| name | deep-planning |
| description | Use when formulating the best technical approach for a task โ before writing implementation plans or code. Triggers on architecture decisions, complex features, refactors, or when the user asks how to approach something. Investigates current codebase, questions existing patterns, researches alternatives, and presents approaches with trade-offs. |
/deep-planning - Deep Planning
Overview
Find the best approach to a task โ not just a good one. Investigate the codebase deeply, question whether existing patterns are optimal, research how the industry solves this, and present approaches grounded in reality with honest trade-offs.
This skill answers: "What is the BEST approach?"
Usage
/deep-planning [task, feature, refactor, or design decision]
/deep-planning --scope .work/dev/5096/
--scope <path>: When provided, read goals.md from the given directory as the task input instead of expecting the task description in the prompt. The goals document contains success criteria, verified claims, user experience requirements, code landscape, and constraints โ use all of these to inform your planning. When --scope is absent, the skill works exactly as it does today (interactive, prompt-driven).
- If adjacent skills like
brainstorming or writing-plans are available, use them as complements.
- If they are not available in the current environment, continue with this skill. It is self-contained.
Related Skills
brainstorming answers "What are we building?"
writing-plans answers "What are the exact steps?"
challenge-plan answers "Are there problems with this plan?"
deep-planning answers "What is the optimal technical approach, and why?"
Checklist
You MUST create a task for each of these items and complete them in order:
- Define success criteria โ What does "good" look like? Performance targets, maintainability bar, constraints.
- Investigate the codebase โ Launch Explore agents if available. Otherwise inspect the codebase directly. Map current patterns, utilities, abstractions. Find reusable code.
- Question the status quo โ For each existing pattern: Why was it done this way? What are its limitations? Is there a fundamentally better approach?
- Research externally โ When the problem domain warrants it, use web search. How do popular libraries, frameworks, or competing tools solve this? What are current best practices? (Skip for purely internal/domain-specific tasks.)
- Formulate approaches โ Design 1-3 approaches. Each must reference specific files and code paths. Include at least one approach that challenges existing patterns (if the status quo has limitations).
- Compare and recommend โ Trade-off matrix. Clear recommendation. Honest about costs.
- Present to user โ Use the output format below. Be concise.
Process Flow
digraph deep_planning {
"Define success criteria" [shape=box];
"Investigate codebase" [shape=box];
"Question status quo" [shape=box];
"Domain warrants research?" [shape=diamond];
"Research externally" [shape=box];
"Formulate approaches" [shape=box];
"Compare and recommend" [shape=box];
"Present to user" [shape=doublecircle];
"Define success criteria" -> "Investigate codebase";
"Investigate codebase" -> "Question status quo";
"Question status quo" -> "Domain warrants research?";
"Domain warrants research?" -> "Research externally" [label="yes"];
"Domain warrants research?" -> "Formulate approaches" [label="no"];
"Research externally" -> "Formulate approaches";
"Formulate approaches" -> "Compare and recommend";
"Compare and recommend" -> "Present to user";
}
Critical Rules
Question the Status Quo (Baseline agents skip this 100% of the time)
For every existing pattern you find, ask:
- Why was it done this way? (commit history, PR context, or infer from constraints)
- What are its limitations? (performance ceiling, maintainability debt, missing capabilities)
- Would a fundamentally different approach serve better? Not just a variation โ a different mental model.
Do NOT accept existing patterns as given. They may be historical artifacts, premature abstractions, or simply the first thing that worked. Your job is to evaluate them critically.
Research Externally (Baseline agents never do this)
When the problem domain is not purely internal (caching strategies, rate limiting, streaming protocols, state management, etc.), you MUST research:
- How popular open-source projects solve this
- Current industry best practices
- Existing libraries that could be leveraged
Use web search. Do NOT rely solely on training knowledge โ it may be outdated or incomplete.
Skip research when: The task is purely internal (renaming, reorganizing existing code, fixing a bug in project-specific logic).
Include a Challenging Alternative (Baseline agents stay conservative)
If the status quo has limitations, at least one of your approaches MUST challenge existing patterns โ not just extend them. Ask: "What if we did this completely differently?"
This doesn't mean the radical approach must win. Often the incremental approach is correct. But you must consider the alternative so the recommendation is informed, not default.
Exception: When there is genuinely only one good approach, say so explicitly and explain why alternatives are worse. Don't manufacture fake alternatives.
Be Concise (Baseline agents write 2000+ words)
Your analysis should be scannable. The comparison table is the centerpiece โ not prose.
- Non-recommended approaches: 2-3 sentences max. Just enough to understand what it is and why it's not recommended.
- Recommended approach: 4-6 sentences. Include specific files, the key design decision, and why it wins.
- Current State: Focus on limitations, not exhaustive description. 3-5 bullet points.
- External Research: Key takeaways only, not summaries of articles. 2-4 bullets with source links.
Target: 400-700 words for the final output. Complex tasks may need 800-1000 โ but if you're over 1000, you're writing an implementation plan, not a planning analysis. Cut prose, expand the table.
Output Format
## Deep Planning: [Task Description]
### Success Criteria
[What "good" looks like โ performance targets, constraints, maintainability bar]
### Current State
[How the codebase handles this today โ specific files, patterns]
[Limitations of the current approach]
### External Research (if applicable)
[How popular solutions handle this โ with sources]
[Key takeaways for our context]
### Approaches
#### Approach A: [Name] (Recommended)
[4-6 sentence description with specific files/code changes]
#### Approach B: [Name]
[2-3 sentence description]
#### Approach C: [Name] (if applicable)
[2-3 sentence description]
### Comparison
| Dimension | A | B | C |
| --------------- | --- | --- | --- |
| Performance | ... | ... | ... |
| Maintainability | ... | ... | ... |
| Complexity | ... | ... | ... |
| Migration cost | ... | ... | ... |
### Recommendation
[Which approach and why โ 2-3 sentences]
[Key risks to watch for]