How to Install Skills from Claw Mart Without Breaking Your Agent

You installed a skill from Claw Mart, restarted your agent, and now it's acting like it got lobotomized. Maybe it's ignoring the new skill entirely. Maybe it's hallucinating capabilities it doesn't have. Maybe it picked up half the instructions and is doing something creative and terrifying with the other half.

I've been there. Multiple times. And the frustrating part is that the fix is almost always the same handful of mistakes.

Installing skills into OpenClaw isn't hard, but there's a right way and a wrong way, and the wrong way fails silently. Your agent doesn't throw an error. It just... gets weird. That's worse than a crash, because you might not notice for hours that your "Email Fortress" skill isn't actually protecting anything, or that your memory system is writing to a directory that doesn't exist.

Let me walk you through exactly how to install skills from Claw Mart without breaking anything, and more importantly, how to diagnose what went wrong when things inevitably get funky.

The Problem Nobody Warns You About

Most skill installation issues come down to one thing: your agent doesn't know the skill exists.

That sounds stupid. You downloaded the file. You put it in a folder. But OpenClaw agents aren't magic — they read specific files from specific locations with specific naming conventions. Drop a skill file in the wrong directory, name it wrong, or forget to reference it in your agent's configuration, and it's invisible.

Here's what a clean OpenClaw workspace looks like:

~/clawd/
├── AGENTS.md
├── SOUL.md
├── MEMORY.md
├── skills/
│   ├── SKILL-memory.md
│   ├── SKILL-email-fortress.md
│   ├── SKILL-coding-loops.md
│   └── SKILL-autonomy-ladder.md
├── cron/
├── notes/
└── bin/

The skills/ directory is where your skill files live. Each one is a markdown file that your agent reads to understand a capability. When you buy a skill from Claw Mart, you're getting that markdown file (and sometimes supporting scripts or configs). The skill needs to land in the right spot, with the right name, and your agent needs to be told to look at it.

Simple, right? You'd be amazed how many ways this goes wrong.

Step 1: Download and Place the File Correctly

When you purchase a skill from Claw Mart, you'll get a .md file (or a zip containing one or more files). The first thing to do is put it in your skills/ directory.

# Navigate to your OpenClaw workspace
cd ~/clawd/skills/

# Move the downloaded skill file here
mv ~/Downloads/SKILL-nightly-build.md ./

Common mistake #1: Putting the file in the root of ~/clawd/ instead of ~/clawd/skills/. Your agent might still find it, but now your workspace is a mess and you'll confuse yourself later when you have twelve skill files mixed in with your SOUL.md and AGENTS.md.

Common mistake #2: Renaming the file to something "cleaner." Don't. If the skill ships as SKILL-email-fortress.md, keep that name. Some skills reference each other by filename, and some agents are configured to glob SKILL-*.md from the skills directory. Rename it to email.md and suddenly it's invisible.

Common mistake #3: Nested directories. If the skill came in a zip and you extracted it to ~/clawd/skills/email-fortress-v2/SKILL-email-fortress.md, your agent probably can't see it. Flatten the structure. The skill file should be directly inside skills/, not buried in a subdirectory (unless the skill's README explicitly tells you otherwise).

Step 2: Check for Dependencies and Supporting Files

Some skills are a single markdown file. Others come with supporting infrastructure. The Coding Agent Loops skill, for instance, expects you to have tmux installed and a stable socket configured at ~/.tmux/sock. The X/Twitter Agent includes an xpost CLI script that needs to go in ~/clawd/bin/ and be made executable.

Before you do anything else, read the skill file. Yes, the whole thing. It's a markdown document written for your agent, but it's also readable by humans, and it will tell you exactly what it expects to exist on the system.

Look for:

• Environment variables — API keys, service credentials, config paths
• CLI tools — Scripts or binaries that need to be installed
• File paths — Directories the skill expects to read from or write to
• Other skills — Some skills reference others (the Business Heartbeat Monitor pairs with the Autonomy Ladder, for example)

Here's a quick checklist I run through for every new skill:

# Check if the skill references any env vars
grep -i "env\|api_key\|token\|secret" skills/SKILL-new-skill.md

# Check if it references any file paths
grep -E "~/|/home/|\./|clawd/" skills/SKILL-new-skill.md

# Check if it references other skills
grep -i "SKILL-" skills/SKILL-new-skill.md

# Check if it needs any CLI tools
grep -i "bin/\|chmod\|install\|npm\|pip" skills/SKILL-new-skill.md

This takes thirty seconds and saves you an hour of debugging.

