Further Reading

A curated set of high-quality external sources to go deeper, grouped by theme. Every link below was checked and resolves as of July 2026, and points to a primary, authoritative source — official vendor documentation, the original research, or the tool's own maintainers. Each entry has a one-line note on why it's worth your time.

Two caveats. First, because the field moves fast (see limitation #3: outdated knowledge), treat any specific model name, price, or benchmark number as perishable — re-check before relying on it. Second, Anthropic's documentation has been consolidating onto platform.claude.com (API/platform), code.claude.com and docs.claude.com (Claude Code and guides), and anthropic.com/engineering (deep dives), so some older docs.anthropic.com URLs now redirect.

---

Prompt engineering & best practices

Source	Why it's worth reading
Anthropic — Prompting best practices	The living reference for prompting Claude: clarity, examples, XML structuring, role prompting, thinking, and agentic prompting, with model-specific notes.
Anthropic — Prompt engineering interactive tutorial (GitHub)	A hands-on, 9-chapter, exercise-driven course; the fastest way to build real prompting intuition by doing rather than reading.
OpenAI — Prompt engineering guide	The six core strategies (clear instructions, reference text, splitting tasks, time to think, external tools, systematic testing); a useful vendor-independent cross-check.
Google — Gemini API: prompt design strategies	The same fundamentals from a third major lab, with Gemini-specific notes on where to place critical instructions.
Anthropic — Extended thinking	How adaptive (always-on) thinking works and how the effort parameter trades reasoning depth against latency and cost on current Claude models.

Context & limitations (incl. the "lost in the middle" research)

Source	Why it's worth reading
Liu et al. — Lost in the Middle: How Language Models Use Long Contexts (arXiv)	The primary research behind "lost in the middle": models use information best at the start and end of a long context, worst in the middle.
ACL Anthology — same paper (TACL 2024)	The peer-reviewed, citable version of the paper above.
Chroma Research — Context Rot: How Increasing Input Tokens Impacts LLM Performance	Controlled tests across many frontier models showing reliability degrades well before the context window is full — even on simple tasks.
OpenAI — Why language models hallucinate	The lab's own account of why models confabulate — standard training and evaluation reward confident guessing over admitting uncertainty. The primary-source case for why verification is not optional.
Anthropic — Effective context engineering for AI agents	The deepest single piece on managing the context window — the discipline at the heart of working around limitation #1.
Anthropic — Context windows (Claude Platform Docs)	The mechanics of how the context window works, including with extended thinking and prompt caching.

Models & leaderboards

Source	Why it's worth reading
Artificial Analysis — Models leaderboard	Independent, continuously updated benchmarks across quality, speed, price, and context window — the practical first stop for model choice.
LMArena (Chatbot Arena) — text leaderboard	Crowdsourced head-to-head human-preference (Elo) rankings; measures real-world chat quality rather than narrow benchmarks. (The former lmarena.ai redirects here.)
SWE-bench — leaderboards	The standard execution-based benchmark for coding agents: real GitHub issues resolved by passing the repo's tests. The most-cited number for agentic coding.
Anthropic — Models overview (Claude Platform Docs)	The current Claude lineup with context windows, pricing, and knowledge cutoffs side by side.

Claude Code & agentic tools

Source	Why it's worth reading
Anthropic — Claude Code best practices	The most authoritative guide to working effectively with Claude Code: context hygiene, planning before implementing, CLAUDE.md, and the writer/reviewer pattern.
Claude Code — official docs	The reference for the CLI, IDE extensions, settings, skills, plugins, MCP, and subagents.
Anthropic — Building Effective Agents	The conceptual foundation: workflows vs. agents, the five workflow patterns, and the case for starting simple and adding complexity only when it pays.
Anthropic Cookbook (GitHub)	Runnable notebooks and copy-able recipes, including reference implementations for the agent patterns above.
Superpowers (obra/superpowers, GitHub)	The community agentic-skills framework behind the brainstorm → plan → execute → review chain referenced in the AI-assisted Tooling section.
Trail of Bits — Claude skills (GitHub)	A security-focused marketplace of Claude Code skills and plugins from Trail of Bits — smart-contract entry-point analysis, clarifying-question guards, and more; the source of this guide's security plugins.
Matt Pocock — skills (GitHub)	A practitioner's collection of composable, model-agnostic agent skills for real engineering — aligning with the agent, domain modeling, TDD, code review, architecture — a strong reference set to borrow from when writing your own.
everyinc — compound-engineering-plugin (GitHub)	Skills built around one idea: each unit of engineering work should make the next one easier, favoring planning and review over raw execution. The clearest articulation of the compounding theme behind Reusable Skills and Automation.
avoid-ai-writing — SKILL.md (GitHub)	A well-crafted example skill that detects and rewrites "AI-isms" — the tells that make prose read as machine-generated — with detect-only, edit-in-place, and rewrite modes. Worth reading as a model of skill authoring, not just for its output.
OpenAI — Codex CLI docs	The official reference for OpenAI's standalone terminal coding agent — install, AGENTS.md, approval modes, and MCP. A useful counterpart when you want a second model's take.
OpenCode — docs	The official docs for the open-source, provider-agnostic terminal agent — one tool across Claude, GPT, Gemini, and local models.

