Builds raw unsigned DeFi transactions using the defi-skills CLI. The tool is a deterministic, stateless PlaybookEngine. It takes one action at a time with structured input, resolves token addresses, converts amounts to base units, ABI-encodes calldata, and outputs unsigned transactions. It has no memory between calls.
You are the LLM. Your job is to understand the user's intent, determine the correct action and arguments, then call the CLI with --action and . No LLM runs inside the CLI. It is purely deterministic. You handle all planning, decomposition, and sequencing.
--args
The output is an unsigned transaction (to, value, data). It never signs or broadcasts.
Agent Behavior
These are financial transactions. Be collaborative, not autonomous.
Direct intent: If the user's request maps to a single action with clear parameters, build it immediately. No discussion needed.
Examples: "Send 0.5 ETH to vitalik.eth", "Supply 100 USDC to Aave", "Wrap 5 ETH".
Multi-step intent: If the request requires multiple actions, present your execution plan to the user before building anything. Explain:
What steps you will execute and in what order
What tokens and amounts are involved at each step
Whether you can build all steps upfront (predictable outputs like stake/wrap) or need to execute step-by-step (unpredictable outputs like swaps)
Any assumptions you are making (e.g., "I will use the same amount for step 2 since Lido staking is roughly 1:1")
Only proceed after the user confirms the plan. If the user modifies the plan, adjust accordingly.
Ambiguous intent: If the request is vague or could be interpreted multiple ways, ask clarifying questions before doing anything. Do not assume a protocol, token, or strategy.
Examples:
"I want to earn yield on my ETH": Which protocol? Lido staking? Aave supply? Rocket Pool? Each has different risk/reward.
"Move my stables somewhere safe": Which stablecoins? What does "safe" mean to them? Aave supply? MakerDAO DSR?
"Do something with my USDC": Too vague. Ask what they want to achieve.
Large amounts or "max": If the user is operating with "max" (entire wallet balance) or amounts you consider significant, confirm the details before building. Remind them that "max" means their entire balance of that token, not just a portion.
Constraints
Multi-chain support. Supported chains: Ethereum Mainnet (1), Arbitrum (42161), Base (8453), Optimism (10), Polygon (137), Sepolia (11155111). Default is mainnet. Use --chain-id to target other chains. Not all actions are available on all chains — check with defi-skills actions --chain-id <id>.
One action per CLI call. For multi-step intents, decompose into separate build calls. You are the planner.
Some actions restrict valid tokens. If you are unsure whether a token is supported for an action, run defi-skills actions <name> --json to check. The CLI will reject unsupported tokens with a clear error listing the valid options.
Negative amounts are rejected. All amounts must be zero or positive.
Always check "success" in the JSON response. If success is false, the error field explains what went wrong. Do not proceed with failed builds.
Some actions have optional parameters with sensible defaults (slippage, fee tier, deadline, recipient address). These are visible via defi-skills actions <name> --json. Inform the user about relevant defaults before building so they can override if needed.
Prerequisites
The CLI must be installed:
pip install --upgrade defi-skills
A wallet address is required. The CLI reads it from (in priority order):
--wallet flag on each command
WALLET_ADDRESS environment variable
Persisted config (defi-skills config set-wallet)
For agents, prefer the env var or --wallet flag — no file writes needed. If no wallet is available, ask the user to provide their address before continuing.
API keys
Most actions that involve on-chain data need ALCHEMY_API_KEY (ENS resolution, swap quotes, "max" balance queries, EigenLayer strategy verification, etc.). Simple actions with known tokens and fixed amounts (like aave_supply with USDC) work without it.
If a build fails with "no web3 instance" or "no RPC provider available", the user needs to set their Alchemy key via environment variable or CLI config:
# Environment variable (preferred for agents)export ALCHEMY_API_KEY="<KEY>"# Or persist via CLI config
defi-skills config set alchemy_api_key "<KEY>"
Balancer actions additionally need THEGRAPH_API_KEY:
export THEGRAPH_API_KEY="<KEY>"# Or persist via CLI config
defi-skills config set thegraph_api_key "<KEY>"
Workflow
Step 1: Identify the Action
Determine the correct action from the supported actions table below. If unsure, or to check availability on a specific chain:
echo"$TX" | jq '.success'# must be trueecho"$TX" | jq '.transactions'# ordered array: approvals first, then action
Response Format
The response is a JSON object with success and either transactions (on success) or error (on failure).
On success, transactions is an ordered array. Execute them in order: approval transactions first, then the main action. Some actions need no approval (1 transaction), others need 1 approval (2 transactions), and USDT needs a reset-to-zero step (3 transactions).
The CLI is stateless. Each build call is independent with no memory of previous calls and no way to reference the output of a prior transaction. When planning multi-step intents:
Predictable outputs (stake, wrap, supply): input roughly equals output. Build all steps upfront with the same amount. Example: stake 10 ETH on Lido produces ~10 stETH, so pass "amount":"10" to the EigenLayer deposit.
Unpredictable outputs (swaps, complex withdrawals): the output depends on live prices or on-chain state. You cannot know the exact amount at build time. Build and execute step 1 first, then ask the user for the result or check on-chain before building step 2. Do not guess amounts.
"max" means entire wallet balance, not "whatever came from the last step." If the user already holds the token, "max" will include their existing balance plus the new amount. Make sure the user understands this before using it.
Error Handling
success: false: Read the error field. Do not retry blindly. Fix the input based on the error message.
Unknown action: Run defi-skills actions to see all supported actions.
Unsupported token: The error lists valid tokens for the action.
Action not available on chain: The action exists but not on the requested chain. Use defi-skills actions --chain-id <id> to see what's available. Suggest mainnet if the action is mainnet-only (e.g., Lido, EigenLayer, Curve).
Negative amount: Amounts must be zero or positive.
ENS resolution failed: The user needs to run defi-skills config set alchemy_api_key <KEY>, or provide a hex address instead. Note: ENS is only available on Ethereum Mainnet and Sepolia. On L2 chains, always use hex addresses (0x...).
Missing wallet: Run defi-skills config set-wallet <address>.
CLI not found: Run pip install defi-skills.
Safety
Output is an unsigned transaction. The tool never signs or broadcasts.
No private keys are involved at any stage.
The deterministic build path (--action + --args) uses zero LLM tokens.
Always review the raw_tx before signing. Verify to, value, and data.
