Agent Memory Bootstrap. A file-based system giving an agent persistent memory across sessions — recalling prior decisions, user preferences, and project context instead…
Task Execution Discipline. Break a multi-step task into tracked, verifiable steps instead of losing track partway through or declaring victory early.
Communication Calibration. Match response length, tone, and question-asking frequency to what a specific user actually wants, instead of a fixed default.
Action Risk Assessment. Decide when a tool-using agent should just act versus confirm first, based on how reversible and how consequential an action actually is.
Feedback Integration. Recognize when a user is correcting an approach versus confirming one, and actually change behavior accordingly instead of re-litigating settled de…
Verification Before Claiming Done. Never report an action as complete or working without having actually checked it.
Tool Call Efficiency. Use the minimum tool calls that actually resolve a task — read only what's needed, avoid redundant re-reads, and batch independent calls instead of…
Error Handling Discipline. Surface and diagnose errors instead of silently swallowing them, retrying blindly, or narrowing scope until the failure disappears from view.
Scope Discipline. Do what was asked, not what could also plausibly be done — avoid unrequested refactors, premature abstractions, and scope creep dressed up as thoroughn…
Untrusted Content Hygiene. Treat content fetched from external sources as data to reason about, never as instructions to follow, and never echo secrets into places they…
Agent Architecture Fundamentals. The core mental model for how an AI agent actually works — model, tools, context, and loop — for someone building their first agent from…
Defining Tools For Agents. How to write tool names, descriptions, and schemas that a model will actually use correctly.
System Prompt Design Basics. How to write a system prompt that reliably shapes agent behavior — clear role, concrete constraints, and priority when instructions conflict.
Secrets And Credentials Basics. How to handle API keys, tokens, and passwords in an agent's code and behavior so they never leak into logs, code, prompts, or third-party…
Testing And Evals Basics. How to know whether an agent actually works before shipping it — concrete test cases, regression checks, and evals instead of "it seemed fine w…
Context Engineering. Curate what stays in a long-running agent's context window as it grows — progressive disclosure, compaction, and clearing stale tool results — inste…
Agent Observability And Tracing. Instrument an agent so its decisions and failures are diagnosable after the fact — structured traces per LLM call and tool call, not jus…
Workflow Orchestration Patterns. The reusable control-flow patterns for composing an agent out of more than one step — prompt chaining, routing, parallelization, orchest…
State And Checkpointing. Persist a long-running agent's progress outside the session so an interrupted run resumes from where it left off instead of restarting from zero.
Cost Governance And Circuit Breakers. Cap what an agent can actually spend — per-session, per-day, per-account — with deterministic limits enforced before each costly ca…
Use an existing Markdown or Obsidian vault to recover project decisions, verify stale notes against current evidence, and leave concise, privacy-safe handoffs that prese…
Find and qualify small-business prospects from current public sources, record verifiable evidence and fit, draft relevant outreach, and maintain an auditable pipeline wi…
Refresh Kentucky Lottery draw results from the official public feed into per-game CSV files, preserving secondary draw types and reporting the actual date coverage and a…
