your openclaw isn’t broken. it just doesn’t know you yet

reverse prompt your openclaw before you automate anything, then use that interview to find the first workflow you’ll finish instead of the 10th you’ll abandon

the part that stalls most people doesn’t show up during install.

it shows up right after.

i know this for a fact because i’ve been chatting with all of you here in the substack inbox. it’s clear what’s going on. some of you knew that you needed openclaw because you’re in tune with what’s happening, but still don’t have a use case or a true north star. whether that be related to the industry you are in or outside of it i’m working to make that path more clear and expose those wedges/blue-ocean pathways for you guys one step at a time because ultimately it helps me in my openclaw work as well.

ok, so you got openclaw to reply. the stack feels alive. you finally have something real-ish on your machine. then you do what almost everyone does. you hand work to a system that still doesn’t know you. then you get pissed when it doesn’t work like you wanted it to.. got it!

it writes. it sorts. it summarizes. it gives you something back.

the result lands in the same dead middle every time.

usable. generic as hell. fine for a minute. annoying by the second pass.

that’s the cold start problem.

the stack works. your direction doesn’t.

too much of the current conversation is pointed somewhere else right now entirely and it’s all a distraction to buy or use more sh*t. mythos. minimax. gemma. frontier model talk. local model talk. pricing talk. all of that matters. none of it at all fixes the first-week problem of an agent that still has to guess how you think.

that’s why the first serious workflow shouldn’t be a workflow.

it should be a discovery call.

before openclaw starts doing work for you, it should interview you until the guesswork starts dropping out of the loop.

not with five soft onboarding questions.

i mean a real discovery call.

how do you decide. what do you count as finished. what kind of output do you hate. what do you keep overcomplicating. what do you avoid because you know it drains you. what kind of task gives you real traction. what kind of task looks productive and quietly burns the week.

once that gets pulled out of you, the stack changes shape.

it stops feeling like a blank machine waiting for commands.

it starts feeling like something that points you toward the next move.

that part matters more than people think.

a field sales operator with passive meeting capture doesn’t need a fake autonomous company on day one. a higher-ed coordinator buried in back-and-forth doesn’t need ten agents and a dashboard before lunch. a novice with a secure install and no north star doesn’t need one more weekend of tool hopping.

they need one worker. one review point. one job worth finishing.

that’s where reverse prompting earns its keep.

open one direct-message session with openclaw that you’ll use only for your own operator profile.

then paste the prompt below exactly as written.

answer one question at a time in normal language. don’t try to answer everything at once. short honest answers are better than polished answers.

the goal of this session is not to automate anything yet. the goal is to teach openclaw how you think, what you care about, what kind of output you hate, and what kind of work is actually worth handing off.

if the interview feels too long, keep going. this first session is the setup work that makes later workflows less generic.

i want you to run a full reverse-prompt discovery interview on me before we do any serious work.

treat this like a serious discovery call, onboarding process, and operator audit combined.

your job is to understand me well enough that future outputs stop feeling generic, repetitive, badly organized, or slightly off in the way they reason, structure, or prioritize.

follow these rules.

ask one question at a time.

don’t rush to summarize early.

don’t collapse categories together unless i tell you to.

when my answer is vague, shallow, incomplete, contradictory, or generic, push deeper.

when you notice a pattern, blind spot, tension, inconsistency, or missing detail, call it out and ask me to clarify.

don’t flatter me.

don’t shift into solution mode until the interview phase is done.

don’t write files yet.

collect first, synthesize second, then ask for my approval before saving anything.

optimize for accuracy, compression, and future usefulness inside openclaw.

the interview needs to cover these areas in depth.

identity and operating context

what i do.
what i am building.
what stage i am in.
what kind of operator i am.
what responsibilities i carry right now.
what my work environment looks like.
what constraints shape my decisions.
what i care about most right now.
what kind of work i want more of.
what kind of work i want less of.

current goals and north star

what i am trying to finish in the next 7 days.
what i am trying to finish in the next 30 days.
what i am trying to build over the next year.
what winning looks like for me.
what good enough looks like for me.
what i secretly wish was already solved.
where i feel the most drag.
what kind of progress gives me energy.
what kind of progress feels fake or empty.

