Launching today
Cleo
The AI PM that runs your team
367 followers
The AI PM that runs your team
367 followers
Cleo is the AI product manager for founders and lean teams. It lives in Telegram and Slack - learns your tone, knows your team, and runs the PM work (standups, follow-ups, decisions) while you ship the product. What's different: every fact Cleo learns is transparent - you see the source, the confidence, and can confirm or correct it. No black-box memory. Five trust levels, from observer to operator. Free in Telegram. 1 min setup
Cleo
Cleo
@rahimwws Happy to launch with you mate
Sleek Pay
@rahimwws I'm cool to use it in Telegram / Slack, but is it available in Discord or Whatsapp?
Cleo
@daniel_baum soon, If you want to follow where this is going you can join to waitlist in our website trycleo.ai
Product Hunt
Cleo
@curiouskitty Three layers:
Detection - Cleo watches for linguistic markers ("let's go with", "decided", "we'll ship X instead of Y") and structural ones (a question raised + an answer accepted in the same thread). Both signals together = high confidence it's a decision. One alone = surfaced for confirmation, not auto-captured.
The "why" - every decision card stores the trade-off considered (what we picked, what we didn't, what was the constraint). Cleo pulls this from the same thread, not from a separate prompt. If the trade-off isn't explicit in the conversation, the card flags it as missing.
Staying fresh - decisions have an owner and a domain (architecture, hiring, pricing). When a contradicting decision appears later, Cleo links them and asks if the older one is superseded. That's how the log stays clean: not auto-archive, not silent overwrite - explicit linking.
Big congrats on the launch! The transparency model (source + confidence + fix button on every fact) is the right instinct, but I'd push further: there's a failure mode where an auditable AI quietly becomes a verification queue. If I'm confirming and correcting facts to keep Cleo accurate, I've just swapped standup overhead for memory-maintenance overhead. What's the design that keeps the audit asymmetric, where I only look when something's wrong, not to keep it right? Does Cleo surface only low-confidence facts for review, act silently on high-confidence ones, and shrink the queue as it learns my corrections?
Cleo
@ferdi_sigona Sharp catch. This is the failure mode we feared most and explicitly designed against. Three things keep the audit asymmetric:
1. Confidence determines visibility, not just labeling. High-confidence facts (>85%) act silently — Cleo uses them, doesn't ping you. Mid-confidence (60-85%) gets surfaced only when about to be used in a consequential action. Low-confidence (<60%) goes to a review queue. You're not approving facts — you're approving the small subset where Cleo isn't sure yet.
2. Confirmations compound, corrections decay. A fact you confirm once moves up the confidence ladder. A fact that survives 30 days without contradiction or correction graduates to "settled" — invisible in normal flow. You converge to a state where most of memory operates without you.
3. Corrections become rules, not patches. When you fix a fact, Cleo extracts the pattern ("Mark doesn't own ingestion anymore, Sarah does — and ownership transitions happen in #eng-leads channel"). The next 100 facts of that type don't need correcting.
The math we watch: confirmation-events per active user per week. If that number doesn't drop month over month, the design is failing. Right now it drops ~40% from week 2 to week 6 for active users. That's the signal we'd kill the product over if it inverted.
Nice launch! Quick q - is Cleo built specifically for agencies, or would it work for a small SaaS team too? Most of my "PM" pain is internal standups and follow-ups, not client stuff.
Cleo
@ahmetaz Yes - works for small SaaS teams. Agencies are the wedge because the tool stack is consistent and the pain is acute, but the operational primitives (memory, learned rules, integrations) generalize.
what is the difference between Cleo and handing over a few docs for guidelines on any ai agent? genuinely interested.
also, quick feedback on your comment, when you use AI to write it, remove the em dashes and make it personal! good luck!
Cleo
@math_biomes thanks!
Cleo
@math_biomes
1. They're a snapshot. Mark owned ingestion latency last month. Sarah owns it now. You'd have to remember to update the doc every time something changes, and you won't. Cleo updates itself from the actual signals (a Slack message saying "Sarah is taking over") and flags both for confirmation.
2. They don't carry source. Doc says "we decided to ship X". Three months later you forget why. Cleo stores the decision with the conversation it came from, the trade-off considered, who was in the room.
3. They don't act. A doc is read-only context. Cleo reads the doc, then drafts the message, runs the standup, catches the follow-up. The doc is the prompt. Cleo is the worker.
the confidence scoring model is the right approach. most AI tools just act on everything with the same certainty and you only find out it was wrong after it already sent the message. curious how the 30 day correction window works in practice when the team structure changes fast
Cleo
@tina_chhabra Static decay breaks for fast-moving teams — by the time a fact ages out, the org chart has changed twice. So decay is event-driven, not time-driven.
Facts have a type that determines how they age. Structural facts (ownership, roles) decay fast on contradiction -one Slack message saying "Sarah's taking over from Mark" drops the old fact's confidence below threshold and queues both for confirmation.
30 days is the fallback for facts with no contradicting signal. For active teams, most facts get refreshed or contradicted long before that.
Congrats on the launch! How is it different from OpenClaw with some PM skills? You mentioned OpenClaw doesn't remember what you said yesterday, but from my experience it remembers quite well.
Cleo
@greatll Fair point - Openclaw does have memory now, and it works well for what it's built for. The distinction we draw is between personal memory and operational memory.
Openclaw remembers things about you - your preferences, your style, what you've said in past conversations. That's powerful for a personal assistant.