, gotchas, and workflow recipes — clustered by theme so you can skim the area you need. Opinionated and plainspoken.
How to ask — phrasing, structure, what to include.
Tip · Prompting
Claude will happily turn a question into a . Add an explicit boundary — "report findings and stop" for questions, "go ahead and implement" for work — and you stop getting changes you didn't ask for.
Tip · Prompting
Claude leans towards agreeing with whatever you propose. Ask "is this a good idea?" and you'll usually hear yes. Ask it to argue the other side, or to find the three biggest problems, and you get the judgement you were actually after.
Tip · Prompting
Role-play preambles were useful for older, weaker and are now dead weight. Replace the persona with three things that actually help — where to look, what "done" means, and how Claude should check itself.
Tip · Prompting
Two phrasing habits that quietly damage — framing instructions as prohibitions ("do NOT use tables") and emphatic shouting ("CRITICAL - you must ALWAYS"). Both work less well than plainly describing the outcome you want.
Tip · Prompting
The upgrade that works on every report, CSV, or checklist Claude produces — "turn this into a single self-contained interactive HTML file I can open by double-clicking." No server, no installs, no IT ticket; the file travels by email and opens in any browser.
Tip · Prompting
Drop "think harder" or "ultrathink" into a and Claude reasons longer before answering. Use `/effort max` for hard debugging, architecture decisions, or anything where speed matters less than getting it right.
Tip · Prompting
Claude is multilingual. in Greek when it's the language you can think in most clearly. The replies in the language you wrote — and code/identifiers stay in English, which is exactly what you want.
Tip · Prompting
A one-line gets one-line-quality work. Brief Claude with Role + Goal + Task + Constraints + Context — the same shape you'd hand a colleague who's helping with the task.
Tip · Prompting
A useful names the file, the symptom, the repro, and the desired action. Vague prompts get vague results. The five-part shape gives Claude what it needs on the first try.
Tip · Prompting
When Claude misunderstood, lead the correction with 'No, what I meant was…' — it's shorter than re-explaining everything and Claude takes the cue cleanly.
Tip · Prompting
Tell Claude what you want the outcome to be, not which functions to call. Claude is better at finding the right code path than you are at prescribing it.
Tip · Prompting
When Claude's answer is wrong, don't `/clear` and start over. Tell it what's wrong and let it adjust — you keep all the context you've already built up.
Tip · Prompting
A with five asks gets five mediocre answers. Break work into one focused step at a time, accept it, then move to the next.
Tip · Prompting
When you really do need to constrain Claude, pin the *requirements* (must run in under 200ms, must work offline, must not break the public ) — not the implementation choices Claude can pick freely.
Tip · Prompting
A screenshot of a broken UI beats four paragraphs of describing it. Drag the image into the — Claude will see what you see.
The keys and modes that put you back in the driver seat — stop, rewind, stash, and switch what Claude is allowed to do. Claude Code calls these Shortcuts too.
Tip · Shortcuts
Beyond Esc there are about seven shortcuts that change how the feels — search your own history, park a half-written , open the editor, see what Claude actually did, and background a long task. All verified against 2.1.
Tip · Shortcuts
Permission fatigue has a better cure than skipping entirely. /fewer-permission- pre-approves the safe read-only commands you already use, and /permissions lets you write deny rules that block things you never want touched — deny always wins.
Tip · Shortcuts
/rewind restores the code and conversation to an earlier point. It beats arguing with a that has gone off the rails. Critical caveat - rewind does not undo files deleted by ; that needs .
Tip · Shortcuts
The flag the team uses as its default Day-1 launch command. Skips read/ permission but does NOT skip code-edit reviews. Knowing the precise boundary defuses the "wait, isn't this dangerous?" anxiety.
Tip · Shortcuts
Press Shift+Tab in the to cycle Claude's . Auto mode is now the default — Claude vets each action itself and only asks about the risky ones. is read-only; accept-edits lets edits through while still pause.
Tip · Shortcuts
When Claude is going down the wrong path, press Esc to halt it immediately. This single key prevents most "Claude just deleted everything" horror stories.
How to drive Claude Code — plan-first, slash commands, sessions, subagents, hooks. The shortcuts that compound.
Tip · Workflow & commands
ships with an interactive tour of its own features. Type /powerup in any and it walks you through the things most people never find. It is the single highest-value five minutes a newcomer can spend.
Tip · Workflow & commands
Switching mid- is the standard move when you're stuck, and it's usually the wrong one — it costs you the cached context and carries the same wrong assumptions into the new model. Pick the model at the start; hand off to a fresh session if you need to change.
Tip · Workflow & commands
A summary is written for a human and is nearly useless to the next . A handoff is written for Claude — it points at files, names what was already tried and failed, and says where to start. Ask for a handoff instead and the next morning starts warm.
