up properly. Similarly, the sales team will come knocking on your door if you start hosting very large files, but for a simple static site with nothing more complex that some HTML</code> and images, you’re good to go.

Cloning in your classroom Max Bruges — Sun, 04 May 2025 00:00:00 +0000 Dolly the sheep</a> (left). “But what’s the right answer, sir?” I’d barely been teaching a week before being brought to a dead stop, mid-lesson. I’d done all the right things: pre-taught the vocabulary, broken down the context, painstakingly annotated</a> the iambs and trochees. But that had all been vanity</a>. Rafael wanted an answer. And he wanted the right one. No amount of “well, one could argue…” or “a possible interpretation might be…” or any critical expostulation would satisfy him. If Maths and Science and French and ICT could answer their questions with a tick or a cross, why couldn’t English? Spinning the flywheel🔗</a></h2> Rafael was right, of course. This is the problem with English. Our answers are too fuzzy. There isn’t one, immutable, correct response to ‘How does Shakespeare portray the protagonist</a> as a tragic hero?’ - and Harold Bloom will rise from his grave and throttle you if you even suggest it. Not angry, just disappointed. But that is also the point: we do not teach knowledge. We teach the encoding and decoding of language, in all its fuzzy nuance. This makes the essential iterative loop of learning - answer --></code> assess --></code> improve --></code> re-answer</code> - a slow, manual process. Every answer from a student in English needs to be weighed on Thoth’s</a> scales of criticality, and only the teacher in the room is fully qualified to make that judgment. We look in envy at our STEM colleagues with their binary marking and autonomous testing; for every sentence we evaluate for the nuance of language usage, they can mark a dozen sums right</code> or wrong</code>. The maths student receives an order of magnitude more indicators of success and correctness that the English student does, even before accounting for the ease of self-assessment. The bottleneck to learning in Arts and Humanities is the teacher’s capacity to assess and feedback. But that could be about to change. Fuzzy bots, firm answers🔗</a></h2> Above everything, LLMs excel at fuzzily assessing language, which makes them the perfect candidate for ‘cloning’ an English teacher. Yes I know they aren’t clones</a>. They may hallucinate facts</a> or forget data, but their ability to read and extrapolate from written text is dependably good. Ultimately, that’s what we need in English. We learn it the same way the LLM does: ingesting as much good quality writing</a> as possible so that we can produce our own, mimicking styles and structures whilst changing the content. The classroom of tomorrow can leverage this technology to finally widen that bottleneck; a digital twin</a> of the teacher to serve as a low-stakes feedback machine. They don’t need to be perfect, only ‘good enough’ to give an indication of if the student is the right direction: finally giving English some semblance of the STEM binary; though rather than correct/incorrect</code>, we can ask for better/worse</code>. Cautious optimisation🔗</a></h2> But tread softly, and never mistake eloquence for accuracy. The key distinction here is between evaluation and feedback. The bots are rubbish at the former: too easily swayed by context and weighting to reliably give accurate marks and levels for written work. In my own rudimentary experiements, the bots showed the sort of capricious inconsistency that would get an examiner booted from a marking pool in minutes, with grading that varied wildly even from paragraph to paragraph. Running the same exam script through the AI three times would produce three distinctly different results. There have been some promising steps in recent months</a> to improve on this, but at this stage the tech is simply too inconsistent to take on the quantitative aspect of marking. We may get there with a big dollop of fine-tuning and RAG-ing, but for now the auto-marking bot remains but a twinkle in every GCSE assessor’s eye. “The candidate has demonstrated an ambitious use of vocabulary, commensurate with AO3 Level 4.” Feedback, on the other hand, is ideal. Load up the AI’s context window with a script of pre-written responses (‘Vary your sentence structures’, ‘Consider your paragraphing’, ‘Incorporate more ambitious vocabulary’, etc.) and they can usually be counted on to pick the right one for the right text. And for obvious reasons: it’s doing the job it was made for, to extrapolate text. Assessment frameworks may come and go, marking criteria may change with phases of the moon, but good, broad feedback is forever. For low-stakes assessment, where the practice rather than the outcome is what matters, this has the potential to work brilliantly. A quick nudge in the right direction can keep momentum up and spin the learning flywheel; nudges that you don’t always have the time or space to deliver. In truth, the only way to know is to do. Run some student work through ChatGPT, with a strict prompt limiting them to your pre-written comments. See what it comes up with. Feed in some more tailored comment templates. Tweak, repeat, iterate. It may not be classroom-ready today, but you may have a feedback clone to help you tomorrow. Revenge of the knowledge-based curriculum Max Bruges — Sun, 16 Feb 2025 00:00:00 +0000 The ever-interesting Ben Thompson has been playing with OpenAI’s new Deep Research</a> tool. His write-up</a> of how he used it to generate a report on Apple’s earnings announcements is worth reading, but this section stood out to me: The issue with the report I generated […] is that it completely missed a major entity in the industry in question. This particular entity is not a well-known brand, but is a major player in the supply chain. It is a significant enough entity that any report about the industry that did not include them is, if you want to be generous, incomplete. It is, in fact, the fourth categorization that Rumsfeld</a> didn’t