Anthropic cut new-hire technical onboarding from 3 weeks to 2 days—not by hiring smarter people, but by building a better “context engineering” system.
Today, Boris Cherny (head of Claude Code) publicly shared the workflow their team uses every day. I cross-checked it against my own quant system and found a serious gap.
━━━ The most valuable one ━━━
On the very first day of a new project, do not write a single line of code—decisively.
First, have the AI translate the Git history and figure out why this codebase looks the way it does. Why is the API designed like this? Why does this class exist? What known pitfalls are there?
Against my own system: whenever a new AI comes into a fresh session, I have to re-understand the 4,000 lines of brahma_core.py, re-guess the architectural taboos, and step through the same pitfalls that were already stepped through before. The waste isn’t just time—every decision is made using worse context each time.
The fix is simple: place a CLAUDE.md in the root directory of the module and clearly document:
→ the core entry-point workflow of this module
→ the test commands that must be run before making changes
→ the architectural taboos you must never touch
→ the pitfalls you’ve encountered in the past
Boris said this step reduces new-hire ramp-up time from 3 weeks to 2 days. Today I added this file to my own system, and it feels like leaving a truly useful instruction manual for a stranger.
━━━ Another one worth remembering ━━━
After correcting a mistake once, write the rule into the system immediately. Don’t rely on “remember it next time”—you won’t remember next time.
This isn’t prompt-engineering trickery; it’s an engineering habit. Step into a pit once and you lose money. Step in again and it’s just waste.
Correct once, compound forever.
#Claude #AI工程 #quant trading