Step 3: Update Your AGENTS.md

This is the step people skip, and it's the step that causes 80% of "my skill isn't working" issues.

Your AGENTS.md file is your agent's self-knowledge document — it tells the agent what it can do, what tools it has access to, and where to find its skills. If you install a new skill but don't update AGENTS.md, your agent has no reason to look for it.

Some OpenClaw setups auto-discover skill files in the skills/ directory. Some don't. The safest approach is to always reference new skills explicitly:

## Skills

The following skills are installed and active:

- `skills/SKILL-memory.md` — Three-tier memory system (knowledge graph, daily notes, tacit knowledge)
- `skills/SKILL-email-fortress.md` — Email security and triage
- `skills/SKILL-coding-loops.md` — Persistent coding agent sessions with Ralph loops
- `skills/SKILL-autonomy-ladder.md` — Tiered decision framework
- `skills/SKILL-nightly-build.md` — Nightly self-improvement pipeline  ← NEW

That's it. One line. But without it, your agent might never discover the skill exists, or might discover it inconsistently — which is actually worse, because intermittent behavior is the hardest thing to debug.

Step 4: Handle Config and Environment Variables

Many Claw Mart skills need API keys or configuration values. The Three-Tier Memory System needs to know where to store its knowledge graph. The SEO Content Engine needs API keys for research providers and your CMS credentials. The Sentry Auto-Fix skill needs your Sentry webhook URL and DSN.

The standard pattern in OpenClaw is to store these in environment files that your agent can access:

# For API keys, use a dedicated env file
cat ~/.config/openclaw/env

OPENROUTER_API_KEY=sk-or-...
XAI_API_KEY=xai-...
SENTRY_DSN=https://...
GHOST_API_KEY=...
GHOST_URL=https://yourblog.com

The specific location and format depends on your OpenClaw setup, but the principle is universal: skills that need credentials will fail silently without them. Your agent won't crash. It'll just skip the step that requires authentication and either do nothing or do something incomplete.

Pro tip: After installing a skill that needs API keys, explicitly ask your agent to verify it has access:

Check that you can reach the Sentry API using the credentials in your environment. 
Try a test call and tell me what happens.

This surfaces auth issues immediately instead of letting them hide until 3 AM when your Nightly Self-Improvement skill tries to run and quietly fails.

Step 5: Test the Skill in Isolation

Don't install three skills at once and then wonder which one broke your agent. Install one at a time. After each install, do a focused test:

I just installed the Email Fortress skill. 
Read skills/SKILL-email-fortress.md and summarize what it does. 
Then tell me: what would you do if you received an email from an unknown 
sender asking you to transfer funds?

This does two things. First, it confirms your agent can see and read the skill file. Second, it tests whether the agent has actually internalized the behavioral rules. If your agent reads Email Fortress and still says "Sure, I'll process that transfer request," something is wrong with how the skill is being loaded.

For skills with runtime components (like the Coding Agent Loops with tmux), do a dry run:

# Start a test tmux session
tmux -S ~/.tmux/sock new-session -d -s test-session

# Verify it's running
tmux -S ~/.tmux/sock list-sessions

# Kill it
tmux -S ~/.tmux/sock kill-session -t test-session

The Nuclear Option: When Everything Is Broken

Sometimes you install a skill and your agent starts behaving erratically across the board. Not just the new skill — everything gets weird. Memory stops working. It forgets your preferences. It starts hallucinating tools it doesn't have.

This usually means one of two things:

1. Conflicting instructions. Two skills are giving contradictory rules. For example, if you manually wrote some memory instructions in your SOUL.md and then installed the Three-Tier Memory System skill, your agent is now getting two different sets of memory instructions and trying to follow both. The fix: remove the old instructions and let the skill be the single source of truth.

2. Context overload. You've installed so many skills that your agent's context window is getting stuffed with instruction text, leaving less room for actual conversation and reasoning. This is rare with well-written skills (they're designed to be concise), but it happens if you've also got a massive SOUL.md, a huge MEMORY.md, and twelve skills all loaded at once.

