CLAUDE CODE · BEST PRACTICES

Claude Code best practices
your own sessions call for.

Anthropic's best practices come down to four: tell Claude what it can't see, give it a check to run, enforce the rules that can't bend and change a default that costs too much. Zest's plugin reads your sessions and suggests which to try next.

Zest, a Claude Code plugin for engineers and teams, picks the best practice your sessions call for.

Start free Explore the team demo

Free for 3 seats · No credit card · The demo needs no signup

Free for 3 seats · No credit card · Setup: 4 commands in Claude Code

04Fix agentic loop bottlenecks

Winding / Agentic loop Sample week

Delivery loop

Team Zest

41 merged PRs

Code→Check→Review→Merge→Live

Pipeline stageMedian

CI wall time 12.4mRuns per PR 2.5Test / verify activity 18h / week

Approve → merge28m per PR

The largest delay comes after approval: the work is ready, but the merge is still waiting.

Give the agent a clear merge rule: merge when approval and all required checks are green.

About this sample

Sample from Zest’s pipeline report · Stage medians overlap and do not add up. Test / verify activity is a weekly session-flow estimate, not pipeline wall time.

USED BY ENGINEERING TEAMS AT

The four Claude Code best practices

Anthropic's guide starts from the defaults: a plain prompt in a fresh session usually does the right thing. Setup pays off where your project has knowledge, rules or constraints Claude can't see from the code. When a session goes wrong, it is almost always one of these four:

From Anthropic's best practices for Claude Code, read November 4, 2026.
What went wrong	The practice	Where to start
Claude built something else, or rediscovers the same fact every session	Tell Claude what it can't see	Name the file, the symptom and what fixed looks like. Keep `CLAUDE.md` to what the code can't show, under 200 lines
Claude said it was done and it wasn't	Give Claude a check to run	Anything that comes back pass or fail: a test file, a build's exit code, a linter. Put the command in `CLAUDE.md`
It did what must never happen, or skipped a step that must always happen	Enforce the rules that can't have exceptions	Move the rule out of `CLAUDE.md` into a hook, a deny rule or a managed setting
It got it right, but cost too much time, too many tokens or too many prompts	Change a default	Run `/usage` to see where it went, set the level with `/effort`, and `/clear` between unrelated tasks

Advice you can stop following

Older tips were workarounds for what earlier models couldn't do reliably. Anthropic now lists these as safe to drop:

Managing the context window aggressively: most current models have a 1M-token window and compact on their own. /clear between tasks is still worth it, for better answers and lower cost.
Adding IMPORTANT or all caps to a line Claude keeps missing: make a rule that can't bend a hook or a deny rule, and shorten the file if the line is getting lost.
Writing "think hard" or "be thorough" in the prompt: effort is the control for that.

Which one does your week call for?

A list can't see your sessions. Zest's Ship More tips can: Zest reads your past week of coding-agent sessions and merged pull requests, and picks up to three tips. A tip appears only when something in your own work triggers it, and each one ends with a line to try in your agent. On a day when nothing fires, Zest sends no tips.

Sample tips from Zest's Ship More catalogue, with the Try-it lines as written. A real tip names your own work.
What Zest saw	The tip	Try it
Agents kept waiting on or re-running a slow test suite	Faster test loop	`List the 10 slowest tests and say which to merge, mock or move out of the default run.`
Plans were rewritten before any working piece shipped	Plan one vertical slice	`Plan the smallest vertical slice of this I can test end to end, then build and ship only that slice.`
A layout problem turned up after the change was called done	Let the agent test it for real	`Run this change in a real environment, check it at desktop and mobile width, and put the screenshots in the pull request.`

Get tips from your own sessions

The tips arrive in your daily standup, which Zest writes at 6pm your time from that day's sessions and pull requests: in the app, and in Slack or by email when those are on.

Zest is AI coding analytics for Claude Code, and it arrives as a plugin. Run these in Claude Code; /zest:login creates your account:

Install the Zest plugin in Claude Code

/plugin marketplace add https://github.com/Winding-Labs/zest-claude
/plugin install zest
/reload-plugins
/zest:login

TRY IT ON YOUR TEAM

See what your team ships with AI.

Starter is free (3 seats, no credit card). Pro is $49 a month with 5 seats and Max is $149 a month with 10 seats — priced per workspace, not per engineer, with extra seats at $19 a month. See pricing

Start free Explore the team demo

A FEW THINGS TO KNOW

Questions

What are the most important Claude Code best practices?+

Anthropic's guide comes down to four: tell Claude what it can't see, give it a check to run, enforce the rules that can't have exceptions, and change a default when a task costs more than it should.

How soon will I see a tip?+

In your next daily standup after your sessions show one of a tip's triggers. Some days none fires, and that day has no tips.

Is this only for teams?+

No. Each engineer's tips come from their own sessions, so one person can use it alone. Starter is free for 3 seats, with no credit card.

Does Zest upload our source code?+

No code diffs are uploaded. The Claude Code plugin sends the text of the conversation, not tool inputs or outputs, after redacting secrets and personal data on the engineer's machine. Message text is kept for at most 90 days.