Turn your old work into skills
Get an LLM to look at your old body of work and turn it into skills that you can use for new work. That way you'll start adding value much faster
I started a new fractional gig this week. I’m still taking it slow - understand what’s happening, get myself comfortable with the data and pipelines, figure out requirements from my customer’s customers, etc. etc.
Like you might have imagined, I’m using LLMs (Codex) rather heavily for my work so far. The existing codebase is all in languages I have no clue of (typescript / react / postgres… ), and if I’d taken on this work 2-3 years ago I would’ve spent months just getting up to speed with the tech and understanding what’s happening. Now, in a matter of a couple of days, I already have SOME clue.
I realise that I’m also able to start contributing much sooner, and this is without actually writing any code. Sometime ago, I just told codex “help me figure out and explore the existing data and pipelines. You know how I analyse data”.
The last bit was key - sometime ago I had created a skill based on all the old analyses that I have done. I just pointed an LLM (I now forget if it was Claude or Codex - I use both) to the “work” folder on my mac and asked it to go through all the code that I’ve written and create a skill.
I just asked Codex to summarise the skill “in 3-4 bullet points”. This is it:
Exploratory notebooks should be practical and sequential: load data, inspect schema, transform, plot/table, then move to the next question. Use html_notebook and short chunks.
Write like an analyst thinking aloud: plain-English notes between chunks, including dead ends, not polished report prose or generic EDA boilerplate.
Prefer tidyverse/tidytable style: %>% pipes, right assignment for finished objects, simple helpers, absolute paths for RStudio chunk-by-chunk work, and minimal abstraction.
Explore with quick checks and plots: raw previews, missingness, counts, group sizes, proportions, demographic cuts, simple ggplot2 charts, and move on when something is uninformative.
So, while working on the new project today, when I said “help me understand the data, you know how I work”, Codex immediately pulled up this skill (what I’ve pasted above is just a summary. It’s far more elaborate than that), and then started plugging it into the project.
Because I said I need to understand the data, it has now written a quick R script to analyse it (though there was no R in the repo until then). It has created a bunch of visualisations in my style (I have a skill for that as well - this one is public). It has created notebooks in my style.
Soon, I’ll be doing “real work” for this customer. Of course this will involve a lot of vibe coding. But because the vibe coding platforms have been enriched with my skills (my way or working, along various aspects), the output that it will produce will be enriched with all of my pre-existing skills (pun intended) and experiences. I’m starting to compound from my pre-existing work even in terms of code, not just how I think.
In effect, thanks to these skills, in any new engagement I’ll be able to “hit the ground running”. Over a period of time I plan to keep these skills updated, and also create new skills that will help me get started with any new work easier, and help me add value quicker.
One note - it helps that I have a huge body of pre-LLM work. And this is a very solid base to get started from. Now, if the entire body of work that you have is vibe-coded stuff, then I don’t know how much value will be there in creating stuff like this. Or maybe there is still value in capturing the choices you’ve made thus far!