Practical workflows for delegating, reviewing, and constraining AI coding agents.
I added an LLM review stage to an unattended publishing pipeline and committed it as verified. The verification: five Japanese posts, all approved. Here is what an all-approve result proves about a gate, and what it does not.
I went after the question of where the 0.62% install-to-purchase funnel dies. A code search produced fourteen matching apps. Exactly one was an actual defect.
I picked five new APIs from the latest SDK and wired them into several apps. All five were reverted — and two days later the three I had marked 'confirmed' collapsed too.
I ran a code review at the end of every task. All of them passed. A single review over the whole thing found blockers — the defects lived between tasks, not inside them.
I knew the shared auth trigger writes a profile row into every app's schema. Then I used exactly that table as the denominator for a retention cohort.
Typing Korean through the input action lands some syllables as different characters. The tool reports success, and the corrupted characters appear in the returned log exactly as they landed. Reading the log cannot catch it.
Two AI sessions were running in one repo. I staged 73 files and was writing the commit message when the other session committed its own work. All 73 of mine went out inside that commit.
I ran a script that prints how long each app has been waiting in review and one app looked far older than the rest. That date wasn't the submission date, it was the version creation date. The real wait was 20 hours — and I had already attached the word 'unusual' to it.
I launched a code review agent in the background and kept editing the same files. The result no longer referred to any tree that existed on disk.
I filtered on length >= 15 to catch label echoes. "Write 40 to 80 characters, at most one emoji" is also longer than 15 characters. That wasn't a hook — it was the instruction I gave the model.
To stop prompt injection through a name field I banned periods. Commands without a period had been getting through all along.
I recolored 26 apps automatically and ran a WCAG contrast audit across every route. Zero failures. Then a user said the text was barely visible. What the audit never measured was not contrast — it was which color was paper and which was ink.
The full concept prompt and background asset prompt used to redesign damta.ootssu.com toward the bungeo.ootssu.com quality bar, plus the constraints that made the result shippable.
The full text of the three prompts I use to generate iOS, Android, and game UI. Then I grepped the shipped code to see how many of those rules actually made it. Three rules survived. Three were quietly ignored.
Every day I fed bot state to an LLM and got a five-line summary. For three weeks it said 'running to plan.' Over the same window, simply holding the same budget would have returned 20pp more. The model never lied — my JSON just had nothing to compare against.
I autonomously delegated mechanical tasks to a CLI agent, and instead of working it opened a web mockup or wrote a plan document and exited. The agent hadn't failed — it was doing a different job: an auto-loaded skillset read 'build something' as brainstorming. But the guard that fixes this is useless for analysis work.
A prompt that costs nothing stripped two rounds of self-deception this week alone — a traffic dashboard and a self-improvement gate. It works because it's not a clever prompt, it's a forcing function.
The biggest lever on a coding agent wasn't the model or MCP — it was a small library of prompts and skills that re-fire every turn. A brevity mode, a YAGNI mode, an investigate-first contractor prompt, and a one-word go-autonomous signal.
A two-agent workflow: Claude orchestrates (plan, verify, deploy), Codex generates (design, content). The handoff is a manifest, and the whole thing lives or dies on verification gates.
A fan-out research workflow with adversarial verification: many search agents, then a 3-vote refutation gate. I pointed it at 'is YouTube automation monetization real?' — and it killed most of the claims I wanted to believe.
One spec, handed to Codex, covering a full design-system redesign across dozens of apps — then batch-deployed by an orchestrator with a regression gate on the last app before shipping.
How I let an agent generate content in bulk without free-running it: a SPEC.md with schema and quality bar, one verified pilot batch, then parallel batches gated by a strict schema parse.
The single biggest quality lever with coding agents isn't the model — it's a process that forces brainstorming before design, a written plan, and verification before any 'done' claim.
What actually using Model Context Protocol servers looks like in a real workflow — driving a browser, capturing result cards with Playwright, running Supabase DDL, and the gotchas each one has.
One fact per file with frontmatter, an index loaded each session, and links between notes. A simple, greppable memory that lets an agent carry context forward without a database.