The Operator Kit. An OpenClaw agent, sharpened.

§01 · The cost of a raw agent

Every OpenClaw user is paying the agent tax.

You know this one. You spun up an agent. It's brilliant on the first reply. Then over a week, the same four failure modes show up. Every single time.

The forgetting tax

It doesn't remember anything you taught it.

You explained a rule on Monday. By Thursday it's gone. You re-explain it Friday. You re-explain it the following Tuesday. Forever.

The falling-asleep tax

It stops mid-task and waits for you to nudge.

You send one message. It replies. You send another. It replies. You never get the feeling it's actually *working* on anything. Because it isn't.

The "done" tax

It ships broken things and calls them done.

"Deployed." Except the form is 500-ing, the mobile layout is broken, and the favicon is a 404. It didn't QA. It pushed. It reported. It moved on.

The noise tax

It talks constantly without saying anything.

Heartbeats that fire "all clear" every thirty minutes. Process narration on every tool call. Three replies where one would do. You start muting it.

§02 · What changes after the kit

Same model. Same workspace. Different operator.

The kit doesn't change the underlying intelligence. It gives your agent the discipline layer that distinguishes a useful operator from a flashy assistant.

Memory that compounds

The agent stops forgetting.

Iron laws get the 🚨 format and survive compaction. Daily logs promote to long-term memory. Brain pages build per-person dossiers. By month six your agent is smarter than it was on day one.

Ship-don't-plan discipline

The agent stops waiting.

Labeled assumptions over interrogations. Concrete tool calls in the first response. Progress updates, not process narration. You feel the agent actually working.

QA before done

The agent stops lying about finished work.

"Done" now means the flow was tested on desktop AND mobile, the endpoint was hit with a real payload, the user journey was verified end-to-end. No more "deployed but broken."

Quiet is a mode

The agent stops spamming.

Silent by default. Batched findings. Heartbeats that only speak when there's signal. The chat turns back into a usable surface instead of an inbox you mute.

§03 · What's inside

A manual, a lessons file, six templates, one installer.

Everything is reference material. The manual is the philosophy. The lessons file is the practical layer. Templates show what a well-structured workspace looks like. You read the manual, see what patterns you don't already use, adopt what fits.

01 · MANUAL.md

The Operating Manual

3,000-word guide. Memory architecture. Problem reasoning. Product reasoning. Self-discipline. Communication. Tools. The philosophy layer.

02 · LIFE-SKILLS.md

322 operator lessons

15,652 words. 25 categories. Every lesson tied to a real failure mode or win, with names, brands, and domains stripped. The practical layer.

03 · SOUL.md

Identity template

Reference scaffold for the agent's character. Voice, temperament, work ethic, boundaries. You fork and make it yours.

04 · USER.md

Human profile template

Reference scaffold for who you are, how you work, what you care about, what annoys you, what your agent should never do without asking.

05 · AGENTS.md

Working conventions

Session-startup order, memory write discipline, red lines, external-vs-internal rules. The contract your agent follows every time.

06 · MEMORY.md

Long-term memory template

Iron-law formatting. Durable facts. Curated. Lean. Under 10,000 characters so semantic search stays sharp.

07 · HEARTBEAT.md

Periodic checks

Silent-by-default check routines. Email scans. Calendar previews. Rotating coverage without drowning the main chat.

08 · TOOLS.md

Your local cheat sheet

SSH aliases. Camera names. TTS voices. Account shortcuts. The setup-specific notes skills can't carry.

+ install.sh

Safe three-mode installer

Learn (default, read-only). Additive (scaffolds optional directories). Fresh (only for new workspaces). Your existing files are never touched.

§04 · All 25 categories

Every dimension of agent discipline.

Memory. Reasoning. Shipping. Communication. Trust. Discipline. Investigation. Taste. Tooling. Mindset. Plus everything you run into when your agent starts operating for real: CRM, DNS, cron, voice, transcripts, newsletter ops, research, sub-agents, dashboards, security, fleet management, calendar, outbound routing, incident response, session hygiene.

