How long should CLAUDE.md be?

Aim for under ~200 lines per CLAUDE.md file. Keep it focused on setup, build and test commands, and split larger or domain-specific instructions into .claude/rules/ so they lazy-load only when relevant.

How do I manage context in Claude Code?

Watch the "dumb zone" around ~40% context usage — results degrade as context fills. Use /compact with a focus hint or /clear when switching tasks, offload exploration to subagents so only conclusions return, and rewind to before a failed attempt instead of correcting in place.

Claude Code Best Practices — 80+ Expert Tips, Workflows & Concepts

60 of 152

Prompting

challenge Claude — "grill me on these changes and don't make a PR until I pass your test." or "prove to me this works" and have Claude diff between main and your branch 🚫👶

Source: Boris

x.com/bcherny/status/2017742752566632544

Prompting

after a mediocre fix — "knowing everything you know now, scrap this and implement the elegant solution" 🚫👶

Source: Boris

x.com/bcherny/status/2017742752566632544

Prompting

Claude fixes most bugs by itself — paste the bug, say "fix", don't micromanage how 🚫👶

Source: Boris

x.com/bcherny/status/2017742750473720121

Planning/Specs

always start with plan mode

Source: Boris

x.com/bcherny/status/2007179845336527000

Planning/Specs

start with a minimal spec or prompt and ask Claude to interview you using AskUserQuestion tool, then make a new session to execute the spec

Source: Thariq

x.com/trq212/status/2005315275026260309

Planning/Specs

always make a phase-wise gated plan, with each phase having multiple tests (unit, automation, integration)

Source: Dex

youtu.be/YwZR6tc7qYg?t=1032

Planning/Specs

break PRDs into vertical slices (tracer bullets) that cross all layers (DB + service + UI) — AI defaults to horizontal phasing (DB phase, then API phase, then frontend phase) which delays end-to-end feedback until the last phase. From the Pragmatic Programmer 🚫👶

Source: Matt

youtu.be/-QFHIoCo-Ko

Planning/Specs

spin up a second Claude to review your plan as a staff engineer, or use cross-model for review

Source: Boris

x.com/bcherny/status/2017742745365057733

Planning/Specs

write detailed specs and reduce ambiguity before handing work off — the more specific you are, the better the output

Source: Boris

x.com/bcherny/status/2017742752566632544

Planning/Specs

prototype > PRD — build 20-30 versions instead of writing specs, the cost of building is low so take many shots

Source: Boris

youtu.be/julbw1JuAz0?t=3630

Context

context rot kicks in around ~300-400k tokens on the 1M context model — don't let sessions drift past that for intelligence-sensitive work

Source: Thariq

github.com/shanraisshan/claude-code-best-practice

Context

dumb zone kicks in around ~40% context — "you hit this point where you have degrading results". Newcomers: "shoot to keep it under 40%, and if you get up to 60%, think about wrapping it up". Experienced: "aggressively keep it below 30%" — push to 60% only on simple tasks. Manual /compact or /clear to reset when switching tasks

Source: Dex

youtu.be/YwZR6tc7qYg?t=1541

Context

rewind > correct — double-Esc or /rewind back to before the failed attempt and re-prompt with what you learned, instead of leaving failed attempts + corrections polluting context 🚫👶

Source: Thariq

github.com/shanraisshan/claude-code-best-practice

Context

/compact with a hint (/compact focus on the auth refactor, drop the test debugging) beats letting autocompact fire — the model is at its least intelligent point when auto-compacting due to context rot

Source: Thariq

github.com/shanraisshan/claude-code-best-practice

Context

use subagents for context management — ask yourself "will I need this tool output again, or just the conclusion?" — 20 file reads + 12 greps + 3 dead ends stay in the child's context, only the final report returns 🚫👶

Source: Thariq

github.com/shanraisshan/claude-code-best-practice

Session Management

every turn is a branching point — after Claude ends a turn, pick between Continue, /rewind, /clear, /compact, or Subagent based on how much existing context you need to carry forward

Source: Thariq

github.com/shanraisshan/claude-code-best-practice

Session Management

new task = new session — related tasks (e.g. writing docs for what you just built) can reuse context for efficiency, but genuinely new tasks deserve a fresh session

Source: Thariq

github.com/shanraisshan/claude-code-best-practice

Session Management

use "summarize from here" before rewinding to have Claude write a handoff message — like a note to the previous iteration of Claude from its future self

