Mých 10 tipů pro Claude Code (a vývoj s AI)

Na minulých setkáních se objevovaly různé tipy a střípky týkající se vývoje s AI, ať už ve formě lightning talků či příspěvků v diskuzích. Inspirovalo mě to ke shrnutí poznatků a tipů, které se mi v praxi osvědčily.

Tipy se budou týkat různých oblastí: zlepšení bezpečnosti, rozhraní pro práci s více paralelními agenty, prompting, pomocné skilly, pracovní postupy a další. Naprostou většinu využijete i s jinými modely a nástroji!

Transcript

1. Pyvo #52 Mých 15* tipů pro Claude Code ft Jan Smitka Tech Lead & So ware Engineer
2.   Ptejte se v průběhu
3.  Diskutujme po přednášce
4.  Setup
5.   1. Sandbox
6. • Zápis pouze do složky projektu, vše ostatní pouze pro čtení. • Zakázán přístup do internetu - localhost a interní IP adresy povoleny. Whitelist domén pomocí proxy. • Zakázán přístup k UNIX socketům. • Platí i pro příkazy, které agent spouští.
7. ~/.claude/settings.json { "sandbox": { "enabled": true, "network": { "allowedDomains": ["github.com", "api.github.com", "*.npmjs.org"], "allowUnixSockets": ["/tmp/.s.PGSQL.5432"], "allowLocalBinding": true }, "f lesystem": { "allowWrite": ["~/.cache", "~/.npm", "~/.pip", " tmp"], "denyRead": ["~/.ssh", "~/.kube", "~/.conf g/gcloud"] }, "excludedCommands": ["docker *"] } / / i i }
8. ~/.claude/settings.json { "permissions": { "defaultMode": "auto" } }
9.   2. Amphetamine
12.   3. Multiplexing
13. cmux
14. herdr
15. T3 Code
16.   Prompting
17.   4. Give me options
18. Review the implementation of `bulk_insert_product_inputs` and give me options how we can speed it up.
19.   5. Use your judgement
20. Implement @specs.md. Commit during your work, then create a PR. Use your judgment.
21.  6. Use subagent
22. Subagenti • Izolovaný kontext - "nezanáší" hlavní vlákno. • Agenti mohou běžet paralelně. • Claude rád spouští agenty na exploraci. • Mohou být různých typů - různé globální instrukce.
23. 150k dumb zone kvalita velikost kontextu
24. Use a subagent to implement the score column - use the same backing data as seen on the project report view. Add the logic, data to API, and update the frontend.
25. For all coding tasks use your judgement to decide an appropriate lower power model and run that in a subagent Simon Willison - Fable's judgment simonwillison.net/2026/jul/3/judgement/
26.   7. Use workflow
27. Workflow • JavaScript řídí spouštění subagentů a předávání výstupů mezi nimi. • Skript je deterministický, běh agentů lze pozastavit a pak zase spustit.
29. Validate and resolve pending review comments on the current PR from CodeRabbit. Read the comments and group same issues on di erent files together. Then use workflow to resolve each issue group sequentially, so the subagents never interfere with each other. Use "Prompt for AI Agents" from each comment as instructions for each subagent. Changes made by each subagent will be commited as a single commit. If a review comment is invalid or rejected, prepare a reply to the corresponding comment on GitHub, but do not post it yet. ff ft A er all fixes are in, push the changes and post ALL comment replies at once. Don't mark any comments resolved - it is expected that CodeRabbit will check the comments and resolve the comments.
30. I need to find user-facing strings in the project that are currently not marked as translatable. These can appear: •… Explore the folder structure, divide it to clusters containing related files, and then use a workflow with Sonnet 5 agents on high reasoning to find untranslated strings in parallel. Make user-facing strings translatable; report other strings that are passed to templates for manual review. Ignore dynamic strings, such as model fields etc.
32. Use a small workflow, 5 agents max …
33.   8. Claudese
34. THE WHEELHOUSE · #2 Fable Likes an Idea yegge.ai/comics/fable-likes-an-idea/
35. /config > Output Style Concise: Claude leads with the result, skips preamble and narration, and keeps responses short by default, while doing the engineering work as thoroughly as in the Default style.
36. ~/.claude/output styles/<name>.md --name: My style description: My custom style keep coding instructions: true --- . . - . - - - Lorem ipsum
37. Prompting Claude Fable 5.1 - Writing Density https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/ prompting-claude-fable-5-1#writing-density Mannered prose substitutes metaphor and flourish for direct statement. Instead of "a parameter worth varying," the mannered writer produces "a dial worth turning." Instead of "this point still matters," they write "this point earns its keep." The phrases exist to display the writer, not to convey the idea, and readers can tell. That is why mannered prose irritates: it makes the reader work harder so the writer can perform. It is also imprecise. Metaphors drag in connotations the writer did not choose and cannot control. The fix is to say what you mean. When a literal phrase is available, use it.
38. Attention Span pay attention, not tokens github.com/alexgreensh/attention-span
