A first workflow
Anthropic’s best-practices page, as of November 2026, comes down to four points: tell Claude what it cannot see, give it a check to run, enforce the rules that cannot have exceptions, and change a default only when a task costs too much. As a workflow for a real task:
- Write a specific prompt. Name the file, the symptom and what “fixed” means. Not “fix the login bug”, but “users report login fails after the session times out. Look at token refresh in src/auth/, write a failing test that reproduces it, then fix it.” Type
@and a path to include a file, paste a screenshot, or pipe output in withcat error.log | claude -p "what's going on here". - If you do not know exactly what you want yet, let Claude interview you. “Interview me until you could write the spec, then write it to SPEC.md.” Then start a fresh session to build from that file, with a clean context and a written definition of done.
- Give Claude a check. A test file, a build, a linter, a script that compares output, a screenshot next to the mockup. In the prompt: “run the auth tests after, and keep going until they pass.”
- Ask for evidence. The command it ran and the output. “All tests pass” on its own tells you little.
- Get a second pair of eyes.
/code-reviewreviews the diff in a fresh subagent that never saw your conversation. When you want a check against your plan, ask a subagent to compare the diff with SPEC.md and to skip style. More on the options in Claude Code review. - Commit, then
/clearbefore the next unrelated task.
If you want the fuller version of a feature, say “be ambitious”: Claude tends to build the smallest thing that satisfies a request, because it cannot see how much you hoped for.
Steer in the first minute
Three controls are worth knowing by heart:
| Control | What it does |
|---|---|
Esc |
Stops Claude mid-action and keeps the work so far, so you can redirect |
/rewind (or Esc twice on an empty prompt) |
Rolls code, conversation or both back to an earlier prompt |
/clear |
Starts over; after two failed corrections, a fresh prompt that includes what you learned works better than a third correction |
/btw answers a quick side question without adding it to the history. /rename oauth-migration names a session so claude --resume oauth-migration picks it up days later.
Where each instruction belongs
The page splits instructions in two. A prompt, CLAUDE.md or a skill is read by the model, which usually follows it but decides itself. A hook, a deny rule or a managed setting is applied by Claude Code every time.
CLAUDE.md: short, only what the repo does not show. Test and typecheck commands, code that misleads, decisions made outside the repo, things you had to correct twice./initdrafts one; then the work is mostly deleting lines. How it loads is on Claude Code memory..claude/rules/with apaths:filter for rules that only matter in one part of the codebase.- Skills for team procedures such as a release. They only load fully when the task needs them. See Claude Code skills.
- Hooks and deny rules for what must always or never happen: never read
.env, always run the formatter. See Claude Code hooks.
A sign from the docs: if you keep adding emphasis to a CLAUDE.md line and Claude still misses it, that line belongs in a hook.
The misunderstanding: the old advice still applies
A lot of Claude Code advice online was written for older models. Anthropic’s page now has a table of things you can stop doing:
- Managing the context window aggressively. Most current models have a 1M-token window and compact on their own.
/clearbetween tasks is still worth it, for focus and cost. - IMPORTANT or ALL CAPS. Current models follow a plain line; emphasis makes Claude over-apply the rule.
- “Think hard” in the prompt. Effort is the control, through
/effort. - Describing your architecture in
CLAUDE.md. Claude reads the repo itself. - Delegating to a subagent for a better answer. Same model, same task. A fresh reviewer helps; plain delegation does not.
What this means for a team
Most of what makes Claude Code work well is information only people have: decisions, procedures, the check that defines done. A short committed CLAUDE.md, shared skills and hooks in .claude/settings.json carry that to every developer. What they do not carry is what one developer’s session learned today, or what was decided in a thread outside the repo, until someone writes it down. For unattended runs, the same points apply more strictly: a definition of done, a check, limits on what it may touch, a cap on turns or spend, and a person who approves the result.
Where Bombos fits
The best-practices page keeps returning to one gap: Claude cannot see what the team knows outside the code. Bombos is a plugin that gives Claude Code and the team’s other AI tools one shared memory with that knowledge, the rules the team learned, and what already lives in mail, documents and repos, each answer traceable to its source. One developer can try it on their own agents first.
Related questions
Do I need to set anything up before my first prompt?
No. Anthropic's best-practices page says a plain first prompt in a fresh session usually does the right thing: sessions start in auto mode, Claude delegates to subagents on its own and compacts long conversations automatically. Setup pays off once the project has knowledge Claude cannot read from the code.
Should I write think hard or be thorough in my prompt?
It no longer raises how hard Claude thinks. Effort is the control: /effort high for tricky work, /effort low for quick back-and-forth. ultrathink in a single prompt asks for deeper reasoning on that one turn.
When should I use /clear and when /compact?
/clear between unrelated tasks, so the next task does not pay for the previous one's history. /compact with a focus, such as /compact keep the API changes, when you want to keep going but trim. /compact is itself a large request on a long session.
How do I undo what Claude just did?
/rewind, or Esc twice on an empty prompt, rolls back the code, the conversation or both to an earlier prompt. It covers files Claude edited with its own file tools, not changes made by Bash commands, so keep using git.
What should go in CLAUDE.md?
Only what Claude cannot find by reading the repo: the exact test and lint commands, misleading spots in the code, team decisions, and anything you corrected twice. /init writes a first version; the memory guide recommends staying under 200 lines.