mention: “the unknown known.” Anyone who read the report that Deep Research generated would be given the illusion of knowledge, but would not know what they think they know. Ben Thompson, Deep Research and Knowledge Value, 2024</a> </blockquote> </blockquote> Unseen and unknown🔗</a></h2> Sound familiar? In teaching, we face that illusion of knowledge twenty times a day. The stuck student is easy to identify and intervene with; it’s the confident but compromised one that’s hardest to fix. We’re too easily fooled by flowing paragraphs and correct responses to prompted questions. Often, one only spots the depths of the misunderstanding in high-stakes, unscaffolded assessment on unfamiliar content, where the fallback on routine and skills leaves glaring gaps in answers. “However, to conclude, it can clearly be seen that the allegorical play is…” The unseen poetry portion of GCSE English is the purest crucible for this: students diligently marching their way through PETAL paragraphs but totally missing the central conceit of the dramatic monologue or allegory, simply because they lack the knowledge of those forms. Most heart-breaking of all is the shock that the grading of these sorts of answers is usually met with: the student doesn’t know what they didn’t know. From their perspective, their perfectly planned answer used all the right sentence stems</a>, ticked all the AOs and followed all the rituals from the classroom. Because when it comes down to it, knowledge trumps all else. The curriculum pendulum may be about to swing back from knowledge to skills</a>. But in a world full of fluent-but-fallible bots, that would be a mistake: our students desperately need a solid grounding in facts to spot when hallucinations creep in. Trust but verify🔗</a></h2> I’ve no quick fixes for this in English, other than that cowbell</a> of Literature study: read moar. But it does prompt a useful, meta-cognitive lesson for a students to take and apply outside of the classroom and exam hall. Large-Language Models are going to - and should! - play a central part in their working lives. Teaching them to use the tools properly, modelling safe and accurate usage, is a duty we as educators have to embrace today. As Ben notes above, the models are extremely good at behaving like over-confident GCSE candidates, producing fluent paragraphs that miss key information. If our students can become domain experts, then they can be equipped to spot when these Rumsfeldian knowledge gaps emerge, and challenge them. To out-smart the machine is pretty a awesome experience for anyone, let alone a neophyte dipping their toes into the pool of academia for the first time. What better confidence boost can there be than telling ChatGPT it’s got it wrong, and having it agree? The best way to learn is to teach, so teach that chatbot who’s boss. Gell-Mann and GPT🔗</a></h2> Above all, an interaction like this can pierce the veil of infallibility and teach students an essential lesson in Gell-Mann Amnesia: You open the newspaper to an article on some subject you know well. […] You read the article and see the journalist has absolutely no understanding of either the facts or the issues. Often, the article is so wrong it actually presents the story backward—reversing cause and effect… In any case, you read with exasperation or amusement the multiple errors in a story, and then turn the page to national or international affairs, and read as if the rest of the newspaper was somehow more accurate about Palestine than the baloney you just read. You turn the page, and forget what you know. Michael Crichton, 2007</a> (yes, that</a> Michael Crichton) </blockquote> </blockquote> Swap ‘newspaper’ and ‘journalist’ for ChatGPT, and you’ll see exactly what Ben Thompson is driving at. Of course, Crichton’s point is that we forget this once we leave our domain of expertise, slipping back in to trusting the authority of fluency. But for a brief, shining moment, our students can know how right they are, and know that these tools exist to augment their knowledge, not replace it. Launching the A.I. Academy Max Bruges — Fri, 24 Jan 2025 00:00:00 +0000 The best thing about my day job is the boredom. Not my boredom (how could I be bored when I’m literally paid to talk about words</a> and witchcraft</a> all day?), but the boredom on the other side of the desk: the students’. There is no creature on this Earth more motivated, more inquisitive, more hungry to learn what’s next than the child who’s bored with what they’re doing right now. If you can manage it, bottling that boredom lightning is the best part of being a teacher, and the best way to teach. It’s in this spirit that we - a few foolhardy Year 11s and I - have launched (probably) the city’s first After-School AI Academy. Building for us, learning to last🔗</a></h2> An LLM running on one of our Macbooks We’re starting small</a>, with a simple question: how can we make this incredible new technology our own? As a child of Web 2.0 and ICT lessons, the last thing I’d want is to lock another generation into walled gardens. They need to see this technology as something they can play with and make their own, not a pay-to-use, locked-down utility managed by someone else. No, our Academy is all about building, running, tinkering and tweaking our own AI tools, on our own computers, under our own steam. Open and free🔗</a></h2> Our first sessions have been all about learning what models are, how they’re created, and - above all - the value of building and learning out in the open. We couldn’t do what we do without the miraculous open-source free-for-all that is Ollama</a>. Everything we do carries on in that spirit, sharing code as collaborators, not as students led by a teacher. The group is already buzzing with applications and ideas, from processing nutritional information to virtual tour guides. I’m sure I’ll be showcasing some of their work in the near-future. If you’d like to join them or share ideas, you can drop us a line</a> and follow along</a> below. And if you’re reading this, academicians, keep it up. Happy hacking.