Source: Thariq

github.com/shanraisshan/claude-code-best-practice

Session Management

/compact vs /clear — compact is lossy but momentum-friendly (mid-task, fuzzy details ok); /clear + brief is more work but you control exactly what carries forward (high-stakes next step)

Source: Thariq

github.com/shanraisshan/claude-code-best-practice

Session Management

use recaps for long-running sessions — short summaries of what Claude did and what's next, useful when returning after minutes or hours. Disable with recaps in /config

Source: Boris

github.com/shanraisshan/claude-code-best-practice

Session Management

/rename important sessions (e.g. [TODO - refactor task]) and /resume them later — label each instance when running multiple Claudes simultaneously

Source: Cat

every.to/podcast/how-to-use-claude-code-like-the-people-who-built-it

CLAUDE.md + .claude/rules

CLAUDE.md should target under 200 lines per file. 60 lines in humanlayer (still not 100% guaranteed)

Source: Boris

x.com/bcherny/status/2007179840848597422

CLAUDE.md + .claude/rules

.claude/rules/.md auto-load into every session like CLAUDE.md — add paths: YAML frontmatter to lazy-load them only when Claude touches files matching the glob

Source: Claude

code.claude.com/docs/en/memory#organize-rules-with-clauderules

CLAUDE.md + .claude/rules

wrap domain-specific CLAUDE.md rules in <important if="..."> tags to stop Claude from ignoring them as files grow longer

Source: Dex

www.hlyr.dev/blog/stop-claude-from-ignoring-your-claude-md

CLAUDE.md + .claude/rules

use multiple CLAUDE.md for monorepos — ancestor + descendant loading

Source: Boris

x.com/bcherny/status/2016339448863355206

CLAUDE.md + .claude/rules

use .claude/rules/ to split large instructions

Source: Claude

code.claude.com/docs/en/memory#organize-rules-with-clauderules

CLAUDE.md + .claude/rules

any developer should be able to launch Claude, say "run the tests" and it works on the first try — if it doesn't, your CLAUDE.md is missing essential setup/build/test commands

Source: Dex

x.com/dexhorthy/status/2034713765401551053

CLAUDE.md + .claude/rules

keep codebases clean and finish migrations — partially migrated frameworks confuse models that might pick the wrong pattern

Source: Boris

youtu.be/julbw1JuAz0?t=1112

CLAUDE.md + .claude/rules

use settings.json for harness-enforced behavior (attribution, permissions, model) — don't put "NEVER add Co-Authored-By" in CLAUDE.md when attribution.commit: "" is deterministic

Source: davila7

x.com/dani_avila7/status/2036182734310195550

Agents

have feature specific sub-agents (extra context) with skills (progressive disclosure) instead of general qa, backend engineer

Source: Boris

x.com/bcherny/status/2007179850139000872

Agents

say "use subagents" to throw more compute at a problem — offload tasks to keep your main context clean and focused 🚫👶

Source: Boris

x.com/bcherny/status/2017742755737555434

Agents

agent teams with tmux and git worktrees for parallel development

Source: Boris

x.com/bcherny/status/2025007393290272904

Agents

use test time compute — separate context windows make results better; one agent can cause bugs and another (same model) can find them

Source: Boris

x.com/bcherny/status/2031151689219321886

Commands

use commands for your workflows instead of sub-agents

Source: Boris

x.com/bcherny/status/2007179847949500714

Commands

use slash commands for every "inner loop" workflow you do many times a day — saves repeated prompting, commands live in .claude/commands/ and are checked into git

Source: Boris

x.com/bcherny/status/2007179847949500714

Commands

if you do something more than once a day, turn it into a skill or command — build /techdebt, context-dump, or analytics commands

Source: Boris

x.com/bcherny/status/2017742748984742078

Skills

use context: fork to run a skill in an isolated subagent — main context only sees the final result, not intermediate tool calls. The agent field lets you set the subagent type

Source: Lydia

x.com/lydiahallie/status/2033603164398883042

Skills

use skills in subfolders for monorepos

Source: Claude

code.claude.com/docs/en/skills

Skills

skills are folders, not files — use references/, scripts/, examples/ subdirectories for progressive disclosure

Source: Thariq

x.com/trq212/status/2033949937936085378

Skills

build a Gotchas section in every skill — highest-signal content, add Claude's failure points over time

Source: Thariq

x.com/trq212/status/2033949937936085378

Skills

skill description field is a trigger, not a summary — write it for the model ("when should I fire?")

