CTO plans once, then delivers gate by gate
BeforeOne agent gets the whole plan and drifts
AfterWorkers get one brief each; the next gate starts only after this one passes
Each worker stays on its own small job, and nothing moves on early.
Plan once, then ship gate by gate. A separate agent checks every gate, and the next one starts only after it passes.
/plugin marketplace add nhattrung0911/cto-team /plugin install cto-team@cto-teamGitHub
Workers build, a verifier checks, a fixer repairs, then the gate passes.
BeforeOne agent gets the whole plan and drifts
AfterWorkers get one brief each; the next gate starts only after this one passes
Each worker stays on its own small job, and nothing moves on early.
BeforeFrontend and backend each guess the API
AfterOne contract file the CTO owns; workers may not edit it; the verifier checks the app against it
Both sides build to the same API, and the verifier checks the app against it.
BeforeYou hand-assign agents; two agents edit the same file
AfterYou pick 1–5 Sonnet workers per gate; the CTO gives each disjoint files
You only choose the team size; the CTO keeps workers out of each other’s files.
Before“Tests pass”, says the agent that wrote them
AfterA separate agent proves every acceptance criterion, runs the full suite, and breaks one line the tests must catch
The check comes from an agent that did not write the code.
Beforeassert True slips through; huge input crashes
AfterFake tests fail the gate; every number gets an overflow test
Fake tests are stopped at the gate, and every number is tested for overflow.
v0.1 crashed on a 20-digit amount; fixed from v0.2
BeforeBugs across features are found by users
AfterEvery 3–4 gates Opus reviews the batch
Opus reviews several gates together, not just one at a time.
Found a real race condition the per-gate checks missed
BeforeLong-lived “member” agents grow huge
AfterEvery agent is one-shot and starts small; Sonnet does most of the work
Small, one-shot agents keep your token cost down.
12k vs 42k start context; 0.9M vs 3.8M tokens per agent
BeforeSlow recursive grep, noisy output
AfterA hook points agents to ripgrep (faster, respects .gitignore)
Searches run faster and skip what .gitignore excludes.
BeforeContext reset = redo or lost work
AfterState lives in docs/cto/; delivery resumes from the last checkpoint
A reset does not cost you the work already done.
BeforeA long chat log
AfterOne HTML report: what works, how to try it, tech stack, architecture diagram, known gaps
You read one page instead of scrolling a chat.
Quality reference: on a 10-task test app every version passed 14/14 hidden tests.
Paste these two lines into Claude Code.
/plugin marketplace add nhattrung0911/cto-team /plugin install cto-team@cto-team
Tell Claude what you want to build. The CTO plans it with you once.
/cto-team:plan
Clear the chat, then hand over the plan. Choose 1–5 Sonnet workers per gate. Delivery runs gate by gate, then you read one report.
/clear /cto-team:deliver <plan>