| name | mine.research |
| description | Use when the user says: "research adding X", "feasibility study", "evaluate approach", or wants a focused investigation before committing. Dispatches the researcher agent for codebase investigation and presents a structured brief. |
| user-invocable | true |
Research
Deep investigation of a codebase to evaluate a proposed change, new pattern, or architectural direction. Heavy on questions — surfaces what the user truly needs, not just what they initially say. Produces a structured research brief.
How This Differs From Other Skills
| Skill | Question it answers |
|---|
/mine.research | "What would it take to do X in this codebase?" |
/mine.design | "How should we build X?" |
Research comes before design docs and plans. It's the investigation that makes those possible.
Arguments
$ARGUMENTS — the proposal to investigate. Can be:
- A change proposal:
/research "add SQLite database and command pattern"
- A technology question:
/research "should we use SQLAlchemy or raw sqlite3?"
- A pattern evaluation:
/research "would event sourcing work here?"
- A migration question:
/research "what would it take to move from files to a database?"
- A broad direction:
/research "this app needs persistent state — what are our options?"
- Empty: ask the user what they're considering
Phase 1: Understand the Ask
Do not start exploring the codebase yet. First, understand what the user actually wants and why.
Initial questions
Use AskUserQuestion to probe motivation, constraints, and prior thinking. Ask 2-3 questions in a single call (multi-question, not multi-select).
The goal is to distinguish between:
- "I've decided to do X, tell me how" vs. "I'm considering X, help me decide"
- "I want specifically SQLite" vs. "I need persistence and SQLite is one option"
- "This is urgent and I want to ship it this week" vs. "This is exploratory"
Example opening questions — adapt the Motivation options to match the proposal's domain. The snippets below are items within the AskUserQuestion: questions: array. They show how the structure varies across different types of proposals:
Persistence/data layer proposal (e.g., "add a database"):
- question: "What's driving this change?"
header: "Motivation"
options:
- label: "Data is getting lost"
description: "State doesn't survive restarts, crashes, or context switches"
- label: "Growing complexity"
description: "In-memory data structures are getting unwieldy"
- label: "New feature needs it"
description: "A planned feature requires persistent or queryable data"
Architecture/pattern proposal (e.g., "should we adopt a monorepo?"):
- question: "What's driving this change?"
header: "Motivation"
options:
- label: "Coordination overhead"
description: "Cross-repo changes are painful — too many PRs, broken integrations"
- label: "Inconsistent standards"
description: "Each repo drifts on tooling, linting, testing conventions"
- label: "Deployment coupling"
description: "Services need to ship together but repos make that hard"
Technology choice proposal (e.g., "switch from REST to GraphQL"):
- question: "What's driving this change?"
header: "Motivation"
options:
- label: "Over-fetching / under-fetching"
description: "Clients need data from multiple endpoints or get too much back"
- label: "API evolution pain"
description: "Adding fields or endpoints is getting fragile"
- label: "Developer experience"
description: "The current API is hard to work with or document"
The Flexibility question is consistent across all proposals:
- question: "How committed are you to the specific approach mentioned, vs. open to alternatives?"
header: "Flexibility"
options:
- label: "Exploring options"
description: "I mentioned one idea but I'm open to whatever works best"
- label: "Leaning this way"
description: "I have a preference but could be convinced otherwise"
- label: "Decided"
description: "I've already thought this through — I want to know how, not whether"
Generate domain-appropriate Motivation options based on the user's proposal — the examples above are illustrations, not templates. If the user mentioned multiple things (e.g., "SQLite + command pattern"), ask whether those are linked or separable. If the proposal is vague, ask more. If it's specific, ask fewer.
Follow-up questions
Based on the answers, ask 1-2 targeted follow-ups to fill in gaps:
- Scope: "Should this cover all data in the app, or just [specific area]?"
- Constraints: "Any hard requirements — must be zero-dependency, must work offline, must be reversible?"
- Timeline: "Is this something you want to prototype now, or plan carefully for later?"
- Experience: "Have you used [proposed technology] before, or would this be new territory?"
Do not ask more than 2 rounds of questions before moving to Phase 2. If you still have uncertainties, note them as open questions to revisit after seeing the code.
Phase 2: Investigate
Dispatch the research to a researcher agent. This runs the heavy codebase exploration, web research, and synthesis outside the main context window.
- Run
get-skill-tmpdir mine-research to get a temp directory.
- Determine the research depth from Phase 1 answers:
- Quick — user said "Decided" + narrow scope (single module or specific technology question). Use 2 Explore subagents, focus on feasibility of the chosen approach.
- Normal (default) — user said "Leaning" or moderate scope. Use 3-4 Explore subagents.
- Deep — user said "Exploring" + broad scope (architecture, migration, multiple systems). Use 4 Explore subagents + web research.
- If scope is unclear from Phase 1 answers, default to Normal regardless of Flexibility.
