Tips & Tricks · v1.2.10

12 tips to make Claw work better for you

Give enough context, it's more reliable than you think.
These aren't tricks — they're the right way to use Claw.

01 Context

Be specific · Let Claw plan · Break tasks down

02 Verification

Set standards · Define audience · Correct directly

03 Memory

Use @ for files · Write rules in config · Organize context

04 Efficiency

Check docs first · /powerup · Main Claw thinks, sub-Claw executes

These tips come from real usage experience.
Each one has been tested and proven — not theory, but lessons.
Perfect for those who've installed Claw and want to use it better.

Context determines Claw's ceiling

Claw doesn't truly "understand" — it relies on the context you give. The more you give, the better it aligns with your direction.

Be Specific

Vague instructions = vague results. When describing a task, include your background, goals, and tech stack. Claw can't read minds.

✗ Help me fix this code

✓ Fix the null pointer in the login module, using Python, error appears on line 5

Source: Claude Code / CNBlog

Let Claw Plan First

For complex tasks, let Claw list the steps first before acting. When direction is right, execution matters; when direction is wrong, effort is wasted.

✗ How should I build this feature? Let's just do it

✓ How should I build this feature? First list 3 steps, I'll confirm the direction, then you can proceed

Source: Claude Code founder Boris

Break Tasks into Chunks

Break large requirements into small steps, confirm each step, each step is verifiable and rollback-able. Throwing too much at once leads to uncontrollable quality.

✗ Refactor the entire backend for me

✓ First tell me how many parts this divides into, we'll start with the auth module and go through it piece by piece

Source: Cursor lead designer Ryo Lu

Clear verification standards = less rework

Claw doesn't know your standards. Specify verification criteria, and it knows what to change and what to keep. Less back-and-forth, less rework.

Specify Verification Criteria

Tell Claw "what to focus on" instead of "check this for me". With standards, Claw can give the right direction.

✗ Check this code for me

✓ Check this code, focus on memory leaks and error handling, explain what you changed after completing

Source: Claude Code community

Define the Audience

"Write a technical proposal" ≠ "Write a report for non-technical leadership". Different audiences, completely different depth and tone from Claw.

✗ Write an email for me

✓ Write an email, recipient is Finance department, explain why server costs increased 30%, tone should be objective, no finger-pointing

Source: AI prompt articles

Correct Directly

"You're wrong" is more effective than explaining. Correct once, Claw remembers once. Use your corrections as training data.

✗ Re-describe the requirement, hoping Claw understands this time

✓ You're wrong, I said X, not Y. Remember: in future when you encounter this, handle it the Y way

Source: Cursor lead designer

Use files correctly, use memory well, avoid detours

Context accumulates in files, not in conversations. Remember once, save effort next time.

Use @ to Reference Files

Much more precise than "check if the XX config in the project has problems", saves tokens, saves time, Claw won't go off-track.

✗ Check if the XX config in the project has problems

✓ Check @ this config file (around line 23) for problems

Source: Claude Code docs

Write Rules to Config File

For rules corrected once, tell Claw "remember this, handle it this way in the future". No need to repeat in next conversation, more rules accumulate, the better it understands you.

✗ Always add "by the way, I use Y version, don't use X version methods"

✓ Write this rule in: when version selection is involved, prioritize LTS versions

Source: Claude Code CLAUDE.md

Regularly Organize Context

When conversation gets too long, say "summarize current progress". Claw solidifies key state, frees up space, new task response quality stays high.

✗ Wait for Claw to get stuck, then figure it out

✓ Let's pause, summarize what we've completed so far, make a list, I'll add any misses

Source: Claude Code /compact

Use it right, double the efficiency

Claw has many built-in features. Knowing and not knowing makes it two different tools.

T10

Check Docs First When Problems Occur

When encountering errors, check tech docs §5 first, often there are ready answers. Those records are from real experience, more efficient than describing to Claw.

✗ When encountering errors, check docs first, if not found tell Claw: docs section X says... but I tried and it didn't work

✓ Check docs first → Not found → Then ask Claw, with "docs section X says... but I tried and it didn't work"

Source: Landing page self-value

T11

Try /powerup

10 levels take you through core features quickly. Learn step by step, like Duolingo for languages.

✗ Fumble around not knowing where to start

✓ Type /powerup, go through level by level

Source: Claude Code official

T12

Main Claw thinks, sub-Claw executes

Gongyan Claw focuses on execution, main Claw focuses on thinking. Two things progress simultaneously, each doing its own thing, no waiting.

✗ Wait for Claw to finish A, then let it do B, serial waiting

✓ Gongyan Claw, go run this task — main Claw continue with next thing

Source: Claude Code founder Boris (multi-session parallel)

Ready to Get Started?

Download Claw Soul and configure your personal lobster in 3 minutes.

Getting Started Download Now Back to Home