Self-contained CLI for OpenAI gpt-image-2 generation/editing with single-shot and --batch modes. Thread-safe concurrent job execution via ThreadPoolExecutor, fail-fast error policy (exits non-zero on first error, never silently retries or swaps models), explicit exit codes (2 for missing env/package), clean argparse interface, and correct file lifecycle management for uploaded references. Reusable as a template for any single-model API wrapper needing deterministic error behavior and parallel batch jobs.

Exemplar of SKILL.md instruction engineering: hard rules with 'do not violate' wording, a fail-fast policy mirrored in its companion script, a JSON batch-job format for parallel generation, a context-gathering protocol (inspect sibling assets, brand tokens, and destination surface before prompting), and explicit responsibility for wiring generated images into the codebase. Demonstrates how to encode non-negotiable constraints and quality gates in agent instructions.

Structured methodology for auditing browser network traffic via the cursor-ide-browser MCP. Classifies findings into failures (4xx/5xx, CORS), performance (slow, duplicate, oversized requests), security/privacy (sensitive data in URLs, missing auth, unexpected hosts, preflight storms), and correctness. Includes a defined reporting format and explicit pairing with console-message checks. The audit taxonomy is transferable to any browser-network QA setup.

Novel problem-solving pattern: define 2-3 distinct strategies for a hard problem, launch each in an isolated git worktree via best-of-n-runner subagents so they run concurrently without interfering, then compare results against explicit success criteria and merge the winner. Encodes key operational details: specific prompts per approach, isolation guarantees, and cleanup of losing branches.

Runs four read-only explore subagents on the same diff — security, performance, correctness, readability — with ready-to-paste per-lens prompts, then synthesizes de-duplicated findings ordered by severity and fix cost. The read-only constraint on parallel reviewers is the key safety insight, and the workflow scales to per-directory review waves for large diffs.

Encodes a reproduce → isolate → hypothesize → verify debugging loop with concrete tools per scenario (git bisect for regressions, binary search over code layers, layer isolation via curl/network checks), a table of seven common bug patterns, and explicit 'never guess' rules. Tool-agnostic and directly extractable as a debugging methodology for any agent or human engineer.

Meta-skill that promotes observed repeatable workflows into new SKILL.md files. Includes the repo's own content pattern (slug directories, frontmatter, lean body sections), a validation checklist, and a decision table for choosing between skills (on-demand procedures), rules (always-on conventions), and hooks (event automation). This self-replicating pattern is transferable to any agent-skills ecosystem.