Build in public Day 10
I spent less time building the product today because I came across Every’s AI Product Management Guide, and it really resonated with me.
One idea from the article stood out:
AI-era product and engineering work should not just be about making agents write code faster. It should connect strategy, ideation, planning, review, and learning, so that every unit of work makes the next one easier.
I strongly agree with that, and I want to apply it to my own project.
But Every’s open-source compound engineering plugin feels a bit heavy for a solo developer. It has many skills, agents, reviewers, product pulse workflows, PR workflows, and organizational context search. Very powerful for teams, but if a solo dev adopts the whole thing at once, it can easily become process overhead.
So today I spent some time making a trimmed-down version for my own workflow.
For now, I’m calling it:
Compound Lite
The core idea is simple:
Keep the compounding engineering loop, but remove the team-level complexity.
The current structure looks roughly like this:
STRATEGY.md
docs/ideation/
docs/brainstorms/
docs/plans/
docs/solutions/
docs/evals/
docs/pulse-reports/
And the workflow is:
Strategy
→ Ideate
→ Brainstorm
→ Plan
→ Work
→ Verify
→ Compound
There are three things I care about most:
1. Ideas should be challenged first
- I don’t want an AI to hear a feature idea and immediately start coding. First, it should ask whether the feature is worth building, whether there is a smaller version, and what long-term maintenance cost it might create.
2. Building and verification should be separated
- The builder handles implementation. The verifier independently reviews the diff, tests, and risks. I don’t fully trust one agent to build everything and then justify its own work from start to finish.
3. Lessons should compound
- Bugs I fixed, agent behavior changes, testing strategies, tool permissions, failure recovery patterns, and other hard-won lessons should go into docs/solutions/ or docs/evals/, instead of being scattered across chat history.
This is not fully autonomous.
For now, it only goes up to Level 1 semi-automation: it can read local files, write code, add tests, run verification, and create local docs. But it does not automatically commit, push, open PRs, deploy, or touch external systems.
I’m going to use it myself for a few days and see whether it actually makes development more stable and clearer, instead of just adding more process.
If it feels useful, I’ll open it up in a couple of days for other solo developers building with AI coding agents.
Also highly recommend reading Every’s article. The ideas around agent-native product management are really worth thinking about:
every.to/guides/ai-prod…
English
