22 Agentic Engineering Hacks to Turbocharge Your AI Projects

1. Plan before you act

When a new idea pops up, the first step is to create a /ce-plan file instead of jumping straight into code. The plan can contain any relevant context: a GitHub issue link, a screenshot of an error ( Cmd+Shift+4 → Ctrl+V), design mock‑ups, or Slack threads. If the idea is still fuzzy, run /ce‑brainstorm to flesh out the problem before committing to a concrete plan.

2. Feeding agents efficiently

Agents read the plan file, which acts as a leash: "Plans are for agents, you silly human." The plan tells the agent what research to do, what inputs are needed, and what acceptance criteria to meet. The agent then executes the work with /ce‑work. If something is unclear, you can ask the agent with TLDR? or eli5 this plan.

3. Using the same loop for non‑code work

The /ce‑plan / /ce‑work loop works for strategic documents, product specs, competitive analysis, board updates, and even complex discussion summaries. The process is always: externalize the vague idea into a plan, then let the agent act.

4. Voice as the primary input

Instead of typing, you can speak to the model using voice‑to‑LLM. The large model can fill in gaps caused by pauses or mis‑pronunciations, making voice a robust primary interface.

5. Running multiple agents in parallel

Use cmux to launch 4‑6 agents simultaneously. Typical daily state includes one agent writing a plan, another building, a third running /last30days research, a fourth fixing bugs, and a fifth handling a new task. Each agent works in its own terminal tab, and you can switch contexts with /ce‑plan or /ce‑work as needed.

6. Context matters

Before deciding between tools (e.g., Vercel agent‑browser vs. Playwright), run /last30days Vercel agent browser vs Playwright to fetch recent community discussions. Feed the results into /ce‑plan integrate agent‑browser so the plan is grounded in the latest real‑world experience.

7. Human signals for agents

Agents supply volume; you supply taste. Provide feedback such as "second version is closer but keep the first version's wording," "address the biggest risk first," or "this paragraph is too long." These signals guide the agent toward the desired outcome.

8. Video generation with the same loop

Write a script.md describing scenes, timing, and subtitles, then let the agent render an MP4 using HyperFrames. Each project lives in its own folder, making the workflow repeatable.

9. Turning notes into an agent knowledge base

Sync personal notes (Bear, Obsidian, gbrain, supermemory) so the agent can read years of meeting transcripts, design decisions, and research. This personal RAG layer lets the agent recall past context automatically.

10. Remote‑first work setup

Use a Mac mini as a permanent workstation and connect via Mosh, tmux, or tools like Hermes and OpenClaw. Sync cookies and .env files with Agent Cookie so sessions continue seamlessly across devices.

11. Collaboration and review

Instead of summarizing a meeting, drop the raw transcript into the agent with /ce‑plan turn this into a product proposal. The agent produces a polished document that can be shared via a generated link, allowing inline comments that flow back into the agent loop.

12. Reusable skills

When you perform a task more than twice, turn it into a skill. Ask the agent to "look at the Compound Engineering skill and help me make one for X". The agent scaffolds the new skill based on the existing one.

13. Open‑source contributions as a loop

Identify a daily‑use tool with a missing feature, then close the gap using /ce‑plan + /ce‑work. The same loop can be used to submit PRs to projects like Python, Go, Vercel Agent Browser, and more.

14. Hardware limits and workarounds

Even a MacBook Pro M5 Max with 64 GB RAM can run out of juice after six concurrent Claude sessions. Mitigate by keeping an Anker power brick handy, using a car charger on the road, and disabling sleep with sudo pmset -a disablesleep 1.

15. CLI for real‑world services (Printing Press)

Wrap web actions (GitHub issues, Tesla pre‑heat, etc.) into a CLI called Printing Press. Authentication is handled by Agent Cookie, which injects your browser session into the CLI so the agent can act with full login state.

16. Final loop: talk, plan, build

The complete workflow is: speak the idea → generate a plan.md → let the agent build. The result can be shared as a link for teammates to review, comment, and feed back into the loop.

17. Additive reminders

Agents are addictive; remember to take breaks, go outside, and work on things that matter to you. The process is powerful but can lead to "AI psychosis" if you forget real‑world relationships.

18. This article itself was generated with the loop

The author wrote a Markdown file, fed it to Monologue, let the agent rewrite, and used last30days for recent context. The final output is the result of the same talk‑plan‑build cycle described above.