https://try-management-app.vercel.app/ai-dev-day?v=4
Pictures from the presentations
There might actually be a video at some point I can link to instead
Better to delete a Claude.md than to pollute your context with outdated info.
Only info worth putting into the claude.md is the things it don't get right. Update it when you have to correct it.
Vibe Coding is now at a level where it is at my level or has surpassed me
Lets me focus on problem solving. The code has always been the medium. Getting the AI to produce beautiful code is almost as satisfying.
Makes it possible to be productive in languages I do not know.
- Shift for Magnus
- Jesse shared blog post on superpowers. Superowers Oct 09
- Skills. Antrhpric released 4.1 or 4.2. Skills introduced.
- A huge shift in quality
Never ever ask the AI to start coding before discussing first. Always produces shit code immediately.
I want to tell it what to do and discuss before it does anything.
90% in brainstorming mode A skill that I use to juggle back and forth with the AI to discuss solutions. Which side effect certain designs have.
Creates plan files using the skill.
He keeps them in the project, but does not commit them. The plans should be committed. Was not sure if he wanted or not, but they should be.
We need to find a way to do that properly.
Håkon has set up all his skills to delete the plan files once they are done.
Magnus really wished there was a simple way of using them as documentation. Feels there is a lot of value in the designs.
Agree that outdated documentation is not great, but it would be good ...
If you do not clear the context at regular intervals, or do too much, you run into implementation plans where nothing of what was asked was actually done.
Important: even though you try to be explicit, you can end up with vibe coding results.
Spend more time in initial phases. Spend extra time there where you know you are actually good.
It will accumulate errors in the context, also when compacting.
Better to start over fresh.
Better to be careful and not run into compacting.
The cleaner yoru context is, the better yoru plans are, the better the results it will end up with.
It will then always have it in the context. It will pollute your context and thinking.
It will transform your workflow.
If anyone needs help getting into it, Magnus can try.
https://github.com/obra/superpowers
https://skills.sh/obra/superpowers/brainstorming https://skills.sh/obra/superpowers/systematic-debugging https://skills.sh/obra/superpowers/test-driven-development https://skills.sh/obra/superpowers/writing-plans
If he could have more arms and windows, he would be more productive. He is the current bottleneck for what could be done.
You have to add a setting to Claude Code. Makes it a complete team with a team lead. Useful for bigger tasks.
Even bigger team than sub-agents.
Magnus: the AI cannot think for itself. I do not want it to think it is the team lead. You need to manipulate the AI to think correctly.
Audience (Martin): never got sub-agents to work, but teams is crazy good. Not necessarily huge tasks. Testing sub-agent that runs the browser window or testing output.
Fullstack engineer
Next and Caching
Adept to your needs
LLMs are proediction engines
humans interact.
No right or wrong. Just good and bad variations.
The most efficient ways of getting agents to do wthat you wnat
From A to B. From prompt to solution.
Consders
- cost
- speed of route
- conenience Takes shortcuts
- Hallucinate APIs
- Use deeprecate code. Trained on old code / library names.
- Quality is perfect, but does the wrong thing.
What it thinks you want, not what it knows you need
- Skips critical parts. avoids complexity
- Pure lazyness.
The best map is what works for your workflow
- Configuration
- ..
- ...
Set of rules for every agents session.
HOME/.claude/CLAUDE.md
Consistency Align with expectations
Minimal For everyone working on the project Respected most of the time Should update quite often Consise as possible
/init
Set of instructions to
- perform actions
- gain knowledge
http://skills.sh
npx skills add skills
Local Claude.md + agent browser = 💨
When performing frontend dev always
- check localhost, if not running, as k the user to start it, do not start yoruself
- use
agent browserskill to verify your work visually - use
agent browserskill to check scope of bug tasks before starting
/clear
users are reporting tahat on the https://raumagarn.no/forhandlere" we see a wierd looking input. Please check it and to try to fix it in code. Deveserver is running
....
Skill: agent-browser
...
/clear
users are reporting tahat on the https://raumagarn.no/forhandlere" we see a weird looking input. Please check it and to try to fix it in code. Deveserver is running. Please. Also use the agent browser.
...
agent-browser opens http://localhost:3000/forhandlere
Screenshot saved to ...
I can see the issue clearly. Search input with two overlapping placeholder texts. 'Søke etter forhandler' and 'Søk etter sted' is overlapping
Fix: changed ...
Stay away from 1M Context models for better performance.
Asking the agetnt to summarize its current work
- clear all clutter
- 190k tokens -> 20k Auto compact
- Task too big
- multiple tasks within same
- youre getting less quality clode
/clear
- start from scratch
LLM assisted scoping
Planning ahead a
Accurate specifications reflecting bheavior
- Kiro (IDE)
- Get Sh#t done
- Superpowers
Minimal can be used for any changes of any size
Workflow propoosal design tasks spec
Multi-task
Spin up new versions of the same project
pwt.skala.sh
bun install -g pwt-cli
Git worktrees let you check out multiple branches simultaneously in speearte directoreies. Instead of stashing or commiting...
PWT simplifies worktree management.
NPM utility. designed for AI agents.
Live demonstration of AI-powered crisis management and media training prototype
STT - speech to text TTS - text to speech
Can use Norwegian number. SIP trunking. Must talk to Telia.