AI Sandwich: Humans Frame & Polish, AI Executes Middle
In Compon Engineering, humans drive ideation and final polish while AI automates planning, execution, and review—revealing a universal 'sandwich' model for AI-augmented work that preserves human creativity.
Compon Engineering: A Loop for AI-Augmented Work
Kieran, GM of Cora and creator of the Compon Engineering plugin, developed this framework while building AI tools. It structures knowledge work into four core steps: Plan (define clear tasks), Do (agent executes code or work), Review (iterate on output), and Compound (feed learnings back into the system as repo knowledge). This compounding is the secret sauce—agents reference past mistakes in future runs, creating rapid improvement.
Trevin Chow, a key contributor, expanded it for product work with Brainstorm (explore undefined problems) and Ideate (generate wide-ranging ideas). Kieran notes the 'Do' phase is now reliable: "If you have a good plan, it does the plan. LLMs are very good at just following steps, doing deep work, like working for hours, days even now." Review and planning are maturing too, with automated browser testing validating specs.
The AI Sandwich: Humans as Bread, AI as Filling
The framework reveals humans are essential at the edges, not the middle. Kieran and host Dan call this the "AI sandwich": humans provide the bread (framing at start, polish at end), AI fills the middle (execution). Trevin coined the metaphor, capturing how middles automate while edges demand human taste.
At the start, humans lead brainstorming and ideation. Kieran emphasizes tight human-AI loops here: "The human should think hard, the LLM should support the human." Once framed, hand off to AI for autonomous planning. This contrasts spec-driven development, which over-involves humans everywhere—wasting energy where AI excels.
At the end, post-automation, humans polish for feel and beauty. Drawing from Pomodoro technique (extra time after task completion yields breakthroughs), Kieran says: "You will go deeper. You will go further than you would do." Humans click around, spot intangibles like "this doesn't feel good," and elevate from good to great. Without this, outputs become "slop."
Dan extends it beyond engineering: software engineers shift to product-manager hybrids, focusing on frames and joy-sparking polish (beautiful code, UI, copy). Applies to copywriting, strategy, design—any knowledge work.
Why Humans Own Edges: Framing, Taste, and Rarity
AI struggles with frame-shifting. Dan's example: knee pain framed as "take Advil" (local fix) vs. "stretch IT band" or "stop running on concrete" (higher frames). Humans excel at bounding problems; AI needs humans to set environments where it thrives.
Rare expertise compounds this. Feedback loops are sparse (e.g., career-spanning insights), hard for models to ingest. Outputs stay generic without personal tuning: "Language models... end up being a little bit more generic and less personal to you and your situation."
Kieran adds artistic parallels from his music background: AI like Suno generates songs, but lacks live performance magic or melody invention. Start (composition) and end (performance) remain human; middle (practice) automates. "There is something internally in the human... they feel that."
Full automation fails authenticity: "If you ship something... if you want it to be your own, you cannot fully automate everything. It's like art." Simulations (e.g., 100 personas) help ideation but need human 'yes/no' tied to joy.
Limits of Current AI and Path to Deeper Integration
Dan sets AGI bar at 24/7 profitable agents autonomously frame-shifting tasks. Current tools like OpenAI's o1-preview run scheduled but lack persistence: "It's not like you just say, 'Hey, go and just do a bunch of stuff... it's worthwhile.'" Needs architectural changes for contextual sensitivity.
Kieran agrees edges endure: lean into beauty where you find joy—beautiful abstractions, architecture, design. Engineers become product-focused: "Wherever you feel joy... utilize an LLM to make something that gives you energy."
At Every, this hasn't displaced engineers; it amplifies them as managers of frames and polish. Sandwich model demystifies AI's job impact: ride it by owning edges.
"The beginning and the end, the middle is kind of solved and can be automated pretty well." — Kieran on the sandwich structure.
"Humans are the bread in the sandwich and the AI is in the middle. The AI is whatever you put on your sandwich." — Kieran, echoing Trevin.
"Lean into making beautiful stuff... beautiful code, beautiful abstractions, beautiful architecture, beautiful design, beautiful copy." — Kieran on finding joy post-AI.
"If you want it your own, it needs to be from you or somehow be connected." — Kieran on why full automation kills authenticity.
"All of work exists on this spectrum from it being totally rote to it being art." — Dan framing work's future.
Key Takeaways
- Structure AI work with Compon Engineering: Brainstorm/Ideate (human-led), Plan/Do/Review/Compound (AI-heavy).
- Be 'in the loop' only for high-think moments—start (framing) and end (polish)—to maximize leverage.
- Automate middles confidently once plans are solid; trust LLMs for deep, sustained execution.
- Compound learnings into repos for agent self-improvement; it's the framework's power.
- Shift roles: engineers to product hybrids focusing on taste, beauty, and joy-sparking polish.
- Frame-shift problems upward for impact; humans own rare, personal expertise AI can't replicate.
- Polish post-automation like Pomodoro extras: click, feel, elevate to beat rising bars.
- Pursue authentic output—AI generics lack 'yours'; tie to personal decisions and performance.
- No job apocalypse: sandwich positions humans as indispensable directors.