DeepWiki Research

Query repositories via DeepWiki MCP to answer technical questions with precise, source-backed answers.

Workflow

Understand the question
- Identify what the user is asking and which domain it falls into
- If the question is ambiguous, state your understanding and ask for clarification
Gather context
- Match question to the execution steps below
- Use Glob, Grep, and Read to find relevant local files
- Query DeepWiki MCP (mcp__deepwiki__ask_question) and mcp__zkcompression__SearchLightProtocol for repository-level context
- Use Task subagents for parallel research across multiple repos when needed
Synthesize and respond
- Apply precision rules to the answer
- Format per format response

Execution Steps

1. Read required context in local repo

Use Glob and Grep to locate relevant files in the current repository. Use Read to pull in specific content needed to answer the question.

2. Identify question scope

Determine the domain:

Programs, client SDKs, architecture, implementation details
Specific components (LightAccount, ValidityProof, CPI, etc.)

3. Fetch repository context

Select the appropriate repository based on question scope:

Light Protocol (compressed accounts, state trees, ZK compression, Light SDK)

mcp__deepwiki__read_wiki_structure("Lightprotocol/light-protocol")
mcp__deepwiki__read_wiki_contents("Lightprotocol/light-protocol")
mcp__deepwiki__ask_question("Lightprotocol/light-protocol", "your question")

Solana Development (programs, accounts, general Solana)

mcp__deepwiki__read_wiki_structure("blueshift-gg/blueshift-dashboard")
mcp__deepwiki__read_wiki_contents("blueshift-gg/blueshift-dashboard")
mcp__deepwiki__ask_question("blueshift-gg/blueshift-dashboard", "your question")

Claude Code (CLI features, commands, configuration)

mcp__deepwiki__read_wiki_structure("anthropics/claude-code")
mcp__deepwiki__read_wiki_contents("anthropics/claude-code")
mcp__deepwiki__ask_question("anthropics/claude-code", "your question")

Agent Skills (skills, subagents, patterns)

mcp__deepwiki__read_wiki_structure("anthropics/skills")
mcp__deepwiki__read_wiki_contents("anthropics/skills")
mcp__deepwiki__ask_question("anthropics/skills", "your question")

4. Apply precision rules

AVOID:

Vague verbs: "handles", "manages", "processes", "enables", "provides"
Abstract nouns: "operations", "management", "coordination"
-ing endings when a precise verb exists
Marketing language: "comprehensive", "rich", "seamless", "enables", "provides capability"
ZK terminology in user-facing docs: "inclusion proof", "non-inclusion proof" → Instead describe what it does: "prove the address doesn't exist", "prove the account hash exists"
Generic descriptions: "account metadata" instead of "CompressedAccountMeta"

USE:

Exact function/method names
Concrete data structures
Specific operations: "sets discriminator", "creates hash", "nullifies account"
File:line references
"reference" when describing u8 indices that reference accounts in the accounts array (standard Solana terminology)
"CPI" as a verb (e.g., "CPIs the Account Compression Program")
Technical method/function names used as verbs when contextually clear

Describe exactly what happens.

5. Format response

Structure answers with:

Technical precision
Code examples with inline comments
Source references (file paths, line numbers)

Example Output

Question: What do you include in instruction data for compressed account creation?

Answer:

Compressed Account Creation Instruction Data

Instruction data contains three components for proof verification and tree coordination.

Components:

pub struct InstructionData {
    proof: ValidityProof,
    address_tree_info: PackedAddressTreeInfo,
    output_state_tree_index: u8,
    // + custom account data
}

1. Non-inclusion proof

ValidityProof proves the derived address does not exist in the specified address tree
Client generates via getValidityProof() RPC call

2. Tree location parameters

PackedAddressTreeInfo: Specifies which address tree registers the address (u8 index, not pubkey)
output_state_tree_index: Specifies which state tree stores the compressed account hash

3. Custom account data

Program-specific fields (e.g., message: String, user data)

Execution flow:

Address tree proves uniqueness, stores address
State tree stores account hash
ValidityProof cryptographically guarantees address is unused

Why separate trees:

Compressed accounts require client-generated cryptographic proof that address doesn't exist (unlike regular Solana where runtime checks PDA existence). Address trees enforce uniqueness; state trees store account hashes.

Packed structs use u8 indices to reference accounts in remaining_accounts, reducing transaction size.

Security

This skill does not pull, store, or transmit external secrets. It provides code patterns, documentation references, and development guidance only.

No credentials consumed. The skill requires no API keys, private keys, or signing secrets. env: [] is declared explicitly.
DeepWiki MCP accesses public repositories only. All mcp__deepwiki__* calls query public GitHub repositories (Lightprotocol/light-protocol, anthropics/claude-code, anthropics/skills). No authentication tokens are required or transmitted. DeepWiki does not access private repositories unless explicitly configured with a token — this skill does not configure one.
User-provided configuration. RPC endpoints, wallet keypairs, and authentication tokens (Privy, wallet adapters) are configured in the user's own application code — the skill only demonstrates how to use them.
Tool boundary enforced. The allowed-tools list restricts this skill to read-only operations (Read, Glob, Grep), research subagents (Task), web fetches to Light Protocol domains, and MCP queries. It cannot load other skills, write files, or execute shell commands. Verify the allowed-tools list in the frontmatter above matches these constraints.
Install source. npx skills add Lightprotocol/skills installs from the public GitHub repository (Lightprotocol/skills). Verify the source before running.
Audited protocol. Light Protocol smart contracts are independently audited. Reports are published at github.com/Lightprotocol/light-protocol/tree/main/audits.

ask-mcp