Claude Code Best Practices (Boris Cherny)

Verification and effort — moved

These two lived here and in CLAUDE.md, which is carried into every session; the same fact in two places is how the anti-pattern lists in this repository drifted apart. They now have one home, in the project's CLAUDE.md: give yourself the cheapest check that can fail, and match reasoning depth to the task (/effort). This file keeps what is reference rather than instruction — consulted on demand, not carried (specs/guardrail-cleanup/carried-material.md).

Recommended Model

"I use Opus with adaptive thinking for everything." -- Boris Cherny

(Since 2026-07-24, Opus 5 is the recommended default for complex work — near-Fable intelligence at half Fable's price.)

Context	Model	Rationale
Complex tasks (default)	Opus 5 (claude-opus-5)	Released 2026-07-24: within ~0.5% of Fable 5's peak scores at half the price ($5/$25 per MTok — same as Opus 4.8), 1M context, configurable effort. The opus alias resolves to it automatically
Extreme niche: the ~0.5% Fable still wins	Fable 5 (claude-fable-5)	~$10/$50 per MTok (2× Opus 5). Since Opus 5, a rare, deliberate choice — see the note below
Audits, analyses, high-volume agentic work	Sonnet (Sonnet 5)	1M context, $2/$10 per MTok — permanent since 2026-08-10 (the planned $3/$15 rise was cancelled) vs Opus $5/$25. Claude Code's default model for subscription seats since 2026-06-30
Simple tasks	Haiku	Fast for trivial operations

Opus 5 (since 2026-07-24): the new default of the opus tier alias (no agent-frontmatter change needed). Same price as Opus 4.8 with greatly improved performance: doubles Opus 4.8 on Frontier-Bench v0.1 and outperforms Fable 5 on OSWorld 2.0 at a third of the cost. Fast mode runs on it at $10/$50 per MTok (~2.5× speed). Opus 4.8 is not deprecated (it serves as fallback) but is no longer the recommendation. (Announcement)

Fable 5 after Opus 5: the "escalate to Fable for hard chantiers" advice is largely obsolete — Opus 5 lands within ~0.5% of Fable's peak at half the price, so Opus 5 absorbs the escalation role. Fable 5 remains relevant only where that last half-percent measurably matters (the hardest long-horizon autonomous runs) and there is still no fable tier alias: select it per-session via /model or --model claude-fable-5, deliberately. Availability note: Fable 5 has been generally available again since 2026-07-01 (the June export-control directive was lifted 2026-06-30; Mythos 5 stays restricted to a subset of US organizations). (Anthropic statement)

Sonnet 5 (since 2026-06-30): Claude Code's own default for subscription seats — near-Opus-4.8 quality on many agentic tasks at roughly a third of the cost, native 1M context, sonnet alias absorbs it automatically. This foundation recommends Opus 5 for complex/critical work (TDD, Audit, architecture) where the quality delta pays off, and Sonnet 5 for audits, analyses, and high-volume agentic passes where its price/perf shines.

Advanced Prompting

Avoid	Prefer
"Fix this bug"	"Fix the null pointer in getUserById when user doesn't exist"
"Make it better"	"Reduce the time complexity from O(n^2) to O(n log n)"
"Add error handling"	"Add try/catch for network errors with retry logic (3 attempts, exponential backoff)"

Techniques: "Grill me on these changes", "Prove to me this works", "Knowing everything you know now, implement the elegant solution".

See docs/guides/PROMPTING-GUIDE.md for the complete guide.

Automatic Memory (CLI 2.1.76+)

Claude Code automatically remembers preferences, decisions, and project context in ~/.claude/memory/.

Memorize (auto)	CLAUDE.md (git)	Rules (auto-activated)
Personal preferences	Project conventions	Per-language rules
Architecture decisions	Mandatory workflow	Code patterns
Team context	Documentation references	Verification checklist

Do not duplicate: if it is in CLAUDE.md, no need to memorize it. Use "remember that..." to force an explicit memorization.

Note (since Code with Claude 2026, May 6): Anthropic also ships Auto Dream / Dreaming, a managed memory feature where a background subagent reviews recent transcripts and consolidates the memory directory between sessions. It is complementary to the file-based system above: Auto Memory captures notes during work, Auto Dream cleans them between sessions. See the Claude managed agents blog post and the Dreams API doc.

Parallel Sessions

"The single biggest productivity unlock." -- Boris Cherny

Use git worktrees for 5+ Claude Code sessions in parallel. See the git-worktrees skill for details.

Context Management

Situation	Action	When
Long session, intact context	/compact	Between phases (Explore → Plan → TDD)
Total topic change	/clear	New unrelated task
Normal session	Let it be	Auto-compaction if needed

Quick Recovery

If a refactoring breaks everything: /rewind (or /undo, equivalent alias) returns to the last stable state. Faster than git stash or git checkout. Checkpoints saved automatically before each modification.

Since CLI 2.1.141, the Rewind menu also exposes a "Summarize up to here" entry that compresses earlier turns while keeping the recent ones intact — useful when the issue is context bloat rather than a broken change.

Session Resume

/recap generates a summary of the current session — decisions made, files modified, work state. Useful to resume a session after a break or a /compact.

Situation	Action
Return after a break	/recap to recover the context
After /compact	/recap to verify what was kept
Onboarding on an existing session	claude --resume <id> then /recap

Configurable via /config (enable/disable automatic recap on resume).

Token Optimization

Prompt Caching 1h (CLI 2.1.108+)

The ENABLE_PROMPT_CACHING_1H variable enables a 1-hour prompt cache instead of 5 minutes. Significantly reduces costs for long sessions.

Enable in .claude/settings.local.json:

{
  "env": {
    "ENABLE_PROMPT_CACHING_1H": "1"
  }
}

Compatible with API key, Bedrock, Vertex, and Foundry. Alternative: FORCE_PROMPT_CACHING_5M to force the 5-minute TTL (useful if telemetry is disabled).

Quick Command

/work:work-commit-push-pr "description" -- commit + push + PR in a single command.