Techniques
Mental models tell you where to push. Techniques are the actual phrases that do the pushing — the openings, tactics, and endgame moves of this craft.
Chess players build a repertoire — openings, tactical patterns (forks, pins, skewers, discovered attacks), endgame studies. Most of it learned from other players, none of it secret. The same vocabulary is forming for prompting. Here are some moves worth adding to yours.
Before you start
What questions should you ask me before starting?
Surfaces ambiguity instead of guessing. The model returns a list of things it's uncertain about; you answer them, then it goes. Saves a lot of clarifying turns later. Especially good with smaller, faster models.
What's the smallest version of this that ships?
Scopes down before delegating. Ten lines instead of a thousand. The model is happy to grow scope; you have to ask it to shrink.
Open the options
Give me 3 options, rank them, name the trade-offs.
Forces breadth before depth. The model defaults to its first plausible idea; this resets the floor. Useful for any decision where you don't already know the right answer — algorithm choice, library choice, schema design.
Argue against your last suggestion.
Surfaces hidden assumptions. Especially useful right after a plan — much cheaper to discover the plan's weaknesses now than after you've implemented it.
Pressure-test it
What did you skip?
After a delegated task, ask what was glossed over. The model often owns things you'd have missed — tests, error handling, the "TODO: revisit" it left in line 47. Cheap two-second move, high hit rate.
How would a senior engineer review this?
Triggers review-mode output. Different vibe than implementation-mode — more critical, more "but consider" tradeoffs, more willing to call out things its implementation-self would have left in.
Pay it forward
Match the style of components/PostCard.tsx.
Anchor on something concrete — a path, a screenshot, a doc, a working example. "Match this" beats "make it look nice" by a wide margin. Models are imitators first; give them something to imitate.
Update CLAUDE.md with what you just learned.
Capture the lesson in the project rules so the next agent (or future-you) doesn't have to relearn it. Five minutes of investment that pays back forever.
Lift them into skills
Once you find a move that works, stop typing it manually. Every serious agent tool lets you write project-level instructions that the agent reads on every task — you write the lesson once, the agent has it forever.
The techniques above turn into one-liners in that file. A useful rules file looks something like:
- Before any non-trivial task, ask what's unclear before writing code. - When asked to plan, propose 3 options ranked by trade-off. - After delegating, list what was skipped or glossed over. - Match the patterns in components/PostCard.tsx for new card UIs.
Now those moves are the agent's defaults instead of yours. You stop carrying the techniques in your head and start accumulating an edge that compounds with every task.
Keep climbing
None of these are hard. Used regularly, they compound. Don't memorize them as a checklist — notice there's a craft here, and the shape is still being mapped.
We're all early at this. The next thousand ELO points are wide open.