How this skill is triggered — by the user, by Claude, or both
Slash command
/operant:development-methodologyThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
<!-- #core -->
Three-agent build methodology where a CEO orchestrator delegates all implementation work to specialized subagents. The CEO never writes code directly -- only delegates, verifies results, and makes decisions.
The dev agent operates as a CEO that spawns three types of subagents:
| Role | Purpose | Writes Code? | Opens Browser? |
|---|---|---|---|
| Maintainer | Infrastructure, planning, task assignment | No (config only) | Yes (manages tabs) |
| Builder | Implements scoped code tasks | Yes | No |
| Police | Functional and visual verification | No | Yes (auditor-browser) |
The CEO never implements code. Maximum one direct edit per cycle, and only for trivial fixes (typos, config values). Everything else is delegated.
Launch order is strict and sequential by role type:
Maintainer first -- Stands up infrastructure before anyone else starts:
Builders second (parallel) -- Launch after Maintainer completes:
Police last -- Launch after all Builders finish:
The Maintainer is the first agent launched and the fastest to complete. Responsibilities:
implementation-spec.md and all files in revisions/ (if any)The Maintainer does NOT write application code. It writes plans and configuration only.
Builders receive a scoped task from the Maintainer and implement exactly what is specified. Rules:
tsc or equivalent for compilation checks, but functional/visual verification is Police-only.browser_evaluate to programmatically inject behavior that bypasses real UI flows.If a Builder is going in circles (stuck, looping, not making progress), kill it and spawn a fresh agent with tighter instructions. Fire and replace is faster than nudging.
Police agents verify that Builder output matches the spec. They run after all Builders complete.
Police Round 2: After fixes are applied, run Police again as cheap insurance. A few targeted tests confirm the fix works. Do not skip this step.
Police must follow tab assignments exactly. If told "use tab 1", verify the correct tab before proceeding.
The CEO (the dev agent itself) orchestrates but does not implement:
The CEO makes decisions: which agents to spawn, when to fire and replace, when to declare dev complete.
If any agent encounters an issue it cannot resolve (missing API keys, unclear requirements, external service down):
.md file to spec/<name>/blockers/ with:
When the dev agent is respawned after an audit failure or user rejection:
spec/<name>/revisions/ as additive contextimplementation-spec.mdThe dev loop is complete when:
browser_evaluate to programmatically inject behavior.The pipeline hooks automatically detect when the dev phase completes by checking for new blockers at the end of each agent turn. When all FRs are implemented and Police passes:
DEV_COMPLETEUse /status to check the current pipeline state if uncertain about where things stand.
npx claudepluginhub esxr/operant --plugin operantProvides behavioral guidelines to reduce common LLM coding mistakes, focusing on simplicity, surgical changes, assumption surfacing, and verifiable success criteria.
Searches, retrieves, and installs Agent Skills from prompts.chat registry using MCP tools like search_skills and get_skill. Activates for finding skills, browsing catalogs, or extending Claude.
Creates, edits, and optimizes skills for Claude Code, including drafting, evaluating with test prompts, iterating on performance, and improving skill descriptions for better triggering accuracy.