- Launch
Agent(subagent_type: "researcher") with a prompt containing:
- The proposal (from $ARGUMENTS or user input)
- The user's answers from Phase 1 (motivation, flexibility, constraints) — use the caller prompt checklist format from
researcher.md
- Depth:
<quick|normal|deep>
- Output file path:
<tmpdir>/brief.md
- After the agent completes, verify the output: read
<tmpdir>/brief.md and check that it exists and contains the # Research Brief: header. If the file is missing, empty, or lacks the expected header, inform the user that the research agent failed to produce a complete brief and offer:
- Retry the researcher agent
- Proceed with a manual investigation in the main context window
Phase 3: Present & Discuss
Ask where to save
Research briefs are forward-looking — they evaluate a proposed change before committing to an approach. They may feed into a design doc (when a direction is chosen) or be abandoned (if the approach is rejected).
The recommended convention is a date-stamped topic directory under design/research/:
design/research/
└── YYYY-MM-DD-topic-name/
├── research.md Main research brief
├── prereq-01-name.md Prerequisite breakdowns (if applicable)
└── ... Additional artifacts
AskUserQuestion:
question: "Where should I save the research brief?"
header: "Output"
multiSelect: false
options:
- label: "design/research/ (Recommended)"
description: "Save as design/research/YYYY-MM-DD-<topic>/research.md — with room for prereq breakdowns"
- label: "docs/research/"
description: "Save as docs/research/YYYY-MM-DD-<topic>.md"
- label: "Just show me"
description: "Display in the conversation, don't save a file"
Create the design/research/ directory if it doesn't exist. If the project already has research in docs/, follow the existing convention.
Copy (never move) the brief from the temp file to the user's chosen location, or display it inline if they chose "Just show me". The tmpdir copy must always remain intact — downstream challenge and design handoffs reference it.
After the save step, set <research_brief_path>:
- If saved to a permanent location:
<research_brief_path> = the saved file path
- If "Just show me":
<research_brief_path> = <tmpdir>/brief.md
Present the key findings conversationally and ask what the user wants to do next.
AskUserQuestion:
question: "Research is done. What would you like to do next?"
header: "Next step"
multiSelect: false
options:
- label: "Challenge these findings first"
description: "Run /mine.challenge on the research brief before committing to a direction"
- label: "Design it (/mine.design)"
description: "Formalize findings into a design doc — the research brief will be passed as prior work"
- label: "Build it (/mine.build)"
description: "Skip design — route straight to implementation"
- label: "I need to think about it"
description: "The brief has what I need — I'll come back when I'm ready"
If "Challenge these findings first" is selected: invoke /mine.challenge --mode=passthrough --target-type=research <research_brief_path>. After challenge completes, loop back to this gate.
If "Design it (/mine.design)" is selected: invoke /mine.design and pass <research_brief_path> so mine.design can use it as prior work and skip its own researcher dispatch.
If "Build it (/mine.build)" is selected: invoke /mine.build with context: "Prior research brief available at <research_brief_path>." This ensures mine.build's prior-analysis detection fires reliably.
Principles
- Questions before code — understand motivation and constraints before exploring the codebase. The user's first description of what they want is almost never the full picture.
- Options, not prescriptions — present trade-offs honestly. Include a "do less" option when the proposal is ambitious. The user decides, you inform.
- Feeds forward — the research brief should contain everything needed to write a design doc or create an implementation plan. No redundant investigation later.
Common Rationalizations
| Rationalization | Reality |
|---|
| "I have extensive prior analysis — skip research" | Prior analysis covers the problem space. Research covers the solution space — competing approaches, existing patterns in the codebase, API constraints. They answer different questions. Skipping research has led to shipping with suboptimal designs that a brief investigation would have caught. |
| "The approach is obvious, no need to investigate" | The "obviously correct" approach is by definition the least-investigated one. Research has repeatedly surfaced better designs — method-based vs persona-based splits, orthogonal dimensions the intuitive approach didn't consider. Research is cheap relative to a design revision after challenge finds gaps. |
| "Research will slow us down" | A 2-minute research dispatch is cheaper than a design revision after challenge finds holes that research would have prevented. The brief feeds forward into design — skipping it means the design phase compensates with ad-hoc investigation anyway. |
| "This question implies we should design/file/ship" | A scoped question ("can X be done?") is not a workflow trigger. Answer the question by completing the brief and presenting the Phase 3 next-step gate — that gate is the escalation offer. Don't silently launch issue filing, design, or shipping from a scoped question. |
What This Skill Does NOT Do
- Make decisions — it informs them. Use
/mine.design to formalize decisions.
- Plan implementations — it assesses feasibility. Use
/mine.build to route to the right implementation workflow.
- Write code — it's pure investigation. No prototypes, no scaffolding, no "let me just try it."
- Audit health — it evaluates a specific proposal against the codebase. Use
/mine.challenge for general health assessment.
- Benchmark or profile — it can identify likely performance concerns from code reading, but won't run benchmarks.
