Read outward without losing the course thread.
HeatStack reviews selected third-party material against a pinned revision, writes a bilingual guide, and maps it back to a Forge module. Clearly licensed works may also have an attributed local reading edition.
Awesome AI Pedia
A Chinese-language collection of AI engineering, tools, and practices. HeatStack publishes only individually reviewed reading guides; it neither syncs the directory nor treats repository Markdown as installable Skills.
The repository root declares MIT, but upstream rights may not be clear for every aggregated article. HeatStack therefore provides rewritten summaries and outbound links, not reproduced bodies.
- 01Open original
Harness Engineering for Long-running Development
Uses planner, generator, and evaluator roles to discuss handoffs, verification, and context resets in long-running work.
- Best used for
- Useful for readers building planner/executor or evaluation loops.
- HeatStack reading note
- Use the separation of roles and independent verification; treat timing and outcome claims as an unverified case, not a general guarantee.
- 02Open original
MCP and Skills: Different Responsibilities
Distinguishes MCP as a connection to external capabilities from Skills as task guidance, and shows how they can work together.
- Best used for
- Useful for readers beginning to separate protocols, tools, and capability packages.
- HeatStack reading note
- Treat it as a simplified mental model; MCP can also expose resources and prompts, while the meaning of Skill varies by platform.
- 03Open original
Maintaining CLAUDE.md
Discusses layering, size control, actionable rules, and ongoing maintenance for project instruction files.
- Best used for
- Useful for people maintaining AGENTS.md, CLAUDE.md, or cross-platform project constraints.
- HeatStack reading note
- Adopt the short, explicit, verifiable, non-conflicting maintenance principles; verify exact loading behavior and numeric claims against current official docs or the cited paper.
AI Agent Book
Bojie Li's systems-oriented AI agent book, spanning foundations, context, memory, tools, evaluation, continual evolution, and multi-agent systems. Under Apache-2.0, HeatStack hosts a pinned bilingual reading edition mapped to the existing curriculum.
The project declares Apache-2.0. The local edition changes only heading attributes, links, asset paths, and page metadata for web rendering; it intentionally omits conflicting experiment-count claims.
- 01
Chapter 1: Agent Foundations
Builds a basic systems model of agents, environments, actions, and feedback loops.
- Best used for
- Calibrate the core concepts before implementing the Forge loop.
- HeatStack reading note
- Compare the abstractions with HeatStack's task-selection boundaries, then turn the loop into testable single-agent state transitions.
- 02
Chapter 2: Context Engineering
Examines how context selection, organization, and budgets shape agent behavior.
- Best used for
- Place prompt techniques inside a complete context system.
- HeatStack reading note
- Return to the Forge structured-output lab and decide what belongs in context versus what tools should retrieve on demand.
Related modulesPrompts, Context & Structured Output - 03
Chapter 3: Memory and Knowledge
Distinguishes retrieved knowledge, session state, and memory across time scales.
- Best used for
- Avoid treating RAG, state, and long-term memory as one storage mechanism.
- HeatStack reading note
- Use the chapter's taxonomy to inspect Forge data boundaries, then verify retrieval, deletion, and recovery semantics locally.
- 04
Chapter 4: Tools
Frames the boundary between agents and external systems through tool descriptions, calls, and feedback.
- Best used for
- Add a systems-design view beneath tool protocols.
- HeatStack reading note
- Do not stop at successful invocation; map every tool to permissions, failure, idempotency, and side-effect control.
- 05
Chapter 5: Coding and General Agents
Compares capability packaging and execution patterns in coding and general-purpose agents.
- Best used for
- Understand why Skills, planning loops, and platform adapters solve different layers of the system.
- HeatStack reading note
- Inspect capability packages, execution graphs, and platform differences against one Forge task instead of collapsing them into one giant prompt.
- 06
Chapter 6: Interaction and Multimodality
Covers agent interaction surfaces and collaboration beyond text-only input and output.
- Best used for
- Add a human-interaction perspective to the unified input contract.
- HeatStack reading note
- Record parsing failures, privacy, and accessibility needs for each modality, then add explicit rejection paths to Forge.
Related modulesModel Mechanics & Multimodal Input - 07
Chapter 7: Evaluation
Turns agent quality into observable and reproducible evaluation problems.
- Best used for
- Expand beyond answer accuracy to process, cost, and failure diagnosis.
- HeatStack reading note
- Map the evaluation dimensions to Forge fixtures, traces, and regression gates rather than substituting a one-off demo for evidence.
- 08
Chapter 8: Post-training
Introduces how post-training changes model capability, preferences, and task performance.
- Best used for
- Provide optional advanced background for model selection.
- HeatStack reading note
- Use it to explain model differences, not to assume a model fits Forge; validate the choice with benchmark tasks.
Related modulesAI Tools, Models & Task Judgment - 09
Chapter 9: Continual Evolution
Discusses how agents improve through feedback, evaluation, and engineering processes.
- Best used for
- Turn a one-off delivery into a maintainable evolution system.
- HeatStack reading note
- Use Forge versions, rollback, and evidence indexes to test whether continual improvement is auditable, not merely whether outputs look better.
- 10
Chapter 10: Multi-Agent Systems
Explains multi-agent systems through roles, communication, and coordination mechanisms.
- Best used for
- Add a broader structure to the multi-agent orchestration lab.
- HeatStack reading note
- For every added role, record the single-agent failure it addresses, its communication cost, and its conflict-handling rule.
Related modulesMulti-Agent Collaboration & Orchestration