Planning-first workflow for coding agents: turn a rough idea into a spec, record architecture decisions, and break the work into small verifiable tasks before any code is written.
A setup for AI coding agents, made by Kitloom. One source, generated for every major agent: Claude Code, OpenAI Codex, Cursor, Google Antigravity, Gemini CLI, GitHub Copilot, Windsurf (Devin), opencode and Aider.
- Download this repository as a ZIP, or use the Check and download button on its Kitloom page, which checks it first.
- Read the files. Nothing here runs on its own: the kit has no hooks and no install scripts.
- Copy the folders and files you want into the root of your project.
AGENTS.mdis the shared instructions file; edit it to fit your project.
grill-with-docs— Stress-test a vague plan with a relentless interview that also records ADRs and a glossary as you go.idea-refine— The idea is still fuzzy; diverge and converge on a sharp concept before writing a spec.interview-me— The request is underspecified and the real intent has to be extracted one question at a time.spec-driven-development— Write the PRD or spec (objectives, scope, acceptance criteria) before any implementation starts.to-spec— Synthesize an already-discussed conversation into a written spec without another interview.architecture-decision-records— A significant technical choice needs a durable, well-structured ADR with alternatives and consequences.planning-and-task-breakdown— Split an approved spec into ordered, estimable tasks with clear acceptance criteria.to-tickets— Turn a spec into thin tracer-bullet tickets that declare what blocks what.wayfinder— The work is too large for one session; map it as decision tickets and resolve them one at a time.incremental-implementation— Pick up the next task from the plan and deliver it as a thin, verifiable slice.
Invoke by name (for example /new-spec).
new-spec— Interview the user and write a spec to docs/specs/.md.decide— Capture one architecture decision as a numbered ADR in docs/adr.break-down— Turn an approved spec into an ordered task list in docs/plans.
spec-reviewer— Delegate to critique a draft spec for ambiguity, missing acceptance criteria and untestable claims.plan-auditor— Delegate to check that a task list fully covers its spec, is ordered correctly and has no oversized tasks.
spec-before-code— Always on: no implementation work starts until a written spec and task list exist for non-trivial changes.planning-docs— Conventions for files under docs/specs, docs/adr and docs/plans.task-sizing— How to size, order and write tasks in plan and task-list files.
| Agent | Instructions | Skills and commands | Rules | Sub-agents |
|---|---|---|---|---|
| Claude Code | CLAUDE.md (imports AGENTS.md) |
.claude/skills/ |
.claude/rules/ |
.claude/agents/ |
| OpenAI Codex | AGENTS.md |
.agents/skills/ |
listed in AGENTS.md |
.codex/agents/ |
| Cursor | AGENTS.md |
.agents/skills/, .claude/skills/ |
.cursor/rules/ |
.claude/agents/ |
| Google Antigravity | AGENTS.md |
.agents/skills/ |
.agents/rules/ |
.agents/agents/ |
| Gemini CLI | AGENTS.md via .gemini/settings.json |
.agents/skills/, .gemini/commands/ |
listed in AGENTS.md |
.gemini/agents/ |
| GitHub Copilot | AGENTS.md |
.agents/skills/, .github/prompts/ |
.github/instructions/ |
.claude/agents/ |
| Windsurf (Devin) | AGENTS.md |
.agents/skills/, .windsurf/workflows/ |
.windsurf/rules/ |
— |
| opencode | AGENTS.md |
.agents/skills/, .opencode/commands/ |
listed in AGENTS.md |
.opencode/agents/ |
| Aider | AGENTS.md via .aider.conf.yml |
— | — | — |
Skills are stored twice (.agents/skills/ and .claude/skills/) because Claude Code reads only its own folder. Agents that read both may list a skill twice; delete the copy you do not need.
The skills come from these open-source repositories, each under its own licence. Their licence and a notice are kept beside every copied skill.
| Skill | Source | Licence |
|---|---|---|
grill-with-docs |
mattpocock/skills at 49dd158 |
MIT |
idea-refine |
addyosmani/agent-skills at aeac41d |
MIT |
interview-me |
addyosmani/agent-skills at aeac41d |
MIT |
spec-driven-development |
addyosmani/agent-skills at aeac41d |
MIT |
to-spec |
mattpocock/skills at 49dd158 |
MIT |
architecture-decision-records |
wshobson/agents at 46891e7 |
MIT |
planning-and-task-breakdown |
addyosmani/agent-skills at aeac41d |
MIT |
to-tickets |
mattpocock/skills at 49dd158 |
MIT |
wayfinder |
mattpocock/skills at 49dd158 |
MIT |
incremental-implementation |
addyosmani/agent-skills at aeac41d |
MIT |
Everything else in this repository (instructions, rules, commands, sub-agents) is by Kitloom under the MIT licence.
Read every file before you use it. Agent instructions can make an agent run commands on your machine. This kit was checked with Kitloom's pre-download check, which looks for warning signs but cannot prove anything is safe.