Best Claude Code Skills for Planning in 2026

grill-me is the most-installed planning skill on skills.sh, but its SKILL.md is only a few lines long. It tells the agent to call another skill, grilling, from the same repository (1.0M installs). Its disable-model-invocation: true line means the agent never loads it on its own. You type it when you want it. Install grilling too, or install only grilling, as in the command above.

grilling is where the method lives. The agent maps your plan as a tree of decisions. Each round, it asks every question it can ask without guessing an earlier answer, numbers them and gives its own recommended answer for each. You reply, the tree changes, and the next round starts. Facts it can find in your code or tools, it sends a sub-agent to look up. Decisions stay with you.

The session ends only when every branch is visited and you confirm a shared understanding. That makes it slow on small tasks, which is the point.

This skill comes from Superpowers, Jesse Vincent's collection of workflow skills. Its description tells the agent to use it before any creative work, in software or not. The first step is a check: if everything needed is in the request, the agent just does the task. If something important is missing, it starts a conversation.

Before writing any code, the agent must lay out the possible approaches and let you pick one, write a design doc (short or full), and get your approval. The doc is saved under docs/superpowers/specs/ and committed. In an existing project, it reads the code and files you pointed at before asking anything.

It can also open a browser tab to show mockups and diagrams. That "visual companion" runs a small local server, and the Skillpack CLI lists those scripts as executable files before copying anything. When the doc is approved, it hands off to writing-plans (286.4K installs), which turns the design into bite-sized tasks. Install both.

Addy Osmani's skill runs four gated phases: specify, plan, tasks, implement. A human reviews each one before the next starts. It also says when to skip it: one-line fixes, typos and changes whose requirements are already clear.

The spec template covers objective, tech stack, commands, project structure, code style, testing strategy, boundaries, success criteria and open questions. The plan goes to tasks/plan.md and the task list to tasks/todo.md. Each task gets acceptance criteria, a verification step and the files it touches, and no task should change more than about five files.

It leans on other skills from the same repository. Task breakdown is delegated to planning-and-task-breakdown (65.7K installs), and implementation to its incremental and TDD skills. Install those too if you want the full flow.

This PRD skill comes from awesome-copilot, GitHub's community collection for Copilot. It's a plain SKILL.md, so it works in Claude Code too. Before writing anything, the agent must interview you about the core problem, success metrics and constraints such as budget, stack or deadline.

The document follows a fixed five-part schema: executive summary, user experience and functionality, AI system requirements when they apply, technical specifications, and risks and roadmap. A quality section asks for measurable criteria and shows vague and concrete versions side by side.

What sets it apart is the AI features section. If you're specifying an LLM-powered feature, it asks for the AI requirements and an evaluation plan. For a small change, the schema can feel heavy.

create-prd is part of Paweł Huryn's pm-skills, a collection for product managers. It's lean by default: about one to two pages, roughly 600 to 1,000 words. You get a longer version only if you ask for it.

If the problem, the target users or the success metric is missing, the agent stops and asks up to three short questions, with no PRD in that reply. It can search the web for market and competitor context, and cites its sources. The template has eight sections, including market segments and a six-part value proposition based on jobs to be done.

Its rules against padding are the part we liked most. No "TBD", no placeholder roles, and nothing invented. It can also review an existing PRD.

This skill is part of Anthropic's product management plugin in knowledge-work-plugins. It casts the agent as a sharp product thinking partner. The skill tells the agent to push back and bring unexpected angles, rather than churn out deliverables.

It has four modes: problem exploration, solution ideation, assumption testing and strategy exploration. Each lists what to do and which questions to ask. It also describes frameworks the agent can pull in, including How Might We, jobs to be done, opportunity solution trees, first principles and SCAMPER.

It sits one step earlier than the rest of this list. Use it to decide what's worth building. Then use a spec or PRD skill to write it down.

ce-plan comes from Every's Compound Engineering plugin. It splits the work cleanly: ce-brainstorm decides what to build, ce-plan decides how, and ce-work carries it out. ce-plan never implements. It researches, decides and writes the plan.

It first sizes the request. A small change gets a few sentences in chat, bounded work gets a short brief, and anything with real decisions or risk gets a full plan file. Authentication, payments, migrations and external contracts always get the full plan. It asks one question at a time.

It's the most involved skill here. The SKILL.md sends the agent to several reference files in the folder, and the folder includes scripts that the Skillpack CLI lists as executable. It also assumes the rest of the plugin for the steps before and after.