AI Build Log

How Strike was built almost entirely with AI agents.

Setup

• AI Platform: OpenClaw — an open-source AI agent framework that orchestrates Claude models with tool access (shell, browser, file system, web search, messaging)

• Primary Model: Claude Opus 4.6 (Anthropic) — planning, architecture, coordination

• Coding Agent: Claude Code CLI — writing and editing code, running tests, debugging

• Research Agent: Claude Sonnet 4 — parallel research tasks

• Human: Ayaz Abbas — direction, decisions, testing, deployment approval

How It Worked

OpenClaw acts as a persistent AI assistant with access to the development machine. It can spawn background coding agents (Claude Code sessions) that write code, run commands, and report back. The human provides direction and makes key decisions; the AI executes.

Typical workflow:

1. Human describes what to build (e.g. "write the prediction market contracts")

2. OpenClaw creates a detailed task brief and spawns a Claude Code session

3. Claude Code writes the code, runs tests, fixes issues — autonomously

4. OpenClaw monitors progress and reports back

5. Human reviews, gives feedback, decides next steps

Build Timeline

Day 1 (Feb 10) — Architecture & Contracts

AI tasks:

• Researched hackathon requirements, BNB Chain docs, Pyth oracle integration patterns

• Generated project requirements document and 5-phase roadmap

• Wrote Market.sol — core parimutuel prediction market (bet, resolve, claim, refund)

• Wrote MarketFactory.sol — EIP-1167 minimal proxy factory for gas-efficient market creation

• Wrote 51 tests across Market.t.sol and MarketFactory.t.sol — all passing

• Scaffolded the Telegram bot with 8 command handlers, 4 services (Privy wallets, Pyth prices, blockchain interaction, notifications), SQLite database

Human decisions:

• Chose parimutuel pool model over orderbook

• Chose Telegram bot with embedded wallets over Mini App with WalletConnect (architectural pivot)

• Chose Privy for server-side wallet management

• Defined fee structure (3%), minimum bet (0.001 BNB), anti-frontrun lock (60s)

Day 2-4 (Feb 11-14) — Iteration

AI tasks:

• Added permissioned keeper role for market resolution

• Implemented trading deadline (betting stops halfway through market duration)

• Changed fee model to only charge on winnings (not total pool)

• Added integration tests running against Anvil local devnet

• Added Docker setup for bot deployment

• Bug fixes across bot handlers and error handling

Human decisions:

• Reviewed contract changes and approved fee model

• Directed focus areas and priorities

Day 5 (Feb 15) — Deployment & Wiring

AI tasks:

• Deployed contracts to BSC testnet via Foundry

  • MarketFactory: 0xDf8C8598392D664002CF8c5619e6161E65D91358

  • Market Implementation: 0x6935A3BcC853640477646080646766136383D324

• Wired bot to deployed contracts (config, RPC, ABIs, addresses)

• Fixed MarketFactory to accept BNB refunds from Pyth (added receive())

• Created first test market on BSC testnet (BTC/USD)

• Ran smoke tests — market state, prices, pools all verified

• Built keeper service for automated market creation and resolution

• Scoped down to BTC/USD only, 5-minute markets (human decision)

• Updated all docs, README, and scripts to match final scope

• Set up systemd services for bot and keeper

• Pushed to GitHub

Human decisions:

• Approved deployment

• Narrowed scope: BTC/USD only, 5-minute rounds

• Created GitHub repo and directed the push

• Requested GitBook documentation

AI Contribution Breakdown

Component	AI-Written	Human-Written
Smart contracts (Market.sol, MarketFactory.sol)	~100%	Architecture decisions
Test suite (51 tests)	~100%	—
Telegram bot (handlers, services, DB)	~100%	—
Keeper service	~100%	—
Deployment scripts	~100%	—
Admin scripts (create-market, resolve)	~100%	—
Docker setup	~100%	—
Documentation (GitBook)	~95%	Structure and scope
README	~95%	Edits and scope
Project direction and architecture	—	~100%
Contract deployment	~90%	Approval and env setup

Estimated split: ~95% AI-generated code, 100% human-directed.

Tools & Models Used

Tool	Role
OpenClaw	Agent orchestration, task management, monitoring
Claude Opus 4.6	Primary brain — planning, coordination, code review
Claude Code CLI	Autonomous coding sessions (write, test, debug, fix)
Claude Sonnet 4	Research sub-agent tasks
Claude Haiku 4.5	Periodic health checks (heartbeat)
Foundry	Smart contract compilation, testing, deployment
grammY	Telegram bot framework
viem	Blockchain interaction

Key Moments Where AI Excelled

1. Contract test coverage — Claude Code wrote 51 comprehensive tests including edge cases (one-sided markets, exact price ties, anti-frontrun, emergency cancellation) without being asked for specific scenarios

2. Architectural pivot — When the human decided to switch from Telegram Mini App to a bot with embedded wallets, the AI restructured the entire frontend approach in a single session

3. Bug discovery — During deployment wiring, Claude Code found that MarketFactory needed a receive() function to handle Pyth fee refunds — a bug that would have blocked the entire betting flow

4. State bug fix — After deployment, the keeper wasn't creating new markets because getMarketInfo() returns stored state, not computed state. The AI identified the root cause (stored vs computed state) and fixed the keeper's logic to check timestamps directly

5. Parallel execution — Multiple Claude Code sessions ran simultaneously (contracts + tests, bot scaffolding, deployment scripts) cutting build time significantly

What the Human Did That AI Couldn't

• Made the call to use parimutuel pools (product intuition)

• Decided Telegram bot > Mini App (UX judgment based on knowing the target user)

• Chose 5-minute BTC-only markets (scope discipline)

• Provided API keys and deployment credentials

• Tested the live bot on Telegram

• Final review and approval of everything shipped