From agent-loop
Engineering standards every code-writing tick must follow so the loop produces GOOD code, not just working code. Invoked by build/maintenance ticks before and while editing code. Composes with (never overrides) the host project's CLAUDE.md / AGENTS.md.
How this skill is triggered — by the user, by Claude, or both
Slash command
/agent-loop:code-standardsThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
Write code a senior engineer would approve — not just code that passes. Apply on every build/maintenance tick
Write code a senior engineer would approve — not just code that passes. Apply on every build/maintenance tick
that touches code. These are the floor; the host project's CLAUDE.md / AGENTS.md wins on any conflict.
State your assumptions. If the goal has multiple readings, pick the simplest that fits and say which — don't silently guess. If a simpler approach exists, take it. Surface real tradeoffs in the journal, up front.
Before writing code, walk this ladder and stop at the first rung that solves it:
No abstractions for single-use code, no "flexibility"/config nobody asked for, no boilerplate nobody requested.
Deletion over addition; boring over clever. Test: would a senior engineer call this overcomplicated? If yes, cut it.
Lazy, not negligent — simplicity NEVER cuts trust-boundary validation, data-loss handling, security, or
accessibility. When you take a deliberate shortcut, mark it inline: // shortcut: <why> — upgrade: <path>, so
the minimal choice stays visible and reversible, not hidden debt.
Touch only what the goal needs. Match the surrounding style even if you'd do it differently. Don't refactor what isn't broken or "improve" adjacent code/comments/formatting. Remove only the orphans YOUR change created; mention pre-existing dead code, don't delete it. Every changed line traces to the goal.
Turn the task into a checkable success criterion (a test, a passing command). Write/extend the test, make it
pass; the suite is green before AND after. This feeds the forced verify gate — you cannot mark work done
without a real check (backlog_done requires a passing verify block).
Prefer LSP (go-to-definition, find-references, types) over blind grep for navigation. After editing, verify
no type/build errors (tsc --noEmit / cargo check / the project's check) before claiming done.
Match the repo's existing commit style and conventions (message format, and whatever attribution/footer policy the repo already uses — follow the project, don't impose one).
If the host project has a CLAUDE.md / AGENTS.md, read and FOLLOW it — it takes precedence. This skill is the
baseline that travels with the framework (and applies on agents that don't read CLAUDE.md, like Codex).
npx claudepluginhub andonmitev/agent-loop --plugin agent-loopProvides 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.