← Nisarg GurjarEmail me
Method

I decide what should exist. AI writes the code. A system of gates decides what ships.

I run software development the way a founder runs a company: AI is the implementation team, I’m the decision layer, and a system of gates has the final say. Every project in this portfolio came out of the same loop: specify → direct the AI → verify against gates → walk the result end to end. The gates are the point. This page is the machinery.

01 — The loop
  1. 1Specify. I write what should exist and why — the behavior, the failure modes, the conventions. This is the part AI can’t do for you.
  2. 2Direct. Claude Code writes the implementation. I review behavior and structure — trained deliberately, with a curriculum I built for exactly this.
  3. 3Gate. Tests must pass. For AI features, adversarial eval suites must pass 100%. Guardrail hooks block the dangerous paths entirely.
  4. 4Walk. Before anything is called done, the real user flow gets walked — logs checked, database verified, the success state actually seen.
02 — The tooling: 12 custom skills

I extended my AI development environment with tooling of my own — repeatable, safe procedures for things that must never be improvised:

/ship-checkWalks the real user flow before anything is called “done” — logs, database, success state.
/deploy-verifyWatches a deploy to completion, then verifies the live page actually renders — not just a 200.
/m4u-sendBakes every business email convention into the send path; forces a preview + explicit YES.
/devteamConvenes 13 engineering personas to critique a change before it ships.
/boardroomConvenes 12 business personas to pressure-test a strategy decision.
/crew-sheetTurns a sales call into a branded crew prep document.
/build-timelineTurns a sales call into a branded event run-of-show.
/bigin-updateReconciles deferred CRM logs — deterministic script, never freehand API calls.
/whats-nextReads every project tracker and recommends the three best next moves.
/goodbyeCloses a session by writing the handoff — status, open items, a resume prompt.
/obsi-cleanFiles captured notes into the knowledge system by intent, never deletes.
/profileSwitches the AI's working profile — different discipline levels for different risk.
03 — The guardrails: 6 hooks against my own AI

The most honest part of the system: hooks that constrain the AI I work with — because a system that emails real clients and touches real credentials must be safe against its own operator’s tools.

secret-guardBlocks the AI from exposing credentials — runs before every tool call.
m4u-send-guardStops real client emails from being sent outside the approved path.
profile-selectForces a working-discipline choice at the start of every session.
skill-obsidian-syncKeeps tooling documentation in sync automatically.
obsidian-indexMaintains knowledge-base indexes after every edit.
plugin-usage-noticeMakes tool usage visible instead of silent.
04 — The gates that caught what tests couldn’t

Ordinary tests prove code does what code says. They can’t prove an LLM won’t invent a fact, follow an injected instruction, or misread a decoy. So every AI feature here ships behind an adversarial eval gate: golden scenarios built to trick it — prompt injection, decoy numbers, calls that only look like sales calls, drafts that bait invented details. The gate must pass 100% before go-live.

Twice, those gates caught failures that every unit test missed. That’s why they exist.

05 — The honest limits

My review isn’t a career engineer’s review — the method doesn’t depend on it being one. The controls are structural: smaller changes, more tests, adversarial gates, guardrail hooks, and walking every result. When something genuinely needs an expert human — security review, novel architecture — the right move is bringing one in, and I say so. What I bring is different: real, verified software shipped at a speed and breadth one person otherwise couldn’t reach — and every claim on this site is checkable.

Want something like this built?

I take on projects where shipping something real — tested, verified, in production — is the whole point.

Email me