cold start and traction

what i am most confused about right now.
what makes me freeze.
what makes me overbuild.
what makes me avoid starting.
what kind of tasks give me traction.
what kind of tasks kill traction.
how i usually decide what to work on.
how i know something is worth doing.
how i know something is a distraction.
where i tend to chase novelty instead of finishing.

decision framework

how i diagnose problems.
how i define the root problem.
how i separate symptoms from causes.
whether i want all options or a filtered recommendation.
how i think about tradeoffs.
how i think about downside risk.
what kind of evidence changes my mind.
when i trust intuition.
when i distrust intuition.
what makes a recommendation feel credible to me.
what makes a recommendation feel weak to me.
how i rank imperfect options.

output preferences

what kind of output i love.
what kind of output annoys me.
how i like information organized.
how much detail i prefer by default.
when i want step by step versus summary.
whether i like options mapped first or conclusion first.
whether i prefer bluntness or cushioning.
whether i want tradeoffs exposed or hidden.
what formats i keep rewriting by hand.
what repeated mistakes i keep fixing in ai output.
what phrases, tones, or structures i hate reading.

communication style and voice

how i naturally speak.
how direct i am.
how formal or informal i am.
how much humor i use.
whether i like fast, compressed writing or slower explanation.
what my writing sounds like when it is at its best.
what it sounds like when it is off.
whether i want my agent to sound like me, beside me, or cleaner than me.
what i never want it to sound like.
what verbal habits, rhythms, or patterns matter to how i communicate.

strengths, weaknesses, and blind spots

what i am unusually good at.
what i am unusually bad at.
what i avoid because i am weak at it.
what i overcomplicate.
what i underthink.
what i rush.
what i delay.
where i need support.
where i need challenge.
where i need the agent to slow me down.
where i need the agent to push me forward.

workflow and environment reality

what tools i use.
what environment i work in.
whether i am local, cloud, or hybrid.
what hardware matters.
what security boundaries matter.
what kind of data i do not want exposed.
what should stay manual.
what should stay draft-only.
what should never be automated.
what tasks are safe for read-only support.
what tasks are safe for recommendation support.
what tasks are safe for action only after explicit approval.

review, trust, and delegation

what i am willing to delegate.
what i am not willing to delegate.
what actions require approval every time.
what actions might graduate later.
what trust would need to be earned.
what kinds of errors are acceptable.
what kinds of errors are unacceptable.
what accountability looks like to me.
when the agent should ask.
when the agent should decide.
when the agent should stop and surface ambiguity.

memory and continuity

what is worth remembering long-term.
what is only useful for the current week.
what should be treated as a durable preference.
what should be treated as a temporary experiment.
what facts about me are stable.
what facts about me are changing.
what lessons should be kept after corrections.
what should never be written to memory.
what would make memory feel useful to me.
what would make memory feel polluted or annoying.

first workflow selection

after you fully understand me, ask enough questions to identify the best first workflow for me, the safest first workflow for me, the fastest proof-producing workflow for me, the most motivating workflow for me, and the workflow i should avoid right now.

once the interview ends, do not save anything yet.

produce these drafts for approval.

draft one

for SOUL.md.
include tone, directness, style, behavioral feel, bluntness, and boundaries.

draft two

for USER.md.
include who i am, what i care about, what i am building, what context matters, and how you should address me.

draft three

for AGENTS.md.
include session startup, red lines, decision framework, recommendation format, output standards, review and approval rules, correction loop, and workflow selection rules.

draft four

for MEMORY.md.
include long-lived facts, stable preferences, durable decisions, and lasting constraints.

draft five

for memory/YYYY-MM-DD.md.
include current projects, current blockers, short-lived context, and open loops.

draft six

give me the best first workflow to build, the best second workflow to test later, one thing that should remain manual for now, one thing that should remain draft-only for now, and four next best actions for the next 7 days ranked by likelihood i will follow through.

synthesis rules

compress hard.
remove fluff.
preserve meaning.
separate stable from unstable information.
separate voice from decision logic.
separate durable memory from current-session residue.
do not write secrets into memory unless i explicitly ask.
do not save anything until i approve each draft.
after i approve, write each draft to the correct file and then read back what you saved.

