13. Best practices: daily routine and anti-patterns

All previous chapters are about mechanics. This one is about discipline. Without it, even a perfect configuration turns into a mess over time: cache misses, skills go stale, hooks fail silently, and you don’t understand why your monthly bill doubled.

13.1. Daily routine for a developer with Claude Code#

This is not dogma, but a guideline. Everyone has their own rhythm. But if your day looks very different and you’re not happy with the results — try comparing.

Morning (10 minutes)#

📘 Concrete steps:

1. git pull --rebase — pull your colleagues’ changes.
2. Open Claude Code in the repo root. CLAUDE.md and MCP will be picked up automatically.
3. /cost — see how much was spent yesterday. If there’s a sharp spike — figure out why before starting new tasks.
4. /usage optionally — breakdown by models for the week.
5. If there’s a TodoWrite list from yesterday — go through it, cross out what’s outdated.
6. Formulate the first task of the day completely, not “let’s start with X”.

During work#

✅ One task — one session. After completing a task, do /clear or start a new session. Context shouldn’t accumulate “everything from the day”.

✅ Use plan mode for the unfamiliar. Before saying “implement”, say “plan”. That’s Shift+Tab or /model opusplan.

✅ Trust, but verify. Look at every PR with your own eyes, especially diffs in files Claude touched. Don’t become a rubber stamp.

✅ Keep a prompt diary. When a prompt works especially well or especially poorly — copy it to notes/prompts.md. In a month you’ll have a personal collection of working formulations.

End of day (5 minutes)#

📘 Steps:

1. /cost — day’s total.
2. If something broke and Claude didn’t fix it — write it in notes/stuck.md. Tomorrow with fresh eyes you’ll often see the solution.
3. Commit unfinished work to a WIP branch so you don’t start from scratch tomorrow.
4. Close Claude Code (this will end the subscription to MCP servers and free up resources).

13.2. Once a week: hygiene#

💡 Once a month: go through .claude/agents/ and ask yourself — do you still need each subagent. “It might be useful” is bad reasoning.

13.3. When CLAUDE.md “overlearned”#

Symptoms:

• Claude mostly answers “I already know” and ignores fresh hints.
• Strange code patterns: uses a library that was deleted long ago.
• Doesn’t use new skills.

📘 Reason — CLAUDE.md has grown to 5-10 KB, and the model can no longer manage it (it fits, but drowns in noise).

Treatment:

1. Open CLAUDE.md, read it top to bottom. How many lines could you delete right now because “it’s obvious for a modern model anyway”?
2. Move long specific rules to .claude/skills/<topic>/SKILL.md — they’ll be loaded only when needed.
3. Use @-imports for fragmentation.

⚠️ For a new employee, CLAUDE.md over 200 lines is an anti-signal “everything here is confusing”. Aim for 50-100 meaningful lines.

13.4. When the monthly bill shocks you#

📘 Main “wallet killers”:

1. A runaway teammate that got stuck in a loop. The TeammateIdle hook will help you notice this.
2. 1M window “by default”. Each new iteration scales payment linearly with input.
3. Cache misses. For example, if CLAUDE.md has {{date}} or other dynamics — every turn the cache is invalidated.
4. gh pr diff at the start of each session. PR grows — prefix grows. Don’t put it in the cacheable part.

13.5. Antipatterns: what NOT to do#

13.5.1. CLAUDE.md as “a dump of everything I know about the project”#

❌ 1500 lines of architecture description, decision history, all the “don’t do this”.

✅ 80-150 lines: stack, teams, important constraints, links to detailed docs.

13.5.2. Skills as “just in case”#

❌ 30 skills, each “might be useful”. The model spends tokens on their description with every request.

✅ 5-10 targeted skills. Deleted is better than unused.

13.5.3. Subagents for everything#

❌ Every code review through subagent. Every search through subagent. Every file through subagent.

✅ Subagent for browse-heavy tasks where summary matters more than transcript. Direct tool calls for everything else.

13.5.4. bypassPermissions as default mode#

❌ claude --bypass-permissions always. “It’s faster”.

✅ bypassPermissions only for specific safe tasks, worst case — for a separate worktree session.

13.5.5. Hooks without logging#

❌ Hook fails silently, and you only notice when the model starts doing weird things.

✅ Every hook logs its run (at least echo "[hook:format] $(date)" >> .claude/hooks.log).

13.5.6. One big prompt instead of iterative dialogue#

❌ “Add auth, implement RBAC, write tests, update docs, create PR” — in one line.

✅ Break it into 4 prompts. After each one, look at the diff. It’s not slower because errors in the first step don’t have time to bury the rest.

13.5.7. Ignoring /cost#

❌ You find out the bill amount at the end of the month.