Open-source landscape (by GitHub stars)

A map of notable open-source AI tools and frameworks, grouped by what they do and ranked within each group by GitHub stars. Counts are a snapshot as of 1 July 2026, pulled from the GitHub API. Treat stars as a rough popularity-and-momentum signal, not a quality ranking — they move fast, and can be gamed, so re-check before quoting a number. Each link points to the repository's current canonical home (several projects have moved orgs).

Coding agents & terminal CLIs

Tool	Stars	What it is
OpenCode	~181k	Provider-agnostic open-source coding agent for the terminal — the project behind opencode.ai.
Gemini CLI	~106k	Google's open-source terminal agent built on Gemini.
Codex CLI	~95k	OpenAI's lightweight terminal coding agent.
OpenHands	~79k	Autonomous AI software-development agent (formerly OpenDevin).
Cline	~64k	Autonomous coding agent available as an SDK, IDE extension, or CLI.
Goose	~51k	Block's extensible on-machine agent that goes beyond code suggestions.
Aider	~47k	Git-native AI pair programming in your terminal.
Continue	~35k	Open-source coding agent and IDE autopilot.
Roo Code	~24k	A team of AI agents inside your editor (a Cline fork).

Token & context efficiency

Tool	Stars	What it is
caveman	~78k	Claude Code skill that trims ~65% of tokens — "why use many token when few token do trick."
ponytail	~70k	Nudges an agent toward the smallest change that works — "the laziest senior dev in the room."
rtk	~67k	Rust CLI proxy that cuts LLM token use 60–90% on common dev commands.

Agent frameworks & platforms

Tool	Stars	What it is
Dify	~147k	Production platform for building agentic workflows and LLM apps.
LangChain	~141k	The most widely used framework for composing LLM apps and agents.
browser-use	~102k	Lets agents drive a real browser to automate web tasks.
AutoGen	~59k	Microsoft's framework for multi-agent conversation and orchestration.
CrewAI	~55k	Orchestrates role-playing autonomous agents into collaborating "crews."
LlamaIndex	~51k	Data framework for RAG and document agents.

Local inference & model runners

Tool	Stars	What it is
Ollama	~175k	Run open-weight models locally with a single command.
llama.cpp	~119k	The C/C++ inference engine underpinning much local-LLM tooling.

Building LLM apps (tool use, structured output, evals)

Source	Why it's worth reading
Anthropic — Tool use with Claude	How to define tools, handle tool calls, and run the agentic loop; client tools vs. server tools. The doc you'll reference most when building an agent.
Anthropic — Prompt caching	Cut cost and latency for large reused context (system prompts, retrieved docs, tool definitions) by caching a stable prefix across calls.
OpenAI — Structured outputs	How to force schema-conformant JSON output (strict mode, refusals, SDK helpers); a clear treatment of the structured-output pattern.
OpenAI — Function calling	The companion to tool use from the other major vendor — useful for seeing the shared concept under a second name.
Anthropic — Create strong empirical evaluations	Practical guidance for building evals, including LLM-as-judge grading prompts and code-based graders.
OpenAI Evals (GitHub)	An open framework and registry for writing and running evals — a concrete starting point for measuring quality instead of eyeballing one run.
Langfuse — Security & guardrails	Practical patterns for defending a shipped LLM app — input/output guardrails against prompt injection, PII leakage, and harmful content, and how to trace them.
Anthropic — Claude Agent SDK	A higher-level toolkit for building agentic apps (tool loops, subagents, file/memory access) on top of the API.

MCP (Model Context Protocol)

Source	Why it's worth reading
Model Context Protocol — Introduction	The official starting point for the open standard that connects AI applications to external tools and data ("a USB-C port for AI").
Anthropic — Introducing the Model Context Protocol	The original announcement explaining the problem MCP solves and the host/client/server model.
MCP — specification & docs (GitHub)	The protocol specification and reference documentation, now community-governed under the Linux Foundation.

Practitioners & ongoing commentary

Unlike the sources above, these are individual practitioners publishing continuously — blogs and videos — rather than primary vendor docs or research; worth following for how the field is actually being used, but treat opinions as opinions.

