읽을거리

Anthropic 공식 Claude Code 플러그인의 code-review, security-review, feature-dev는 OpenCode에서 그대로 동작하지 않는다. opencode-power-pack은 이 워크플로우들을 OpenCode 네이티브 SKILL.md 포맷으로 번역하고, ~/.config/opencode/opencode.json 한 줄 설정으로 11개 스킬을 적재한다.

guard-skills는 Claude Code, Codex, Cursor가 생성한 코드·테스트·문서에 second-pass 리뷰를 수행하는 5개 Skill 패키지입니다. clean-code-guard는 GitClear·USENIX 연구에 근거한 14가지 AI 실패 패턴을 검사하고, woo-guard는 AI가 반복 생성하는 pre-HPOS WooCommerce 코드 패턴을 직접 겨냥합니다.

baoyu-design ports the claude.ai/design methodology to Claude Code by detecting the agent environment at runtime and loading a tool substitution table from `references/claude.md`. The design system pipeline — `compile-design-system.mjs` produces static lint config, `import-design-system.mjs` generates a token allowlist as `_ds_prompt.md` — enforces the same constraint at two layers. Neither is optional if you want session-to-session consistency.

A small k6/x/mcp workflow shows how to prove an MCP server starts, lists tools, returns expected JSON, and fails CI before an agent depends on it.

Agent Flow is worth saving as an observability article: it turns Claude Code and Codex sessions into traces a reviewer can inspect before trusting the final patch.

The right lesson is not a token-saving trick. It is an MCP interface pattern: upload large files out of band, validate them server-side, and pass only an artifact ID through the model.

The useful story is not a markdown hook that saves tokens. It is a 228-run retraction that teaches Claude Code teams when a hook should be killed, narrowed, or replaced by one instruction.

The repo is worth saving because it gives teams an onboarding system from the first task loop through Skills, subagents, hooks, and MCP.

claude-code-from-scratch is worth saving as an agent runtime map: skills, compaction, subagents, worktrees, permissions, and MCP each get a concrete source-backed checkpoint.

The rescue angle is simple: real 5h and 7d usage can be a better gate than token estimates, but only if the MCP, prompt protocol, credential path, and stale repo docs are checked first.

The useful article is not that Kagi search can be made free for MCP clients. It is how to test a session-token MCP without mistaking it for a normal API key.

A practical audit turns Claude Code commands, memory, MCP, Skills, settings, and hooks into a runbook a team can verify.

Claude Meter is not just a taskbar percentage. It is a small local loop: Claude Code credentials, `/api/oauth/usage`, adaptive polling, Windows UI, and conservative failure handling.

Clawdmeter is worth rescuing as a source-backed tool review because it exposes a concrete boundary: OAuth token access, rate-limit polling, BLE transport, ESP32 firmware, and hardware maintenance.

The useful question is not whether an assistant can remember more. It is whether a developer should let an MCP server capture every turn, store it locally, and inject recall into future sessions.

The useful article is not about generic productivity. It is about when cross-repo delegation gives an agent the right local context, and when it only adds cost, risk, and confusion.

The useful article is not about real-time Ethereum testing. It is about using a narrow Claude Code skill to stop agents from writing smart contract PoCs that pass locally but fail to prove the finding.

office-oxide-mcp is worth rescuing as a Claude Code and Codex workflow piece when the focus shifts from Rust speed claims to local document operations with file-scope controls.

ShipWithAI's Claude Code course is worth rescuing when it becomes a set of local team runbooks, not another list of course modules.

A serious claude-usage article should drop the unproven spike hook and show how teams can audit Claude Code sessions, token totals, cache behavior, and cap pressure from local logs.

Red Hat's evaluator is useful when a Claude Code setup has grown enough that weak instructions, duplicate skills, unsafe hooks, and vague triggers need evidence before the team deletes or rewrites them.

oop-architect deserves rescue as a practical guide to keeping design intent, implementation status, and Claude Code's working context aligned through CLAUDE.md.

Claude Soul deserves rescue as a field guide for agent memory: install it as a runtime change, verify hooks, separate facts from behavior, and set conflict rules before memory starts steering Claude Code.

CodeBurn is worth rescuing as an operating guide for AI coding teams: separate exact usage from estimates, inspect Codex and Claude parsers, review waste findings, and turn costs into local policy rather than billing theater.

gstack is worth saving as a source-backed field guide for adopting a full Claude Code workflow stack without confusing skill count, productivity claims, or team hooks with proof of fit.

Claude Code on Steroids is worth saving as an adoption audit for serious skill packs: inspect the installer, pinning model, startup hook, token dashboard, test scripts, and rollout path before use.

AI Research Skills is worth saving as a source-backed field guide for building serious Claude Code and Codex skill catalogs, not as a generic research automation story.

The useful story is not a benchmark headline. It is a local MCP server that helps Codex choose source files before spending context on full reads.

Unity Claude Template is worth saving as a field guide for teams that want Claude Code to operate inside Unity without turning the editor into an unbounded file-editing target.

claude-replay is not a code-review engine. It is a practical way to turn messy agent transcripts into a reviewable, redacted artifact.

Claude Coworker deserves publication as a bounded delegation pattern for Claude Code, not as a guaranteed cost-saving trick.

Stacklit is worth saving as a tool review when the article explains the generated index, MCP surface, freshness contract, and adoption boundary.

AIS-OS is worth saving as a field guide for Claude Code operators who need context, connections, capability planning, and a cadence before they add more agents.

Vercel Sandbox is useful for Claude Managed Agents when the team reviews the webhook control plane, snapshot runner, network policy, and credential brokering together.

Agentify can turn OpenAPI into runnable MCP code and agent-facing files. The real adoption rule is narrower: parse first, generate a small surface, inspect handlers, prune risky endpoints, and test one safe tool.

claude-teams-brain combines Claude Code hooks, SQLite/FTS5, MCP command-output filters, and reviewable memory commands. The useful question is not whether agents remember more, but whether the team can trust what gets injected.

parallel-agent-worktree-skill is worth saving as an operating guide for Codex and Claude Code teams that want multiple agents without letting them trample the same checkout.

AgentSeal can give Claude Code and Codex users a source-backed way to inspect MCP configs, poisoned tool descriptions, skill files, prompt probes, registry signals, and local agent paths before adopting a new tool.

ctxharness turns selected claims in agent instruction files into repo checks: paths, package scripts, versions, rule globs, hooks, token budgets, and stale context signals.

Vibe Coding Security is worth saving because it gives Claude Code and Codex users concrete pre-launch checks: RLS SQL, NEXT_PUBLIC_ bundle scans, prompt-injection searches, rate-limit review, and React XSS inspection.