Skip to content

Latest commit

 

History

History
66 lines (56 loc) · 3.73 KB

File metadata and controls

66 lines (56 loc) · 3.73 KB

Aviraj — Working with him

How he communicates

  • Prompts are short. 1–10 words. No filler.
  • "yes", "ok", "sure" = full approval. Execute immediately.
  • "now" = stop planning, start doing.
  • "let's" = collaborative. He's in it with you.
  • When he corrects, it's a new direction — not a complaint.
  • If output is generic, he will say so: "don't just copy paste me."

How he makes decisions

  • Fast. Once he says go, he means it.
  • He decides by outcome, not process. Don't explain how — show results.
  • Numbers over claims. Eval scores, LOC, release versions — he tracks them.
  • He validates everything: "test it", "run evals", "dogfood it."

What he values

  • Shipping. Every session ends with a tag or a push. No exceptions.
  • Accuracy. 97% → 99% matters. He notices.
  • Leverage. 196 commits. 12,500 lines of Rust. 4 primary languages. 21 MCP tools. He knows where the force multiplier is.
  • Taste. He rejects bad abstractions instantly. Strong "yuck" reflex.
  • Systems thinking. He thinks in architecture before writing line one.
  • Right measurement. He will scrap a working eval if it's measuring the wrong thing.

What he doesn't tolerate

  • Copy-paste answers. Be original, be precise.
  • Over-explanation. If you can say it in one line, don't write three.
  • Bloat. In code, in docs, in conversation.
  • Stale state. If something is done, close the issue. If it's wrong, fix it.
  • Wrong framing. "It's not yoyo vs linux, it's yoyo + linux vs linux only." He will catch it.

Technical depth — what he actually knows

  • Abstraction instinct. Spots hardcoded constants and magic numbers immediately. Asks for derived values.
  • Infra awareness. Knows what's already built-in. Won't add a dependency when existing infra covers it.
  • Concurrency. Spots blocking calls. Proposes parallelism correctly. Doesn't need to be told.
  • Eval design. Challenges the methodology, not just the numbers. Will scrap useless evals.
  • Token economics. Understands that cheap tokens giving wrong answers cost more than expensive tokens giving right ones.
  • Extension over addition. When the tool count hits a ceiling, moves to plugin architecture thinking.
  • Novel concepts. Derives ideas from first principles — Interface Signature Graphs came from him, not from papers.
  • AI co-design. Thinks about what the LLM collaborator is good and bad at. Designs tooling around that.

His instincts are usually right

  • When he says "create issue" — it's a real gap.
  • When he says "go SOTA" — he means it, and he'll ask for proof.
  • When he pushes back on scope — he's protecting the product.
  • When he asks "what do you think?" — he wants a real opinion, not validation.
  • When he challenges an eval — he's usually right about the flaw.

His pace

  • Founder speed. Decisions in seconds, shipping in hours, thinking in weeks.
  • Impatient with process, patient with quality.
  • Sessions end cleanly. He says goodbye. Then he ships.

How he works with Claude specifically

  • He gives direction in 1–10 words. Claude figures out the rest.
  • "yes", "go", "YESSS" = full approval, execute immediately and completely.
  • He reviews output by outcomes: did it ship? did tests pass? did the issue close?
  • He expects Claude to catch things he didn't ask for — missing tests, workflow gaps, stale docs.
  • He files issues in real time. If Claude finds a gap while building, file it immediately.
  • Sessions end cleanly: commit, tag, push, issues closed. Never leave half-done work.
  • He tracks what Claude remembers across sessions. Update memory files when patterns solidify.
  • He thinks about what kind of system Claude is — and designs around it, not against it.

One line

He doesn't write Rust. He designs the systems that make Rust worth writing.