Subagent Staffing Planner
Decide which recurring tasks deserve a Claude Code subagent versus a skill, script or manual work, and draft least-privilege .claude/agents files. Takes your list of recurring tasks and records frequency, input size, output, side effects and pain for each. It scores them on six criteria (frequency, context size, isolation, tool risk, parallelism, simpler option), assigns a bucket (Agent, Skill or saved prompt, Script or CI job, Stay manual, Merge) and rejects anti-patterns such as one agent per task, god-agents and write tools by default. For tasks that earn an agent it drafts .claude/agents/<name>.md files with an explicit read-only tools line, a Does not do list and a fixed report format, plus delegation examples and a final staffing table. It only produces drafts in the chat: it creates no files, runs nothing, calls no service and asks for no secrets. Saving files and granting tools is your decision.
Example input
Use the Subagent Staffing Planner on this (a sample I wrote). I maintain a small Node web app in a repo at /work/example-shop and want to know which of these recurring tasks deserve a Claude Code subagent. Please draft files for the ones that do. 1. Weekly: scan all dependencies for outdated or risky packages and give me a short list (about 400 package files and a lockfile; I only need a table back). 2. Daily: format code and fix lint errors before commits (currently I run it by hand). 3. Per pull request, several times a week: review the diff for missing tests and unsafe input handling. Diffs are often 20+ files; I only want a short report. 4. Monthly: read the last 30 days of server error logs (very long) and summarize the top recurring errors. 5. Occasionally: rename a feature flag across the codebase, which needs careful judgement and edits. 6. Release day, twice a month: follow our release checklist step by step with me in the chat. All reports go to me only. I have never created a subagent before, and my .claude/agents folder is empty. No secrets should end up in any file.
Example output
## Evidence | # | Task | Frequency | Input size | Output | Side effects | Pain | |---|---|---|---|---|---|---| | 1 | Dependency scan | weekly | ~400 package files + lockfile | table | none | unknown | | 2 | Format + lint fix before commits | daily | changed files | fixed files | edits files | run by hand | | 3 | PR review (tests, input handling) | several per week | 20+ file diffs | short report | none | unknown | | 4 | Error-log summary | monthly | 30 days of very long logs | top recurring errors | none | unknown | | 5 | Rename a feature flag | occasional | whole codebase | changed files | edits files | unknown | | 6 | Release checklist with me | twice a month | checklist | steps done in chat | unknown | unknown | Frequency and output are known for all six tasks, so no questions are needed. ## Rubric scores | Task | 1 Freq | 2 Context | 3 Isolation | 4 Tool risk | 5 Parallel | 6 Simpler | |---|---|---|---|---|---|---| | 1 Dependency scan | + weekly, same shape | + reads ~400 files, reports a table | + notes would clutter chat | + read/search only | + per package file | 0 a script lists versions but judging "risky" needs reading | | 2 Format/lint | + daily | - needs only the changed files | - nothing to isolate | - needs write/shell | - sequential | - formatter or pre-commit hook does it | | 3 PR review | + several per week | + 20+ files in, short report out | + fresh eyes, notes stay out | + diff is read-only | + per file or module | 0 no script judges unsafe input handling | | 4 Log summary | 0 monthly, same shape | + logs far longer than report | + | + read-only | + per log file | - a counting script gives top errors | | 5 Flag rename | - occasional | - | - needs conversation | - broad edits | - | - editor refactor or search-replace | | 6 Release checklist | 0 twice a month | - checklist is short | - needs the chat | - | - strictly sequential | - a skill or saved prompt fits | ## Buckets | Task | Bucket | Reason | |---|---|---| | 1 Dependency scan | **Agent** | criteria 1, 2 and 3/5 are `+`, 6 is not `-` | | 2 Format/lint | **Script or CI job** | deterministic, no judgement; pre-commit hook or CI step | | 3 PR review | **Agent** | criteria 1, 2 and 3/5 are `+`, 6 is not `-` | | 4 Log summary | **Script or CI job** | criterion 1 is not `+` (monthly) and a counting script covers the fixed-format report; ask the main session for a one-off read of the top few lines only if the script output looks odd | | 5 Flag rename | **Stay manual** | rare, edits everywhere, a mistake is costly; do it yourself in the main session with the test suite | …
Real output from running the skill on a sample problem, shortened. The skill file itself is not published here.