✅ Check /cost at the end of each big task. It’s zero effort and builds a sense of proportion.

13.5.8. Skills with model: opus for routine#

❌ Skill format-pr-description with model: opus. Why? It’s a copying task.

✅ model: haiku or model: sonnet. Opus is for skills that really solve complex problems.

13.5.9. MCP servers with 100+ tools#

❌ Connected Notion-MCP with 80 endpoints, of which you use 3. All 80 go into the prefix of every request.

✅ Use tools in skill frontmatter or in settings.json permissions.allow so the model only sees what’s needed.

13.5.10. Never reading docs#

❌ “I read it on Twitter”. Anthropic documentation changes fast (new hook events, new env vars). Especially — release-notes.

✅ Every two weeks — /release-notes. Once a month — https://docs.claude.com/en/docs/claude-code/changelog.

13.6. Checklist before deploying an agent to production#

If your Travel Agent or other LLM project is going to production (where users touch it, not just you) — go through:

[ ] System prompt выделен в отдельный файл, версионируется
[ ] Tools описаны через JSON Schema с required-полями
[ ] Каждый tool возвращает компактный результат (нет 100KB JSON)
[ ] Все долгие tools имеют таймауты + чёткие сообщения об ошибке
[ ] Streaming реализован (UX) и есть graceful cancellation
[ ] Логирование: prompt, tool calls, tool results, final output, latency
[ ] Метрики: tokens_in, tokens_out, cache_read, cost, p50/p95 latency
[ ] Rate limit на пользователя (request/sec и tokens/day)
[ ] Защита от prompt injection (sanitize user input в system context)
[ ] Сценарии retry для transient API errors
[ ] Fallback на дешёвую модель при сбое основной
[ ] Audit log: что сделал агент от имени пользователя
[ ] Permissions: какие tools кому доступны (RBAC)
[ ] Тесты для типовых сценариев (vitest + recorded conversations)
[ ] Documentation: как пользоваться, какие limits, что делать если завис

13.7. Checklist for onboarding a new developer to a project#

When a newcomer joins a project where Claude Code is already built into the process:

[ ] Прочитать CLAUDE.md
[ ] Прочитать README.md
[ ] Запустить bootstrap-команду из 12.13
[ ] Открыть Claude в корне, написать "Покажи структуру проекта и ключевые скиллы"
[ ] Прочитать каждый SKILL.md в .claude/skills/
[ ] Запустить /agents и понять каких есть готовых subagent'ов
[ ] Запустить /mcp и убедиться, что все серверы зелёные
[ ] Сделать первую задачу-разминку из onboarding-doc'а
[ ] Прочитать 13.5 (антипаттерны) и сразу осмотреться, нет ли их в проекте

13.8. Security: rules without compromise#

🔒 Rules that can never be broken, even “temporarily for an experiment”:

1. Never give Bash(*) without a deny-list. Minimum: rm -rf *, curl * | sh, wget * | sh, dd if=*.
2. Never give Edit(.env*). That’s a secrets leak.
3. Never run bypassPermissions in a repo with production credentials in .env.
4. Never push branches with --force through Claude. Hook to block it:

   if echo "$INPUT" | grep -qE 'git push.*--force'; then
     echo '{"action": "block", "reason": "Force push needs human approval"}'
     exit 0
   fi

5. Never give an MCP server with file-write permissions access to the system root. Only the project folder.
6. Never trust prompt injection from user input in system context. Sanitize or escape before feeding to LLM.

⚠️ Special case: MCP servers with side-effects on payments (Stripe-MCP, wallets). By default — ask, no exceptions.

13.9. When Claude Code is not needed#

Sometimes it’s worth closing Claude Code and working by hand. Signs:

• Task is one line change in one file you know by heart.
• You’re trying to learn something and need to feel the material.
• System operations (chmod, chown, system updates) — there’s no copy cost and there’s full cost of error.
• Emotionally charged moment (panic after prod) — Claude doesn’t calm you down, better to take a deep breath first.
• You don’t understand the task yourself. If you can’t formulate it — you won’t formulate it for Claude either.

💡 “Using Claude Code” != “using it always”. Best use is when you consciously turn on the tool, not on autopilot.

13.10. Practice: your personal “knowledge capsule”#

📘 Gradually collect your own patterns in ~/.claude/CLAUDE.md (user-level). In 3-6 months you’ll have:

• List of favorite prompt formulations.
• 3-5 universal skills (for typical PRs, for refactoring, for testing).
• Your personal code style preferences (semicolons, quotes, naming).
• Your “never do this” — lessons from your own mistakes.

It transfers between all projects. This is your personal “leveling up” of the tool.

Next → 14. Verification of claims from the original thread