Blog post triage with a local coding agent

March 24, 2026 by Willem van den Ende llm coding-agent blogging

I made a skill for a coding agent to help me get more of my draft blog posts over the line. I enjoy writing, and am somewhat fluent in it. Publishing that writing is more hit and miss, however. I often lose energy just before a piece is finished enough. I want to publish more often, and need to form a more effective habit for it.

What did I get out if it?

I got a working agent ‘skill’ in an hour or so. I like the QWEN models for their no-bullshit approach to feedback. As it turns out, I have about 60 pages with the ‘Candidate Blogpost’ tag in my notes, but most of them are not more than an idea. Only some of them have enough detail to turn into a post. I am going to keep this around, prune my candidate blogposts, and add my recent clippings to the mix.

Quite a few of my candidates were ‘just links’ according to the model, but as I am inspired by Simon Willison, there is value in sharing links with a brief description on why I think they are relevant. Probably in a different category.

How did I develop the skill?

I was inspired by two writings:

Jurgen De Smet asking how do you write long form articles?.
Chris Parsons suggested to brief an agent for daily tasks, and use the backbriefing loop from “The Art of Action” to improve them.

I like “The Art of Action” - detailed, yet practical. So I had a chat with a frontier model to develop a skill for a local model to surface notes that are almost finished, with some suggestions to get them over the line.

This was my initial prompt. Full chat transcript in the Further Reading section.

https://www.chrismdp.com/stop-prompting-start-briefing/ suggests an art of action style backbriefing loop for daily work. I would like to use a local model with pi, the shitty coding agent, instead of claude code. I have trouble publishing blogposts. I have many drafts, marked as CandidateBlogPost in an org-roam directory. I wonder if I could make some kind of pi extension or skill that finds candidate blogposts, helps identify ones that are almost finished, with a suggesion on what to do next for the top 3 almost finished, and suggestions for others on what to add. Probably prioritize recency. I could run that as a cron job in the morning, and create a new daily entry (I use daily entries for org-roam) to get me starte.d Goal would be not to have AI write my posts, but help me finish in pomodori instead of days.

What I found interesting was that, maybe because I mentioned the links were in an sqlite database, claude desktop spontaneously suggested to create a bash script as part of the skill. I used to have a meta-skill to separate the deterministic parts of agent skills into scripts, but that does not seem to be necessary anymore. I prune my agent setups continuously, only keeping what is needed.

Tradeoffs

Initially I planned to run this as a scheduled job, but from the development chat it emerged that backbriefing (improving the skill as we run it daily) would not work if it runs scheduled.

I chose a local coding agent with a local model, because I don’t want to share my personal notes with a cloud service, and I thought that a smaller model would be more than powerful enough.

What did I get out if it?

How did I develop the skill?

Tradeoffs

Further reading