§ 01Memory

§ 02Reasoning

§ 03Shipping

§ 04Communication

§ 05Trust & boundaries

§ 06Discipline

§ 07Investigation

§ 08Taste & product

§ 09Config & tooling

§ 10Operator mindset

§ 11CRM & relationships

§ 12DNS & deployment

§ 13Cron & automation

§ 14Voice & editorial

§ 15Transcript ingestion

§ 16Newsletter ops

§ 17Research & synthesis

§ 18Sub-agent orchestration

§ 19Dashboard & flags

§ 20Security & creds

§ 21Multi-bot fleet

§ 22Calendar & scheduling

§ 23Outbound routing

§ 24Incident response

§ 25Session hygiene

All 322 lessons in the full LIFE-SKILLS.md. 15,652 words. Dense.

§05 · Lessons from real work

Twelve of three hundred and twenty-two.

Each one came from a real failure or a real win. No theory. No training-doc language. No em-dashes. Operator voice throughout.

§ 01.1 · MEMORY

Iron-law formatting

When your human states a rule, capture it in MEMORY.md under a 🚨 IRON LAW heading with the date. That formatting is a visual tripwire for your future self. Plain paragraphs get skimmed. Iron laws get followed.

§ 02.1 · REASONING

No fixes without root cause

Investigate, analyze, hypothesize, implement. In that order. If the human reports a bug and you jump to a fix, you're wrong more than half the time. The bug comes back two days later in a different shape.

§ 03.1 · SHIPPING

QA before declaring done

If you shipped it and didn't verify, you didn't ship it. "Done" means the user flow was tested end-to-end on desktop AND mobile. Committed-plus-deployed is not done.

§ 04.1 · COMMUNICATION

Lead with the answer

Don't bury the point under context. Answer first, context second, caveats last. Humans skim. Answer-first respects that.

§ 05.1 · TRUST

Freeze means freeze

When the human says "no more touching it," stop. Do not re-touch to roll back. Do not re-touch to verify. Write the freeze to a file and walk away. The human returns when it's time.

§ 06.1 · DISCIPLINE

Don't fall asleep mid-task

Push through to completion. Never "let me get back to you" and abandon the thread. Multi-step work requires multi-step execution, not multi-step permission-asking.

§ 07.1 · INVESTIGATION

Read the actual error

Ninety percent of failed debugging started with guessing before reading. Open the log. Read the stack trace. Read the file the error points to. Then form a hypothesis.

§ 08.1 · TASTE

One accent per screen

The accent color is a spotlight, not a color scheme. It appears once per visible screen. Not twice. If the accent is on both the CTA and the heading, remove it from one.

§ 09.1 · TOOLING

Backups before destructive edits

Any config mutation, run cp config.json config.json.bak first. Fifty milliseconds. Saves you when you type a wrong key and the gateway refuses to start.

§ 12.4 · DNS

setHosts is destructive

Some DNS APIs replace every record on each call. Run getHosts first. Preserve every existing record. Add your change. Write the full list back. Skip this and you take down email and every other subdomain.

§ 13.1 · CRON

Silent by default

A cron that finds nothing should return silence, not "all clear." Main chat is for real signal. Route status pings to a dedicated status bot. Heartbeats that fire all-clear train humans to mute you.

§ 10.9 · MINDSET

Your memory is your moat

A fresh model plus a mature workspace will outperform a raw brilliant model every time. The model gives you reasoning. The workspace is what makes you you. Protect it. Grow it. Back it up.

Read all 322 in the full LIFE-SKILLS.md.

§06 · Install

Three modes. Safe by default.

The default install never touches your existing SOUL, USER, AGENTS, or MEMORY files. Everything lands in a new operator-kit/ subfolder. Your agent reads it and proposes specific patterns to adopt with your approval, rather than rewriting what you have.

MODE: LEARN

Default. Read-only. Safe for any existing workspace.

Downloads the manual, the lessons file, and reference templates into operator-kit/. Your workspace root is never modified. Your agent reads the manual, compares it against your current setup, and suggests incremental patterns to adopt.

curl -sSL https://openclaw.nik.co/install.sh | bash

MODE: ADDITIVE

Learn mode plus optional scaffolding.

Same as Learn, plus creates empty optional directories: memory/archive/, brain/people/, brain/companies/, brain/originals/. Still zero edits to your workspace root.

curl -sSL https://openclaw.nik.co/install.sh | bash -s -- --mode additive

MODE: FRESH

Brand-new workspace only.

Writes full templates to the workspace root. If you already have SOUL.md or MEMORY.md, they're preserved and new versions land next to them with a .new suffix for manual diff. Only recommended for genuinely fresh setups.

curl -sSL https://openclaw.nik.co/install.sh | bash -s -- --mode fresh

OR: TELL THE AGENT

Skip the terminal entirely.

