Case study · 2025 – now · ongoing
My agent toolkit
The skills, subagents and hooks that let one engineer review several agents without lowering the bar.
~ $ ls skills/
hebert-rn-archhebert-rn-featurehebert-rn-testinghebert-rn-navigationhebert-fullstackhebert-ship
~ $ claude --agent reviewer
◐ 3 diffs waiting for hebert
~ $
My roleAuthor, only user
Platforms ● Agents
Built withClaude Code · Codex · MCP · Hooks
Source [Repo or write-up link]
The problem
Agents write fast. Someone still has to say no.
[What went wrong before you had house rules: the kind of mistakes agents made, and what reviewing everything by hand cost you.]
Decisions
The calls I made.
Skills per job, not one giant prompt
Architecture, features, tests, navigation and releases each have their own skill, so an agent loads only the rules it needs. [An example rule.]
Every change waits for me
Agents plan, build and verify; a human approves. Commits and pull requests go through dedicated subagents with fixed formats. [Why that matters to you.]
[A call you’d make differently]
[One honest trade-off.]
+ agent: drafted the change
+ agent: tests pass on both targets
− hebert: not like this. see note.
◐ waiting for hebert
Built with agents
Where agents helped, and where I stepped in.
Agents did[What the toolkit automates today.]
I did[What you still never delegate.]
I overrode[The rule you added after an agent got something wrong.]
[N] skills in daily use
[N] projects that use them
[N] [a number you’re proud of]
