From junior-dev
Delegate a junior-level coding task to a local Ollama model to save Claude usage. Use when a task is simple, self-contained, and objectively verifiable — boilerplate, scaffolding, simple pure functions, test stubs, docstrings/comments, repetitive pattern edits, format/regex conversions. Claude triages, writes the spec, delegates generation, runs an objective check, then reports. Do NOT use for tasks needing judgment, broad codebase context, debugging, architecture, or security-sensitive code.
How this skill is triggered — by the user, by Claude, or both
Slash command
/junior-dev:junior-devThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
Offload **generation** of low-judgment, easily-verified code to a local Ollama
Offload generation of low-judgment, easily-verified code to a local Ollama model while Claude stays the orchestrator: triage → confirm → spec → delegate → verify → report. This saves Claude output tokens only when generation is large and the result is cheap to check.
Delegate ONLY if ALL of these hold:
Good fits: boilerplate/scaffolding, simple pure functions, test stubs, docstrings and comments, repetitive edits across a known pattern, format conversions, regex, simple data transforms, config files.
Never delegate: debugging, architecture/design, multi-file changes, ambiguous requirements, security-sensitive code, or anything where "correct" requires reading the surrounding codebase.
If a task mixes junior and senior parts, either do it yourself or split it: keep the judgment, delegate only the mechanical sub-part.
When you spot a junior task, briefly tell the user what you'd delegate and the check you'll verify it with, then ask for a quick OK. Do not delegate silently.
Run the preflight check. Resolve the script path at call time so it works on any
Claude Code version (whether or not the plugin's bin/ is on PATH):
python3 "$(command -v junior-preflight.py 2>/dev/null || find ~/.claude/plugins/cache -name junior-preflight.py 2>/dev/null | head -1)"
If Ollama is unreachable or has no model, tell the user and just do the task yourself.
Write an unambiguous spec for the local model: exact signature, inputs/outputs, constraints, language, and "return ONLY code — no markdown, no explanation." Keep it fully self-contained.
printf '%s' "$SPEC" | python3 "$(command -v junior-delegate.py 2>/dev/null || find ~/.claude/plugins/cache -name junior-delegate.py 2>/dev/null | head -1)"
(The command -v … || find … resolves the script whether or not bin/ is on
PATH.) Returns clean code on stdout (thinking block and ``` fences already
stripped).
Decide thinking per task. Local thinking tokens are ~free (just latency), and a better first pass means fewer expensive Claude-side retries:
Pure boilerplate / mechanical → leave thinking off (default).
Any real logic, edge cases, or correctness nuance → prefix with JUNIOR_THINK=1:
printf '%s' "$SPEC" | JUNIOR_THINK=1 python3 "$(command -v junior-delegate.py 2>/dev/null || find ~/.claude/plugins/cache -name junior-delegate.py 2>/dev/null | head -1)"
Config via env: JUNIOR_MODEL (default: first installed model), OLLAMA_HOST
(default http://localhost:11434), JUNIOR_TEMP (default 0).
Run the check you named in triage — tests / typecheck / compile / lint, or for trivial output a careful read. Never accept the output blind.
JUNIOR_THINK=1 on for the retry). Max 2 retries.Provides 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.
npx claudepluginhub 575-lab/junior-dev --plugin junior-dev