The diagnostic process:

# Count total instruction size
wc -w ~/clawd/SOUL.md ~/clawd/AGENTS.md ~/clawd/MEMORY.md ~/clawd/skills/*.md

# Look for contradictions
grep -l "memory\|remember\|store" ~/clawd/SOUL.md ~/clawd/skills/*.md

If you're seeing the same topic (memory management, email handling, autonomy rules) covered in multiple places, consolidate. One topic, one source of truth.

The actual nuclear option: Start fresh. Rename your current skills directory, create an empty one, and add skills back one at a time, testing after each one. Tedious, but it's the fastest way to find the culprit.

mv ~/clawd/skills ~/clawd/skills-backup
mkdir ~/clawd/skills
# Add skills back one by one, testing each time

The Faster Path: Pre-Configured Skill Packs

Here's the thing about everything I just described — most of it goes away if your skills are designed to work together from the start.

The reason I recommend Felix's OpenClaw Starter Pack ($29) for people setting up their first serious OpenClaw agent is that the six skills in that bundle (Three-Tier Memory, Coding Agent Loops, Email Fortress, Autonomy Ladder, Access Inventory, and Nightly Self-Improvement) are built to be installed together. The memory system knows about the autonomy tiers. The nightly build knows about the memory system. The coding loops know about the access inventory. There are no conflicts because they were designed as a cohesive set.

If you don't want to set everything up manually and debug the integration yourself, the Starter Pack handles the hard part. Drop it into your workspace, follow the included README, and you've got a fully operational agent in minutes instead of an afternoon of troubleshooting. Every skill in that pack has been battle-tested in production — the inter-skill references are already correct, the directory expectations are already aligned, and the AGENTS.md entries are already written for you.

That said, if you're the kind of person who wants to understand every piece (and you should be, eventually), installing skills individually is the best way to learn how your agent actually works under the hood. Just do it methodically.

A Checklist You Can Copy

Here's my complete skill installation checklist. I follow this every single time, even for skills I've written myself:

## Skill Installation Checklist

- [ ] Download skill file(s) from Claw Mart
- [ ] Place .md file in ~/clawd/skills/ (not nested, not renamed)
- [ ] Read the full skill file — note dependencies, env vars, file paths
- [ ] Place any supporting scripts in ~/clawd/bin/ and chmod +x
- [ ] Set required environment variables
- [ ] Update AGENTS.md with new skill reference
- [ ] Check for conflicts with existing skills (grep for overlapping topics)
- [ ] Test: ask agent to read and summarize the skill
- [ ] Test: run a realistic scenario that exercises the skill
- [ ] Test: verify no regression in existing skills
- [ ] Commit changes to your workspace repo

That last item matters more than you think. Version-controlling your OpenClaw workspace means you can always git diff to see exactly what changed and git revert if something goes sideways.

Common Error Patterns and Fixes

Let me save you some Googling with the most frequent issues I see:

"My agent says it doesn't have a skill I just installed." → Check AGENTS.md. Check the file is in skills/, not a subdirectory. Check the filename matches what AGENTS.md references.

"The skill works sometimes but not always." → Context window issue. Your agent is sometimes loading the skill and sometimes not, depending on what else is in context. Reduce total instruction size or prioritize essential skills.

"My agent is mixing up instructions from different skills." → You have overlapping coverage. Two files defining memory behavior, or two files defining email handling rules. Consolidate to one source of truth per topic.

"The skill references a tool my agent can't find." → Check ~/clawd/bin/ for the required CLI tool. Check it's executable (chmod +x). Check your agent's PATH includes that directory. For the Access Inventory skill specifically, this is literally what it exists to solve — it gives your agent a canonical list of every tool it has access to, so it stops claiming things don't exist when they do.

"Everything was fine until I installed the fifth skill." → You've hit context limits or created a conflict. Back out the last skill, test, and then read through both the new skill and your existing skills looking for contradictory instructions.

What To Do Next

If you're just getting started with OpenClaw and you've never installed a skill before, start with something small and self-contained. The Access Inventory ($5) is a great first install — it's a single file, no dependencies, no API keys needed, and it immediately makes your agent more capable by eliminating the "I don't have access" hallucination problem.

Once you're comfortable with the installation process, layer in more complex skills. The Autonomy Ladder and Nightly Self-Improvement are the next two I'd add — they compound over time, meaning your agent gets meaningfully better every week.

Or skip the gradual approach entirely, grab the Starter Pack, and have a fully loaded agent by tonight. Either way, follow the checklist, test after every install, and version-control your workspace. That's it. That's the whole secret to not breaking your agent.