Dan Shipper 📧#466
Thanks to everyone that joined Compound Engineering Camp 3 with me and @kieranklaassen just now!
Some of the things we talked about:
1/ Think of this agentic building era like a sandwich:
Sweat the beginning. Automate the middle. Sweat the end.
Compound Engineering is focused on this idea of a sandwich.
Some users expressed gratitude for the Agentic Building Sandwich Strategy at Compound Engineering Camp 3 because the advice saved them time and tokens.
2/ Better inputs → better outputs
Before handing work to agents, get clear on:
- what you’re building -why it matters -what “good” looks like -what tradeoffs matter
ce-ideate, ce-brainstorm and ce-plan are your new BFFs.
@j_everso @kieranklaassen `/ce-code-review` is a multi agent code review that's really comprehensive, i couple it with `/ce-simplify-code` right before it which is terrific. Autofix all issues.
9/ A tiny but powerful artifact: CONCEPTS.md
Use it to align vocabulary across humans and agents. New thing added to Compound Engineering in the last week.
It has clear nouns and verbs make repos easier to navigate, work with your agents, and onboard into your repos.
Run `/ce-compound-refresh boostrap my CONCEPTS.md`
3/ Taste matters, but AI can help.
- First it was coding. -Then it was planning. - Now it’s product thinking, QA, taste, and iteration.
The human job is shifting toward the edges of the workflow (or sandwich)
4/ The AI engineering ladder:
1. Chat/copy-paste 2. Plan-first 3. Autonomous single PR 4. Parallel workflows 5. Cloud fleets 6. Full LFG handoff
The goal: agents working while you sleep.
5/ To climb the AI ladder, don’t just add agents.
Add skills and focus on how you front load the "thinking" work with your agents, then let the implementation flow. Also add safety nets:
- tests - review agents - browser checks - monitoring - rollback paths
Autonomy only works when the system can catch itself.
6/ Brainstorming is leverage.
Use agents to expand the option space, surface weird ideas, and pressure-test assumptions.
Then apply human taste to choose what’s actually worth building.
7/ For bugs, don’t blindly ask agents to fix a bug. Focus on root cause.
Agents can get super lazy and fix a bug without understanding the root cause. `/ce-debug` forces root cause analysis and will fix that.
Otherwise it may optimize for the fastest patch instead of the real problem.
8/ Compound your knowledge.
After meaningful work, capture what was learned into durable docs.
Future agents should start smarter because of past work.
Power move: add AGENTS.md directive to do this automatically!
10/ Team adoption tip:
Don't focus on evangelizing your tools. tools are personal.
- Ship better artifacts. - Move faster. - Create clearer plans. - Compound knowledge in your repo so everyone benefits regardless of skill/plugin/tool usage
@trevin @kieranklaassen great material! On the call the topic of automated reviews came up. @kieranklaassen said he's only using the code review in the compound eng skill. So no reviewers after PR? I've always heard to use a different tool to review than the one that built it. Thoughts?
@trevin @kieranklaassen Will the replay be available for subscribers?
@trevin @kieranklaassen You just saved me. A bunch of time and tokens. Big love.