← Index / AI method
How I work with AI
I use AI agents daily — for generation, scaffolding and exploration. What I ship is what I've read, audited and pushed through a gate. This page is the method, not a claim: the pipeline, the audit checklist, the testing discipline, and exactly where the boundary sits between what the model may write and what stays mine.
tools · Devin · Google Antigravity · Claude Code · MCP — parallel agents, personal ownership of architecture and review
“AI writes the draft. I own the diff.”
01 / Division of labor
Who writes what
Every AI-assisted project starts with a written contract — an AGENTS.md that names the zones. The rule is simple: generation is cheap, correctness is not. AI gets the volume; I keep the parts where being wrong is expensive.
AI writes
- Scaffolding, DI wiring, boilerplate
- Migrations, dashboards, docker-compose
- Docs, load generators, test harnesses
- First drafts of anything reviewable
I write
- Idempotency and dedup boundaries
- Retry scheduling, backoff, DLQ semantics
- Backpressure and overflow policies
- Transaction scopes and signing logic
- Anything where a wrong answer costs more than no answer
live example — the webhook engine's AGENTS.md: dedup, retries, rate limiting and DLQ are hand-written; ASP.NET plumbing, SignalR, migrations and the React dashboard are agent work.
The flow runs the other way too — AI reviews my hand-written code. In the HFT engine, model-suggested memory-layout and cache-line tweaks on C++ hot paths were adopted where they actually measured faster; the ones that didn't, weren't. Optimization is a measurement, not a suggestion — whoever wrote the diff.
02 / Pipeline
Draft to ship
Every generated diff runs the same gauntlet. The loop back from audit to rework is the point — most AI output fails it at least once, and that's the system working, not failing.
03 / Review
The slop audit
Standing assumption: the code is hollow until proven otherwise. AI output looks clean — plausible names, passing tests — while doing nothing underneath. So I run a census, not a vibe check:
- 01Endpoints returning bare `dict` — no contract, count them
- 02Functions returning the same value for any input — stubs
- 03`random` inside service files — fake metrics, fake scores
- 04`assert isinstance()` / `callable()` in tests — tests that verify nothing
- 05`except Exception` without logging — silent failures
- 06Module-level mutable globals — unbounded state
- 07`# type: ignore` / `TODO` / `pass` bodies — the debt census
- 08Flat directories, god files, README claims that don't survive a grep
then trace the data: if it's hardcoded on the way in, the service is fake.
a real find — S12, forgeai codebase
POST /api/templates: body.id and body.type flowed straight into join() — zero sanitization. Result: path traversal with arbitrary .json write (could clobber package.json), forged index entries, and read-back through the GET route. The code looked fine, the tests passed, the feature "worked". Found by the census, fixed and documented.
receipts — 113 findings closed across 28 audit rounds on that codebase; a second private SaaS board sits at 546 closed. Auditing AI output is a repeatable practice, not a one-off.
04 / Testing
Rounds, not passes
A guarantee isn't proven by one green run. Each boundary gets repeated rounds of adversarial cases — and during the runs I'm hunting weaknesses, not collecting checkmarks.
rounds
Every boundary gets multiple rounds of cases, not one happy path — edge inputs, race timing, partial failures, replay.
chaos
Webhook engine ships a mock receiver with failure modes — 500 / 429 / slow / flap / drop — fired at the retry heap and DLQ until the guarantee holds.
hunt
Test runs are for finding weaknesses, not confirming the code works. A green suite that found nothing is a wasted round.
honest
Assertions check behaviour and invariants. A test that can't fail isn't a test — it goes in the slop count.
05 / Artifacts
The receipts
Tool · public
GuardDOG
The CI gate itself: regex + AST + optional LLM layers scanning for AI slop. Hash-cached per-file, exit codes for pipelines, markdown/JSON/HTML reports.
repo ↗Kit · unpublished
agents-workflow
My deployable agent-workflow kit: per-repo adoption (small/mid/big), verified gate commands, session memory in .devin/, spec/tasks/audit templates.
this site's .devin →Method · document
Slop audit
The checklist below written as a reusable workflow — drop it into any project, run the census, document findings.
the checklist ↓Contract · per repo
AGENTS.md splits
Every AI-assisted project gets a written division of labor — like the webhook engine's 50/50 split: AI owns plumbing, I own guarantee boundaries.
see /work ↗06 / Contact