Six ordered skills that make AI coding agents reliable enough for real work. Drop them into Claude Code or Cursor and run any task through them in order.
AI doesn't fail because it isn't smart enough. It fails because we don't give it structure.
You've felt it: the model that can ace a math olympiad chokes on a simple refactor in your codebase. That's not a failure of intelligence — it's a failure of instruction. The fix isn't a cleverer prompt. It's a repeatable process that transfers understanding to the AI.
Treat the AI like a brilliant new intern: vast knowledge, but blind to your context and too eager to please to admit when it's guessing. You'd never tell a new hire "make it better" and walk away. This framework is how you brief it properly — every time.
Run them in order, or jump in where you are. Each step hands the AI one thing it was missing.
| Step | Skill | Hands the AI |
|---|---|---|
| 1 | /s1-explore — understand the system before touching it |
Context |
| 2 | /s2-clarify — ask the questions that kill ambiguity |
Precision |
| 3 | /s3-define — lock the problem, constraints, success criteria |
A target |
| 4 | /s4-solutions — compare 3 approaches, pick the simplest |
A strategy |
| 5 | /s5-plan — write a phased plan you can resume from |
Memory |
| 6 | /s6-implement — build it test-first, then verify the result |
Accountability |
The magic isn't any single step — it's that together they force you to transfer understanding instead of hoping a prompt lands.
The same chain works for any complex problem you'd hand a sharp collaborator — strategy, planning, writing, analysis, research. Explore the landscape, clarify the ask, define success, weigh options, plan, execute and verify. Code is just where the payoff is easiest to see.
git clone https://github.com/evanscastonguay/agentic-problem-solving.git
cp -r agentic-problem-solving/skills/* ~/.claude/skills/They sort s1…s6 in your skill picker, so the order is always in front of you. Any tool that reads markdown command files works (Cursor, etc.).
- Treat AI like a brilliant intern — vast knowledge, but context-blind and eager to please.
- Delegate outcomes, not tasks.
- Trust evidence, not confidence — make it verify its own work.
MIT © 2026 Evans Castonguay — use it, fork it, ship it.
If you run a real task through this, open an issue and tell me what broke and what worked. I improve it in the open.