A personal kit for working with Claude Code, Codex, and other coding agents.
Install via the plugin marketplace from inside Claude Code:
/plugin marketplace add drimchansky/agents-kit
/plugin install agents-kit@drimchansky-agents-kit
For local development against an unreleased clone: claude --plugin-dir /path/to/agents-kit.
For Codex or non-plugin install of Claude Code, see docs/manual-install.md.
Skills are organized into two groups: an engineering workflow (9 skills) that shapes the development loop, and a set of utilities (3 skills) for ad-hoc tasks.
The workflow runs roughly in order, but you don't need every step — pick whichever fits the task.
explore— Understand existing code or context before changing it. Example:/explore how does the retry queue work?refine-idea— Sharpen a rough idea before planning — assumptions, MVP scope, "Not Doing" list. Example:/refine-idea add a draft mode to the editorresume-task— Brief on, resume, or hand off a task — read.agents/tasks/<slug>/and report in chat. Example:/resume-task auth-jwt-migrationplan-task— A change is non-trivial and needs a contract. Writes paired.agents/tasks/<slug>/<task-slug>.spec.md(acceptance criteria) and<task-slug>.plan.md(steps). Example:/plan-task migrate auth to JWTreview-plan— Confirm the implementation direction is right and still in sync withCONTEXT.md, the spec, and the current codebase; surface any drift between plan assumptions and code reality. Example:/review-plan auth-jwt-migrationimplement-plan— A validated plan is ready to ship. Marks steps[x], writes a*.result.md, and runs an acceptance gate against the spec before flipping the plan todone. Example:/implement-plan auth-jwt-migrationreview-code— Code is written and needs an audit before merge — bugs, blast radius, pattern fit. Example:/review-codeverify-issue— A reported bug needs to be confirmed and root-caused before a fix. Example:/verify-issue users see 500 on signupreview-docs— Audit a doc against the codebase and surface stale references, gaps, and drift. Applies fixes only when you ask after seeing the review. Example:/review-docs README
refine-idea, plan-task, review-plan, and implement-plan share a directory-based contract that lets them hand off cleanly; resume-task reads from the same directory but never mutates it. Each effort lives in .agents/tasks/<slug>/ with a shared CONTEXT.md, paired *.spec.md + *.plan.md files, and append-only *.result.md records. The spec carries the acceptance criteria; implement-plan runs an acceptance gate against it before flipping the plan to done.
See docs/task-directories.md for the full contract.
The kit ships domain checklists in references/engineering/ that engineering skills (especially implement-plan and review-code) consult when relevant.
See docs/reference-checklists.md for the list and conventions.
Standalone skills that aren't tied to the engineering loop:
proofread— Polishing a message, email, or piece of writing for grammar, clarity, and factual accuracy. Example:/proofread(paste text)translate— Moving content between languages while preserving tone and context. Example:/translate to Spanishfact-check— Verifying factual claims against trustworthy live sources on the internet — not against pretraining. Example:/fact-check(paste claim)
For details on how skills load shared rules and how setup.sh orchestrates the install, see docs/how-it-works.md.