Day One
I’m not going to write something every day, but today has been quite productive, so I wanted to capture a few lessons for my own benefit.
I tweaked the project plan earlier and let Claude Opus 4.6 take on the admin work for me. It created new issues, organised the project board, labelled everything up and worked out any dependencies/the best sprint for each task. I didn’t disagree with a single decision it came to. Outsourcing this kind of faffy admin task is a real time saver. It does feel slightly strange being semi project-managed by an LLM, but as all of the outputs are written with my thinking style in mind, any “360 feedback” from me would be really positive. I’ve exhausted my limits with Opus 4.6 most days since launch, and it has really impressed me.
Aside from that, the bulk of the work today involved various compliance tasks and setting things up. So far, so good.
I’m aiming to build a dedicated resource for users of the Freedom of Information Scotland Act (FOISA). FOISA was passed in 2002, yet there is almost no information that is accessible and targeted specifically at requesters outside of what the SIC produces. Where guidance does exist, it is usually explained through how it differs to the UK-wide FOI laws as a bit of an afterthought. I think I can use my experience to create a free resource that is significantly better than what currently exists, and that doesn’t require someone to navigate a decision tree to use it.
As I’ve hinted at before, I’ve also extracted or generated enhanced metadata from all SIC decisions, which has created a resource that I think could be really valuable. My goal is to make this data accessible in a way that is useful to someone appealing a refusal, and not just to data nerds. I’ve got a pipeline set up to keep this updated about once a month, as decisions are not frequent enough to make it worthwhile running more often. I need to be strict about what fields are useful to show by default, but will make a full download available if I can do so without it eating the entire budget.
I used some of my downtime to fix some dead links on WhatDoTheyKnow, having spotted the issues when looking for the FOISA stuff that I wrote for them. Unfortunately, the ICO is really bad at using redirects when moving content, which makes link rot a regular problem. So regular in fact, that I already had a script to find and check these. I let Claude deploy a 5-agent swarm to fix and verify some of the changes. The advantage of using agents like this is that they are fast (run in parallel) and can check if the content changes have rendered any quoted text outdated in a way that’s more accurate than keyword matching. I reviewed what it came back with and it looked solid. I noted the AI’s involvement on the relevant PRs for transparency, but it was good to confirm that this sort of task can be properly automated now. I might need something like this to help keep my own guides from suffering the same fate.