Source	Why it's worth reading
Simon Willison's Weblog	The most consistent day-by-day chronicle of what's actually shipping in LLMs and agentic coding — hands-on, skeptical, and quick to separate real capability from hype.
Rinat Abdullin's blog	Field notes on shipping LLM-driven products in production — evals, schema-guided reasoning, and the engineering realities behind making these systems reliable.
Sean's AI Stories — AI Agent Harness & Loop Engineering in 19 Min (YouTube)	A 19-minute plain-English video tour of the whole agent stack — harness, loop, memory, RAG, tool calling, tracing, evals — covering the same ground as Integrating AI via API; a good primer for non-engineers.

---

Cited across this knowledge base

Every external source cited in the guide, deduplicated and grouped by the section it appears in. URLs are the current, verified ones (Anthropic docs normalized onto platform.claude.com and code.claude.com). Operational artifacts such as install scripts and example endpoints are omitted.

00 · Introduction

• Lost in the Middle (arXiv) — long-context attention research
• Lost in the Middle (TACL 2024) — peer-reviewed version
• Chroma — Context Rot — reliability degrades before window fills
• Context windows — how the window works
• Models overview — current Claude lineup

01 · Foundations

• Prompt engineering overview — when and how to prompt-engineer
• Prompt engineering for business performance — techniques and a case study
• Effective context engineering for AI agents — managing the context window
• Chroma — Context Rot — input length vs. performance
• Extended thinking tips — prompting for reasoning
• Extended thinking — how thinking works
• Structured outputs — schema-conformant JSON output
• Cookbook — structured JSON extraction — runnable extraction notebook
• Artificial Analysis — models leaderboard — independent model benchmarks
• Claude Code — IAM & permissions — identity and access controls
• Claude Code — settings — settings reference
• OWASP — prompt injection (LLM01) — the top LLM risk

02 · Workflows

• git worktree — multiple working trees, one repo
• Claude Code — subagents — delegated, isolated context
• Superpowers (obra/superpowers) — agentic skills framework
• Claude Code — skills — packaged repeatable workflows
• Equipping agents with Agent Skills — the skills announcement
• everyinc — compound-engineering-plugin — each unit of work makes the next easier
• avoid-ai-writing — SKILL.md — example skill; strips machine-generated writing tells
• Context7 (upstash/context7) — up-to-date docs MCP server
• Context7 — hosted service — the hosted Context7 endpoint
• Context windows — window mechanics
• Claude Code — context window — interactive context visualization

03 · AI-assisted Tooling

• Claude Code — overview — what Claude Code is
• Claude Code — quickstart — your first task
• Claude Code — permission modes — controlling autonomy
• Claude Code — memory — CLAUDE.md and auto memory
• Claude Code — settings — settings reference
• Claude Code — skills — creating and using skills
• Claude Code — plugins — bundling skills and config
• Claude Code — discover plugins — finding plugins
• Claude Code — commands — built-in commands
• Claude Code — MCP — connecting external tools
• Claude Code — MCP quickstart — first server end to end
• Claude Code — hooks — shell commands around actions
• Claude Code — environment variables — configuration via env
• Claude Code — subagents — delegated workers
• Official Claude plugins — the official marketplace
• Trail of Bits — Claude skills — security-focused skills marketplace
• Matt Pocock — skills — practitioner's everyday-engineering skill collection
• Model Context Protocol — introduction — the open standard

04 · Integrating AI via API

• Building Effective Agents — workflow and agent patterns
• Effective context engineering for AI agents — managing context
• Prompt caching — cut cost on reused prefixes
• Lessons from building Claude Code: prompt caching — cache-stability field guide
• Tool use overview — defining tools and the loop
• Tree-sitter — syntax-aware code chunking
• Contextual retrieval — better RAG retrieval
• Context windows — window mechanics
• Claude Agent SDK — higher-level agent toolkit
• Structured outputs — guaranteed schema output
• Build strong evaluations — designing and grading evals
• Console evaluation tool — testing prompts in the Console
• A statistical approach to evals — statistical rigor for evals
• Langfuse — observability overview — tracing LLM apps
• Langfuse — get started — instrumenting your app
• Sean's AI Stories — AI Agent Harness & Loop Engineering in 19 Min (YouTube) — plain-English video tour of the agent stack

---

Related

• Glossary — definitions for the terms these sources use
• Checklists — the practical habits these sources support
• AI limitations — the frame the whole knowledge base hangs on
• This knowledge base's own Foundations and Workflows pages are the primary, opinionated reference for teams here; the links above provide depth and authority behind them.