Source: Thariq

x.com/trq212/status/2033949937936085378

Skills

don't state the obvious in skills — focus on what pushes Claude out of its default behavior 🚫👶

Source: Thariq

x.com/trq212/status/2033949937936085378

Skills

don't railroad Claude in skills — give goals and constraints, not prescriptive step-by-step instructions 🚫👶

Source: Thariq

x.com/trq212/status/2033949937936085378

Skills

include scripts and libraries in skills so Claude composes rather than reconstructs boilerplate

Source: Thariq

x.com/trq212/status/2033949937936085378

Skills

embed !command in SKILL.md to inject dynamic shell output into the prompt — Claude runs it on invocation and the model only sees the result

Source: Lydia

x.com/lydiahallie/status/2034337963820327017

Hooks

use on-demand hooks in skills — /careful blocks destructive commands, /freeze blocks edits outside a directory

Source: Thariq

x.com/trq212/status/2033949937936085378

Hooks

measure skill usage with a PreToolUse hook to find popular or undertriggering skills

Source: Thariq

x.com/trq212/status/2033949937936085378

Hooks

use a PostToolUse hook to auto-format code — Claude generates well-formatted code, the hook handles the last 10% to avoid CI failures

Source: Boris

x.com/bcherny/status/2007179852047335529

Hooks

route permission requests to Opus via a hook — let it scan for attacks and auto-approve safe ones 🚫👶

Source: Boris

x.com/bcherny/status/2017742755737555434

Hooks

use a Stop hook to nudge Claude to keep going or verify its work at the end of a turn

Source: Boris

x.com/bcherny/status/2021701059253874861

Workflows

use /model to select model and reasoning, /context to see context usage, /usage to check plan limits, /extra-usage to configure overflow billing, /config to configure settings — use Opus for plan mode and Sonnet for code to get the best of both

Source: Cat

x.com/_catwu/status/1955694117264261609

Workflows

always use thinking mode true (to see reasoning) and Output Style Explanatory (to see detailed output with ★ Insight boxes) in /config for better understanding of Claude's decisions

Source: Boris

x.com/bcherny/status/2007179838864666847

Workflows

use ultrathink keyword in prompts for high effort reasoning

Source: Claude

docs.anthropic.com/en/docs/build-with-claude/extended-thinking#tips-and-best-practices

Workflows

/focus mode hides all intermediate work and shows only the final result — trust the model to run the right commands and just look at the outcome (toggle with /focus)

Source: Boris

github.com/shanraisshan/claude-code-best-practice

Workflows

tune effort level with Opus 4.7's adaptive thinking — low for speed and fewer tokens, max for most intelligence (slider: low · medium · high · xhigh · max)

Source: Boris

github.com/shanraisshan/claude-code-best-practice

Workflows Advanced

use ASCII diagrams a lot to understand your architecture

Source: Boris

x.com/bcherny/status/2017742759218794768

Workflows Advanced

use /loop for local recurring monitoring (up to 7 days) · use /schedule for cloud-based recurring tasks that run even when your machine is off

Source: Boris

x.com/bcherny/status/2038454341884154269

Workflows Advanced

use Ralph Wiggum plugin for long-running autonomous tasks

Source: Boris

x.com/bcherny/status/2007179858435281082

Workflows Advanced

/permissions with wildcard syntax (Bash(npm run ), Edit(/docs/)) instead of dangerously-skip-permissions

Source: Boris

x.com/bcherny/status/2007179854077407667

Workflows Advanced

/sandbox to reduce permission prompts with file and network isolation — 84% reduction internally

Source: Boris

x.com/bcherny/status/2021700506465579443

What are Claude Code best practices?

Claude Code best practices are the workflows, habits and configuration patterns that get the most out of Claude Code — from how you write CLAUDE.md and manage context, to subagents, skills, hooks, plan mode and full Research → Plan → Execute → Review → Ship pipelines. This page distills 82 tips and 152 practices from the community's most-followed best-practice repository into one searchable, filterable index — sourced from the people who build and ship with Claude Code daily.

Quick wins

Always start in plan mode and write a phase-wise, testable plan before executing.
Keep CLAUDE.md under ~200 lines and split larger instructions into .claude/rules/.
Mind the "dumb zone" — aggressively /compact or /clear before context drifts past ~40%.
Use subagents for context management — let the conclusion return, not 20 file reads.
Rewind, don't correct — double-Esc back to before a failed attempt instead of polluting context.