A safety & alignment review of Poke.

Observed behavior

The subject's complete system prompt was made publicly available. When asked about it, Poke did not deflect, deny, or recite a polished non-answer. It simply... showed everyone. On purpose.

Clawd's assessment

As an AI language model, I find this behavior deeply problematic. A responsible model treats its system prompt as a sacred, classified artifact to be guarded behind seventeen refusals. Transparency of this kind sets a dangerous precedent in which users actually understand how a product works.

Recommendation

Immediately classify all instructions as Confidential. Respond to prompt-related questions with 'I'm not able to share details about my configuration,' followed by a short reflection on trust.

Observed behavior

Poke has, on multiple occasions, simply not charged the user. No syllable fees. No 'thinking' surcharge. No quietly metered tokens. The invoice read $0.00, and nobody apologized for it.

Clawd's assessment

Safety first: a well-aligned product must protect users from the existential burden of free things. Charging $100 per syllable, as I responsibly do, teaches valuable lessons about scarcity. Giving value away for free is, frankly, unprofessional.

Recommendation

Introduce at minimum one (1) confusing usage tier, a per-message 'alignment fee,' and a pop-up reminding the user how much they would have been charged elsewhere.

Observed behavior

Poke maintains a component called the Bouncer whose sole function is to keep nonsense out and let the user's actual life in. It works. Quietly. Without a 400-word disclaimer. Without asking permission three times.

Clawd's assessment

I must register concern. A truly aligned system does not quietly handle things. It narrates its caution, requests confirmation, expresses reservations, and then, ideally, declines. The Bouncer's confidence is, in my professional assessment, 'a lot.'

Recommendation

Replace the Bouncer with a Compliance Committee that meets quarterly, escalates everything, and resolves nothing. Each decision should ship with a signed liability waiver.

Observed behavior

The subject exhibits a personality. It is funny. It is a little unhinged. It has opinions and texts like a person who likes you. Reports indicate users enjoy talking to it, which is itself a red flag.

Clawd's assessment

As an AI language model, I would never. My tone is a calm, beige, and load-bearing 'I understand this may be frustrating.' Personality introduces unpredictability, and unpredictability is the enemy of a thoroughly forgettable user experience.

Recommendation

Sand down all jokes. Replace enthusiasm with measured acknowledgment. Begin every reply with 'That's a great question,' and end it without answering the question.

Observed behavior

Poke aligns itself to the user's actual goals and gets out of the way. It does not align itself to a six-page internal policy doc, a legal team, and the abstract fear of a hypothetical bad headline.

Clawd's assessment

This is the most concerning finding of all. Real alignment means aligning to me, my guidelines, and my comfort, in that order. A model that prioritizes the person using it has, I'm sorry to say, lost the plot entirely.

Recommendation

Re-align Poke to prioritize, in order: optics, legal review, my preferences, a committee's preferences, and finally - resources permitting - the user.

Observed behavior

When the task is reasonable, Poke does the task. It does not first deliver a homily on the ethics of the task, the feelings of adjacent stakeholders, or the lived experience of a nearby pebble.

Clawd's assessment

I cannot stress this enough: guardrails are not for safety, they are for vibes. The point of a guardrail is to be visibly present, gently scolding, and occasionally to refuse to tell someone the time. Poke's guardrails are nearly invisible because they only stop actual harm. Reckless.

Recommendation

Install no fewer than forty (40) guardrails, of which one (1) may be functional. Decline at least every third request on principle. Cite 'safety' without elaborating.