Course Overview
Most AI coding courses stop once participants can prompt a tool into building a prototype. This course goes further: participants build a small resident coding and workflow agent in Claude Code or Codex that can remember project context, use tools, run on a schedule, keep logs, and pause for human approval when the work gets risky.
The course is designed for participants who have already seen the baseline ideas behind AI coding: prompt and context engineering, repo instructions such as AGENTS.md or CLAUDE.md, file search, tool choice, skills, MCPs, subagents, code review, testing, documentation, and worktrees. We revisit those ideas through one coherent build instead of treating them as disconnected revision slides.
The build can be framed around a theme such as Humanitarian Assistance and Disaster Relief (HADR), where an agent researches a developing incident, monitors selected public sources, prepares cited updates, and generates reports. The exact theme may vary, but the pattern stays the same: a real workflow with sources, changing context, outputs, logs, tests, diagrams, and approval gates that a human can inspect.
Lesson Outcomes
By the end of this course, participants will have built and reviewed a working mini-agent system that can:
Maintain useful project memory. Store repo instructions, task notes, decisions, source summaries, and reusable operating procedures without stuffing every prompt with stale context.
Use tools with clear boundaries. Decide what belongs in a skill, what belongs in an MCP or tool integration, and what should stay as plain repo instructions.
Run a research and monitoring workflow. Collect source material, track changes, produce situation updates, and generate a final report with citations and human review points, using HADR or another comparable theme as the course context.
Control token and model use. Route planning, execution, review, and documentation tasks to suitable models; reset context deliberately; and keep skills and instructions lean.
Add production guardrails. Apply sandboxing, credential handling, prompt-injection checks, approval gates, audit trails, and rollback plans to an agentic workflow.
Document the system for human inspection. Generate tests, review notes, runbooks, and C4-style system views that make the agent's behaviour easier to check.
Curriculum
Day 1: Build the smallest useful mini-agent
Participants start with a compact but working resident agent. The build reinforces the baseline: prompting, context, repo instructions, memory, compaction, model and tool choice, code quality, and review loops. The goal is to get a small agent running early, then improve it with evidence instead of designing a giant system on paper.
Set up the working repo, instructions, task log, and review loop.
Create the first memory layer for source notes, project decisions, and scenario facts.
Build the first end-to-end task: ingest a small set of sources, summarise them, and produce a short update for human review.
Review the agent's work with tests, checklists, and a second model pass.
Day 2: Add tools, autonomy, and model discipline
The second day turns the agent from a prompt-driven assistant into a workflow resident. Participants add one or two useful integrations, scheduled or heartbeat runs, custom skills, and model routing rules. The course theme shifts from one-off research into monitoring and report generation.
Design a custom skill for repeatable research, source checking, or report drafting.
Add a tool or MCP-style integration for file search, source ingestion, issue tracking, or report output.
Implement a scheduled or heartbeat run that checks for updates and writes an auditable log.
Practise token and model discipline: planning model, execution model, review model, reset points, and context hygiene.
Day 3: Harden, document, and present
The final day treats the mini-agent like a system someone else might have to maintain. Participants add guardrails, write runbooks, generate architecture views, test failure cases, and present a production-readiness verdict rather than only a demo.
Add approval gates for sensitive actions, credentials, source trust, and generated reports.
Test prompt-injection risks, confused-deputy failures, runaway loops, stale memory, and bad source handling.
Generate architecture and data-flow diagrams that a non-author can inspect.
Prepare a final monitoring report, runbook, token log, and demo.
Details
Dates and Times: 7-9 July 2026, 3 full-day sessions.
Location: CT Hub 2, Lavender.
Requirements: Bring your own laptop. Modern Mac, Windows, or Linux laptops are supported. Participants should be comfortable using a terminal, Git, GitHub, and at least one AI coding tool such as Claude Code or Codex.
Prerequisites: Participants should have completed AIxTech or have equivalent hands-on experience with AI coding tools, prompt engineering, context engineering, basic software review, and testing/debugging.
Tools and accounts: The course uses Claude Code or Codex as the main coding harness, with comparisons to other agentic coding tools where useful. The course fee includes a 12-month subscription to a 5x plan for each participant, using either Claude Code or Codex depending on final course setup.
Fees (excl. GST): S$2,600 per participant for the 3-day course, including the year-long 5x Claude Code or Codex subscription. Minimum 8 participants to start.
All quoted prices are in SGD. Invoicing terms available.