39. For a new social app, start with PostgreSQL, and the bar for choosing MongoDB instead is high. Social apps are relationship-heavy by nature: users follow users, posts belong to users, comments belong to posts, likes connect users to posts. That's a web of connections, exactly what a relational database is built for. In MongoDB you either cram related data into one record and ght size limits, or split it out and end up stitching it back together in your app code anyway. The feed problem makes this worse: timelines, "posts from people you follow," and mutual-friend lookups all lean on the kind of cross-referencing Postgres does in a single query. You also get consistency for free, since follow counts and "did I already like this" checks want real transactions… [and it keeps going like this] → Use PostgreSQL. A social app is all relationships (friends, follows, likes), and Postgres handles those natively. Right for ~90% of new social apps. fl Attention-kind · 94 words fi fi fi Default · 430 words → You keep exibility. Its JSONB column stores schema-less data inside the same database, so you get MongoDB's looseness without leaving Postgres. → Pick MongoDB only if your data is selfcontained documents with no links, and you need massive write volume on day one. → The early trap with MongoDB: six months in, ve different shapes of "user" and no clean x. Postgres forces the structure. → Next: start on managed Postgres (Supabase, Neon, RDS).
41. /wait-what github.com/mattpocock/skills
42.   Workflow
43.   9. /grill-me
44. github.com/mattpocock/skills
45. /grill-me I want to build a dashboard ...
46. Round 1. These are the decisions that do not depend on each other. ❓ Q1 - Rendering stack: Rich Live cannot read keys, and phase 4 needs Enter. Options: (a) Rich Live(screen=True) plus a small raw-mode stdin reader using termios/select; (b) Textual app with a canvas widget; (c) … ➡ (a) Rich Live plus a raw-mode stdin reader, … ❓ Q2 - Target terminal: Which terminal emulator and roughly what size will run on the projector laptop? …
47. /grill-me I want to build an admin dashboard ... Write what we agreed on to spec.md
48. /grill-me Refine @spec.md
49.   10. /simplify
50. /simplify /simplify working tree changes /simplify this branch /simplify this branch and create fixup commits
52.  Spouštět v nové session
53.   11. /code-review
54. /code-review working tree changes /code-review --fix /code-review --fix high /code-review --fix this branch and create fixup commits
55. 1. Implementace 2. /simplify 3. /code-review 4. Můj review
56. Implement @spec.md . Use workflow to: 1. Plan the work. 2. Run each task (or chunk of work) as a subagent. Commit as described in the plan. Prefer sequential workflow to avoid conflicts. 3. When done, run a step with /simplify on the entire branch, instruct it to create fixup commits so the final branch lands clean. 4. Autosquash the commits. 5. Run a step with /code-review --fix on the entire branch, instruct it to create fixup commits so the final branch lands clean. 6. Autosquash the commits.
57.  Spouštět v nové session
58.  12. Claude Design
62.  kód  Claude Design Claude Code
63.  13. Slow Loop
64. Každý den: • Vyber jeden úkol menšího rozsahu (PR na 100 - 500 řádek). refactoring, doplnění testů, ošetření edge cases, … • Implementuj ho. • Vytvoř pull request.
65. Kód každý den o kousek lepší
67.  Tuning
68.   14. PEP 758
69. try: . . . . . . # except ValueError, TypeError: logger.exception(" ") app.send_to_flames()
70. ### Python 3.14 syntax notes **Unparenthesized `except` with multiple exception types is valid** (PEP 758, Python 3.14+). `except ValueError, TypeError:` is equivalent to `except (ValueError, TypeError):` — it is **not** Python 2 syntax and will **not** raise a `SyntaxError`. Parentheses are still required when using `as` (e.g. `except (ValueError, TypeError) as e:`). Do not "fix" unparenthesized forms unless an `as` clause is being added.
71. # # 15. Komentáře
72. ### Comments and Docstrings Explain why the code is the way it is. Don't explain why you changed it. - **Why, not what.** A comment earns its place by naming a constraint the code can't show: why this order, why this cast, why not the obvious thing. Never paraphrase the next line. - **No history.** "used to", "previously", "legacy", "this replaces", "the old X" — write as if the current code is the only code that ever existed. Names included. - **Don't relitigate.** Justify the constraint you're living with, not the alternative you rejected. One sentence of "X would break Y" is a why; a paragraph is a commit message. - **Budget one or two sentences.** A multi-paragraph docstring on a helper is a smell.
73.   Bonus
74.   /insights
75. Insights Report • Shrnutí práce: na čem pracujete, jak CC používáte, co se daří, co se nedaří… • Co vyzkoušet: • • Co si přidat do CLAUDE.md • Nastavení hooks • Skilly Další možnosti automatizace úkolů i celých workflows
76.   Závěr
77.   Kvalita, ne objem
78.   Vylaďte si workflow
79. Díky za pozornost! www.linkedin.com/in/jansmitka @jansmitka @jansmitka@fosstodon.org https://lynt.cz/