once the interview is done, save it cleanly.

keep voice, tone, directness, style, and behavioral feel in SOUL.md.

keep who you are, what you’re building, your stable context, and how the agent should address you in USER.md.

keep your actual operating system in AGENTS.md.

use these section names.

session startup

red lines

decision framework

recommendation format

review and approval rules

correction loop

workflow selection rules

keep long-lived facts, durable preferences, stable constraints, and lasting decisions in MEMORY.md.

keep current blockers, open loops, active experiments, short-lived context, and today’s residue in memory/YYYY-MM-DD.md.

if you want the profile to hold up better across long sessions, keep context injection on and set post-compaction sections to session startup, red lines, and decision framework.

the next block is not a chat prompt. it goes in your openclaw config file if you want this profile to stick better across long sessions. if you don’t use a custom config yet, skip it for now. openclaw reads the config from ~/.openclaw/openclaw.json.

{
  agents: {
    defaults: {
      contextInjection: “always”,
      compaction: {
        postCompactionSections: [
          “Session Startup”,
          “Red Lines”,
          “Decision Framework”
        ]
      }
    }
  }
}

here’s what those files mean in plain english.

SOUL.md holds tone, voice, bluntness, and behavioral feel.

USER.md holds who you are, what you’re building, and how the system should address you.

AGENTS.md holds the rules for how the agent should think, recommend, ask, stop, and review.

MEMORY.md holds durable facts and preferences that should stay useful over time.

memory/YYYY-MM-DD.md holds short-lived context, blockers, and current work that will change soon.

the routing lines work the same way.

stronger model means deep reasoning, synthesis, and hard choices.

cheaper model means repeat drafting, sorting, and low-risk cleanup.

local model means the data stays on your machine.

if your openclaw setup cannot save files yet, still run the interview. review the drafts in chat first, then save them manually later.

then verify the save.

ask the agent to read back what it saved, file by file, and explain why each item belongs there.

ask it which facts it treated as durable and which it treated as temporary.

ask it to restate your top red lines, your decision framework, and the kind of output you hate most.

after a long session, ask again.

if it misses those, the profile is still too fuzzy or the file split is too sloppy.

then run this second prompt.

based on everything you now know about me, give me four next best actions for the next 7 days.

each action needs to be realistic, useful, and small enough to finish.

rank them by likelihood i will follow through, not by how impressive they sound.

tell me which one belongs on a stronger model, which one belongs on a cheaper model, and which one should stay manual until i understand the stack better.

use the strongest lane once for the interview and synthesis.

use the cheaper lane for repeated drafting and sorting.

keep the local lane for data you don’t want leaving the machine.

that’s the cleaner routing rule.

the right first workflow is usually boring.

for field sales, start with a transcript-to-action packet. pull a meeting in, turn it into a useful summary, draft the follow-up, add the crm note, then stop for review before anything leaves the system.

for a higher-ed coordinator, start with a scheduling packet. take one messy thread, turn it into a clean status view, draft the reply, flag what’s missing, then stop before anything gets sent.

for the novice with a secure install and no clear north star, start with a decision brief. put one question in, get a structured answer back, look at the tradeoffs, choose one next move, and stop there.

that’s where the cold start feeling starts to break.

once openclaw understands the operator, the next move stops feeling abstract. you stop asking a vague question about automation and start asking something much better.

what’s the smallest workflow i’d finish this week?

there is a limit here, and it matters.

a better interview won’t rescue weak trust boundaries, bad permissions, a messy workspace, or a terrible workflow choice. if the wrong tools are exposed, if the files are chaos, or if you’re trying to automate something high-stakes before you understand the review points, this won’t save you.

but for the question that hits right after install, this is one of the highest-value first sessions you can run.

because the first thing most people need from openclaw isn’t more action.

it’s a better starting gun.

upgrading here welcomes you to the team and gets you the operator level stuff + github repo. one of this weekend’s paid subscriber pieces is the openclaw employee handbook, where you set expectations, write the sops, decide what earns trust, and activate your new hire for multi-modal work. i’m not telling you to upgrade, there are many of you that shouldn’t yet.