Made by David, Built with AI

What I learned, and where I used it

Six Anthropic Academy courses, completed between 14 September and 26 October 2026. Several projects were already under way when I took them, so this page shows where I have practised each topic. It does not claim the courses caused the work.

AI capabilities and limitations

Completed 26 October 2026

A working mental model of how generative AI behaves and why: next token prediction, knowledge, working memory and steerability, and what happens when those properties collide.

Next token prediction: why answers vary
  • Pension Planner: The report is built by plain rules with no AI call, so the same inputs always give the same answer.
  • The ‘Defender’ sites: Claude is called through a forced schema, and its output is validated and retried if it does not fit.
Knowledge: what the model does and does not know
  • The ‘Defender’ sites: Prompts forbid inventing facts, and a second AI fact-checked the content.
  • Star Chaser: Facts that could not be confirmed were left out, and scripts caught fifteen photos of the wrong restaurant.
  • Pension Planner: Tax figures were checked against official sources, and figures that could not be verified are flagged.
Working memory: what stays in the context
  • Recruiter tracking agent: A run log carries what matters from one morning to the next, so each run starts informed.
  • Star Chaser: Research was split into batches sized to what the agents could handle, with a lighter brief for lower-priority work.
  • Pension Planner: A separate session with no memory of the build reviewed the report, so it could not repeat the author's assumptions.
Steerability: guiding behaviour with instructions
When properties collide
  • The ‘Defender’ sites: The appeal pack needs a cautious analysis and a confident statement in one call, so the prompt keeps the two apart.

AI Fluency: Framework and foundations

Completed 8 October 2026

A framework for working with AI effectively, efficiently, ethically and safely, built around four skills: Delegation, Description, Discernment and Diligence.

Delegation: deciding what the AI does and what I keep
  • Recruiter tracking agent: The agent reads, calculates and drafts. I send every message and I alone mark a contact as dropped.
  • Fire TV home screen: A build plan with a rollback for every step, and a rule that no home automation flow changes without asking me.
Description: telling the AI exactly what I want
  • Recruiter tracking agent: A 13-step instruction set with ordered decision rules instead of loose guidance.
  • Fire TV home screen: A structured brief for the vacuum build: scope, guardrails, stage-by-stage checks and a definition of done.
  • ‘butterflai’ website: Written brand rules as the specification for every generated image.
  • Pension Planner: A fix brief written by the reviewer: plan before code, opt-in changes, acceptance figures and regression tests.
Discernment: judging whether the output is right
  • The ‘Defender’ sites: Content fact-checked by a second AI from a different vendor.
  • Pension Planner: A separate Claude asked to review a generated report for errors, plus 154 tests and gov.uk checks.
  • Star Chaser: Verification scripts that caught fifteen photos of the wrong restaurant.
Diligence: taking responsibility for AI-assisted work
  • The ‘Defender’ sites: Scope guards, no legal opinions and no merits scores. The packs state what they cannot do.
  • Pension Planner: Figures that could not be verified are flagged, and the tool says it is not financial advice.
  • Recruiter tracking agent: Preview mode before go-live, a full audit trail and human approval on every outbound message.
The Description and Discernment loop

Claude Code 101

Completed 7 October 2026

Using Claude Code in daily work, from the explore, plan, code and commit workflow to subagents, skills, MCP and hooks.

Explore, plan, code, commit
  • Fire TV home screen: Research of existing projects first, then an approved twelve-step plan, then the build.
  • Star Chaser: A standing rule to show a plan before doing anything non-trivial.
Context management and the CLAUDE.md file
  • Star Chaser: A CLAUDE.md that holds the working rules, build process and naming decisions, so any new session starts informed.
  • Pension Planner: A long CLAUDE.md covering the architecture, report build and Android build.
  • The ‘Defender’ sites: Each site carries its own CLAUDE.md, with security and SEO audit files kept alongside the code.
Subagents
  • Star Chaser: Parallel research agents for lists, synopses and images, on different models for different jobs.
Skills
MCP
Code review
Hooks
Not used in a project yet.

Claude Platform 101

Completed 7 October 2026

Building on the Claude Platform: API requests, the agent loop, tool use, built-in tools, skills, MCP, context management and managed agents.

API requests
Choosing the right model
  • Star Chaser: Larger models for pages that needed careful reading, small fast ones for one-search-per-restaurant sweeps.
The agent loop and context management
  • Recruiter tracking agent: A daily loop that reads, decides, acts and checks its own work, with a run log to carry state between runs.
  • Job search agents: A scheduled loop that scrapes, ranks and reports, with a record of jobs already seen.
Tool use
Managed agents
Not used in a project yet.

Introduction to Claude Cowork

Completed 6 October 2026

Working with Claude on real files and projects: the task loop, standing context, skills, plugins, Claude in Chrome and working safely.

Standing context: global instructions and projects
Multi-step work and scheduled tasks
Claude in Chrome
Working safely
  • Recruiter tracking agent: Read-only apart from drafts, tagged calendar events and treating email content as data, not instructions.
  • Fire TV home screen: Plan approval before changes and no credentials typed into chat.
Plugins
Not used in a project yet.

Claude 101

Completed 14 September 2026

Everyday use of Claude: conversations, projects, skills, connectors and research.

Research mode
Skills
Connectors