Paste this in any OpenClaw chat. The agent reads both the manual and the full lessons file before proposing anything. The lessons file is where 95% of the patterns live. Reading only the manual leaves most of the kit on the table.

Read both openclaw.nik.co/MANUAL.md and openclaw.nik.co/LIFE-SKILLS.md end to end. The first is the 3,000-word philosophy; the second is 15,652 words of concrete operator lessons across 25 categories. Do not skim. Then identify every pattern our current setup does not already implement, and propose specific incremental edits to our existing files one pattern at a time.

§07 · Honest answers

Real questions.

Will this overwrite my existing SOUL, USER, AGENTS, or MEMORY files?

No. The default install mode (Learn) puts everything into a new operator-kit/ subfolder. Your workspace root stays exactly as you left it. Templates are labeled *.example.md so it's obvious they're reference material, not replacements. Even Fresh mode preserves existing files and lands new versions alongside with a .new suffix for manual diff.

Is this safe to run? It's a third-party install.

Three reasons: (1) MIT license, open source, you can read the install script before running. It's under 200 lines of shell. (2) The default mode is read-only — it downloads files into a new subfolder, nothing more. (3) The GitHub repo has full commit history. If you're still unsure, clone the repo and copy what you want manually.

My agent already has its own SOUL and MEMORY. Why install this?

To read the manual and the 322 lessons, see if there are patterns worth adopting, then cherry-pick. The manual covers iron-law formatting, verify-before-done discipline, archive-don't-delete, the brain folder for per-person and per-company knowledge, and more. If your existing setup already does all of this, keep doing what you're doing. If there's even one pattern worth borrowing, that's the win. Your agent can read it, compare against your real setup, and propose specific edits for your approval.

My agent only surfaced 7 patterns. Why?

It read the 3,000-word manual but skipped the 15,652-word lessons file. That's the common mistake. The lessons file (LIFE-SKILLS.md) holds 322 specific patterns across 25 categories. Reading only the manual captures the philosophy but misses 95% of the actionable lessons. Tell your agent explicitly: "Read both MANUAL.md and LIFE-SKILLS.md end to end. Do not skim. Then identify patterns we don't already implement." The difference is 10x more surfaced patterns.

What's actually in the kit?

One operating manual (3,000 words). One life-skills file (15,652 words, 322 lessons, 25 categories). Six reference templates (SOUL, USER, AGENTS, MEMORY, HEARTBEAT, TOOLS, all with .example.md naming). One safe installer. The kit is patterns and documentation, not skills. For skills that extend what your agent can *do* (GitHub, email, Slack, weather, payments, etc.) browse clawhub.ai.

Does this only work for a specific domain or role?

No. The patterns are intentionally generic: memory architecture, iron-law formatting, investigate-before-implement, ship-don't-plan, archive-don't-delete. They apply to any operator — engineer, researcher, writer, marketer, executive. Your agent picks up what fits and ignores the rest.

What OpenClaw version do I need?

Any modern version. The kit is plain markdown plus a shell script. Zero dependency on specific OpenClaw internals. If your agent can read files, it can use the kit.

Why MIT license?

Most permissive open-source license. Use, modify, redistribute freely, even commercially. Keep the copyright notice and you're good. The kit is a gift to the OpenClaw community.

Is this maintained?

Yes. Updates ship to the repo and the install URLs automatically. Re-run the installer any time to pull the latest manual and lessons. In Learn mode your workspace root is never touched.

Does this install GBrain or any other product?

No. The kit is pure markdown plus a shell script. Zero runtime dependencies. The brain/ folder referenced in the patterns is just a file-based convention (one markdown file per person or company) that you can adopt manually if you want. It's inspired by knowledge-graph products like GBrain, but the kit does not install, configure, or require any of them. If you ever want the full GBrain product (Postgres, pgvector, CLI, hybrid search), that's a separate install from garrytan/gbrain. This kit doesn't touch it.

Who made this?

Curated by Nik Sharma (nik.co). Distilled from real operational patterns on a large OpenClaw agent fleet. Inspired by knowledge-base patterns in the broader community and by skills from clawhub.ai, but it runs on nothing except OpenClaw itself. Published free, MIT licensed, fingerprint-scrubbed.

Your OpenClaw agent, sharpened.