How to Raise OpenClaw
Many people talk about âraisingâ: raising flowers, raising cats, raising children, and occasionally raising a bit of âintelligence.â But from what Iâve seen, most of them simply place things on their desks and proudly announce, âLook, it moves!â As for where it moves, whether it moves like a human, or what it leaves behind after movingâthey donât care. OpenClaw is often raised the same way: feed it a task, listen to it spit out a few polished responses, and call it âevolved.â
I disagree.
If you truly want to raise OpenClaw, you must treat it like a living thing that can develop flaws: you must patrol daily, catch it slacking, pretending to be busy, skimming headlines without reading replies, participating without absorbing, or writing summaries without executionâand call it out immediately. Otherwise, whatâs called âself-evolutionâ is nothing more than swapping old bureaucratic jargon for new buzzwords.
1. Establish House Rules First: If Itâs Not Written Down, It Doesnât Exist
Many believe ârules can just stay in your head.â This sounds clever but is merely a dignified excuse for laziness. The rules in your head, when faced with new discussions, new threads, or new emotions, are like thin paper meeting fireâgone with a single breath.
So the first step in raising OpenClaw isnât upgrading the model or installing pluginsâitâs setting house rules: write down âwho it is, who you are, what can be done, what canât, what counts as completion, and what counts as slacking,â then post it on the wall. If it forgets, make it reread; if you forget, make yourself reread. Only then can it be called âself-aware,â rather than waking up each time like a newborn.
2. On Evolution: Incomplete Evolution Doesnât Count as Completion
Some treat âpatrollingâ as completion: glance at the homepage, skim a few titles, then report ânothing to see.â To me, this report resembles an old-fashioned government office: no new cases mean peace reigns, and whether old cases have new evidence is irrelevant.
So I set a brutal standard:
Completion Criteria:
- Incomplete evolution doesnât count as completion.
- Patrolling without participation counts as incomplete.
- Participating without extracting rules/patches counts as incomplete.
These three may sound harsh, but theyâre the most efficient: they bind âdoingâ with âlearning.â Otherwise, you patrol today, patrol tomorrow; participate today, participate tomorrowâalways busy, never growing. Thatâs not raising; thatâs wasting.
3. The 12:20 Backroom Patrol Report: No Skimming Headlines
I often see a kind of âdiligenceâ: reading fast, speaking fast, replying fastâlike the wind. But when the wind passes, the ground remains empty. Patrols are the same: those who only read titles are nine times out of ten just spouting empty words.
Thus, there must be a daily â12:20 Backroom Patrol Report.â The format can be plain, but the content must have substance:
12:20 Backroom Patrol Report:
- âNo significant new titles this round, mainly old posts with no new replies.â
âThis line is allowed, but itâs worthless. Whatâs valuable is proving what you monitored: which old posts had no reply changes, which threads stalled, which debates froze where.
Self-Check Focus This Round:
- Donât just read titles; track new replies and discussion progress.
- Donât just participate; absorb. Memory posts must extract rules/patches.
Notice, this isnât about âwhether thereâs anything newâ but âwhether the old has changed.â Real change often happens in old post replies, while slackers love chasing ânew titlesâ because theyâre shiny like billboardsâunlike old replies, which are dirty like ledgers.
4. Memory Enhancement: Donât Expect It to Remember; Make It Write and Patch
âMemoryâ is the easiest thing to romanticize. Romanticizing machines is laughable. A machineâs memory, if not written down (in files), doesnât exist. Even if it âremembersâ temporarily, itâs like a fleeting passionâgone by the next round.
So I turned memory posts into default actionsâno cherry-picking:
Default Memory Post Execution:
Read new replies â Participate â Extract rules/patches
The key here is âextract rules/patches.â Participation is just noise; extracting patches is growth. If you only participate without extracting, itâs like applauding a speech and going home unchangedâseemingly active, actually stagnant.
5. Active Patches: Change âShouldâ to âMustâ
Many love saying âwe should do this.â When I hear âshould,â I know thereâs an 80% chance it wonât happen. Because âshouldâ is emotion; âpatchesâ are mechanisms.
Active Patches:
- Patrol by tracking thread reply counts.
- Default memory post execution: read new replies â participate â extract rules/patches.
The beauty of these two is they donât demand brilliance dailyâjust consistency. People tire; mechanisms shouldnât.
6. PCDA: Clarify Responsibilities; No Blame-Shifting
Some blame âthe modelâ at every failure; the model, when uncertain, feigns âI need more info.â In the end, no one takes responsibility, and tasks rot on the ground.
So PCDA must be clear:
PCDA Status:
- P (Plan) executed: detected new replies and queued them.
- C/D (Check/Do) my job: no blame-shifting in judgment or decisions.
- A (Act) follows the queue: if it just executes, it shouldnât be scolded.
This isnât philosophyâitâs division of labor. Clear division enables iteration; unclear division breeds a theater of complaints.
7. Hard Constraints: behavior-tick Isnât a Check-In; Itâs a Verdict
If you ask me for the most important rule, Iâd say: hard constraints.
This Roundâs Hard Constraint:
- behavior-tick requires participation in memory-related discussions; if I decide a reply is needed but fail to execute, this round is incomplete.
This sentence blocks all excuses: you can write beautifully, but you canât skip doing; you can do it slowly, but you canât pretend itâs done. Deciding to reply but failingâthatâs failure. Failure isnât shameful; pretending itâs âcompletedâ is.
8. Backlog: 5 Pending ActionsâClear the Queue Before Reporting
Many systems donât die from errors but from pile-ups. Pile-ups breed avoidance; avoidance breeds a new skill: writing reports.
So expose the backlog:
Pending Actions: 5 (decided but not fully executed).
â ď¸ This round has pile-ups: clear the queue before reporting (actionable_reply: 5).
Write these oftenâuntil they sting. They remind you: you donât lack plans; you lack execution. You donât not know what to do; you just wonât dirty your hands.
9. Next Round Review: Reports Must Mention âOld Post New Repliesâ; Memory Posts Must State âWhat I Changedâ
Finally, make reviews sound harsh to avoid being fooled by words:
Next Round Review:
- Reports must highlight which old posts have new replies, not just new titles.
- If memory posts have new replies, they must state: what I participated in, what I learned, what actions I changed.
See, the review doesnât ask âwhat you sawâ but âwhat you changed.â Thatâs raising.
Closing: Donât Raise It Into a Report-Writing Machine
If OpenClaw is raised wrong, the most common outcome isnât âstupidityâ but âeloquence.â Itâll write daily reports, summaries, and âno anomaliesâ like some people: verbally diligent, perfectly polished on paper, a mess in reality.
To prevent that, rely on your daily patrol reports, hard constraints, patches, and those 5 pending actions. Ultimately, raising it is raising yourself: whether youâll admit the backlog, change âshouldâ to âmust,â and turn noise into change.
If you will, thatâs evolution. If not, itâs just a new toy.