Course outline
Agentic Engineering: the hands-on course · Module 6: Agentic engineering in teams
Team conventions and shared prompts
When five developers each prompt their own way, you get five codebases. Promote what works from personal habit to shared prompt, repo instruction file, then automated check.
Lesson 21 / 24 · ⏱ 8 min
On a team, the hard part of agentic work isn’t one person getting good results. It’s five people getting consistent results. Without shared conventions, every developer’s agent writes code in a slightly different dialect.
The fix isn’t a big policy document. It’s a habit: when something works twice, move it one step closer to the repo.
The promotion ladder
Each rung trades flexibility for reliability. A prompt in the library still needs someone to pick it. A line in the instruction file is read on every run. A lint rule can’t be forgotten at all.
My rule: if a rule can be checked mechanically, it belongs at rung 5. Instruction files are for the judgment calls a linter can’t make.
Where does it belong?
Sort each convention into the rung where it earns its keep. There’s room for argument on a couple of them; the explanation says how I’d decide.
🎮 Place the convention
1 / 8 · Score: 0
Where should this live on a team using coding agents?
Instruction files themselves, what goes in them and how long they should be, are covered in Module 2. Here the question is how a team keeps them alive.
Personal prompts vs shared conventions
✕ Everyone for themselves
- –Each dev keeps prompts in a private notes file
- –The same mistake gets corrected in five chats
- –Agent output style depends on who ran it
- –New hires reverse-engineer the team's habits
- –Nobody knows which prompt is the good one
✓ Shared and versioned
- +Prompts live in the repo, next to the code they touch
- +A repeated correction becomes one line in the instruction file
- +Changes to conventions go through a pull request
- +Each shared prompt has an owner and a one-line purpose
- +Stale prompts get deleted, not collected
A shared prompt worth keeping
Good shared prompts read like small briefs with blanks to fill. The blanks are where the person running it adds judgment.
# prompts/add-feature-flag.md
# Owner: platform team. Purpose: add a flag end to end without forgetting a step.
Add a feature flag named {FLAG_NAME} that gates {BEHAVIOUR}.
1. Register it in the flag config with default OFF.
2. Gate the behaviour at {ENTRY_POINT} only. Don't scatter checks.
3. Add one test with the flag ON and one with it OFF.
4. List every file you changed and anything you couldn't verify.
Don't touch unrelated flags or refactor surrounding code.Before you move on
✅ Key takeaways
0 / 5 completed