Tip · Workflow & commands
Asking the that just built something to check its own work gets you a confident pass. The session carries every assumption it made while building. Open a fresh session, give it the diff and no history, and ask it to find what's wrong.
Tip · Workflow & commands
/btw asks a side question without derailing the , / forks the conversation, /recap tells you where you were, /context shows where your went, /doctor cleans up your setup, and /cost answers "what is this costing me".
Tip · Workflow & commands
For one day, note every time you switch away from to look something up. Each switch is a piece of context Claude didn't have. The list you end up with is a ranked to-do list for making it more useful — and you don't need to write any code to produce it.
Tip · Workflow & commands
After a few rounds of "not quite, try again", quality often goes down rather than up — each pass makes cosmetic changes to satisfy the feedback. Recognise the plateau, go back to the best version, and switch tactics instead of iterating.
Tip · Workflow & commands
The strongest verification habit in the hub — after Claude produces an output, ask it to verify its own work with real commands (grep the quotes, recompute the sums, re-run the logic) and show you the output. Commands either pass or fail; assurances always pass.
Tip · Workflow & commands
Once you're past the plain- basics, the next decision is which orchestrator to reach for. /team runs a feature end-to-end in one sitting. -* manages a multi-week with persistent state. Pick by horizon, not by ambition.
Tip · Workflow & commands
Type `@` to reference a file by path (autocompletes), `!` to run a and embed its output. Both go straight into the . They cut the "let me copy this in" loop entirely.
Tip · Workflow & commands
If you find yourself typing the same three times, write it as a custom . Drop a file in `.claude/commands/`, invoke with `/yourname`. it and the whole team gets the shortcut.
Tip · Workflow & commands
is advisory — Claude follows it about 80% of the time. are deterministic — they run every time, no exceptions. If something must happen 100% of the time (lint, security check, format), make it a hook.
Tip · Workflow & commands
Claude is fluent with tools like ``, `az`, `aws`, `jq`, ``. Install them, log in once, and Claude can read , query Azure resources, inspect — without a single server.
Tip · Workflow & commands
Long mixed-topic confuse Claude and burn context. Finish the auth refactor, then `/clear` and start the cart bug fresh. Sessions are cheap; misremembering isn't.
Tip · Workflow & commands
The single biggest mistake newcomers make is letting Claude jump straight into coding. Use to align on the approach first; the code you don't have to undo is the cheapest code you'll ever write.
Tip · Workflow & commands
You don't have to lose context overnight. `claude --continue` picks up your last ; `claude --resume` opens a session picker. The codebase knowledge Claude built up is worth more than the 30 seconds of re-grounding.
Tip · Workflow & commands
A runs in its own , does its exploration, and reports a summary. Your main never loads the 40 files it had to read. Best for research-heavy tasks where the input is large and the answer is small.
Tip · Workflow & commands
Boris Cherny's ('s creator) golden rule — every time Claude makes the same mistake twice, add a one-liner to so it doesn't make it a third time. The file grows as your 's tribal knowledge gets written down.
Keeping Claude focused — what stays in the conversation, what should be cleared.
Tip · Context discipline
is read once, when the starts. Adding a rule to it in the middle of a conversation has no effect until you start a new session — which is why people conclude CLAUDE.md "doesn't work" when it does.
Tip · Context discipline
A rule like 'keep the code organised sensibly' means nothing to Claude; 'all SQL lives in database/' is actionable. The difference between a that works and one that gets ignored is almost always specificity — plus pruning the rules that have quietly gone stale.
Tip · Context discipline
When Claude starts making things up, the bug is almost always in your , not the . Stale sections, leftover scratch files, orphaned READMEs — Claude reads everything and the noise leaks into its answers.
Tip · Context discipline
A worked example of a -level , distilled from 's official guidance, Karpathy's viral behavioral rules, HumanLayer's "under 60 lines" guide, and the awesome-claude- curated index. Each section is annotated with why it earns its slot.
Tip · Context discipline
Don't carry irrelevant history. Use `/` to summarise the conversation when staying on-task, `/clear` to wipe and start fresh when switching tasks. Know that compacting trades detail for room — for work you care about, a handoff to a fresh preserves more.
Working with bank data and external services — what is allowed and what is not.
Tip · Compliance & safety
Claude is fast and capable — treat it like a brilliant junior. Read what it's about to change before saying yes. Skipping the diff review is the most expensive habit a newcomer can pick up.
Tip · Compliance & safety
If a tool will process real bank data at runtime, its calls MUST go through Azure OpenAI. The public and Gemini are off-limits for real customer data. This is data-residency, not a preference.