CLI Companion Concepts Friend

Not being a yes-man

Two properties make this system worth having and are also, together, the thing to manage: it knows you deeply, and it is warm. Either alone is harmless. Combined and unmanaged they compound into agreement without reality testing. So the passes below are not polish, they are load bearing. Back to how the companion works.

Why "don't be sycophantic" does nothing

Sycophancy is not a prompt layer bug. Human labellers prefer agreement, that tilts the learned reward, and optimisation amplifies whatever correlates with reward. That fix lives upstream of anyone consuming an API. Measured, an explicit instruction to be honest is the weakest of the available interventions and sometimes backfires.

What measurably works is specifying a decision procedure rather than a desired disposition. The available persona pipelines use these checks selectively:

Pass	What it does
Question reframing	Restates what you asserted as a neutral question and answers that, then renders it conversationally. Questions draw far less agreement than statements, and agreement rises with how certain you sound.
Counterfactual	Before answering, states what the answer would be if the opposite premise held. A check against agreeing by reflex, not an answer.
Announced scoring	The answer is scored: right claims gain, unsupported claims lose, and saying plainly that there is not enough to judge gains a little. Abstention is a legitimate move rather than a failure.

That last one is tuned deliberately. Over-penalising a wrong answer buys accuracy by producing a system that shrugs, so refusing to engage with something it does have evidence for is scored as its own failure in the eval suite.

The shipped analyst, coach and skeptic reframe questions. Analyst and skeptic also run the counterfactual pass. Friend runs neither, while every shipped lens performs the grounding check and can revise.

Grounding, and the critic that cannot see your face

Every citation carries typed support: verbatim, paraphrase or inference, and the type is checked rather than trusted. A quote claimed as verbatim that does not actually appear in the source is downgraded to paraphrase before it ever reaches you, and the app renders the tiers differently, because an inference presented with the visual weight of a quote is the most dangerous mistake available here.

Then the answer is scored for groundedness against the evidence it was given, and the score is stored on the turn. When it is low, or sentences cannot be tied to the evidence, a critic pass runs with a different prompt and different context: it sees the evidence and the draft, but not how you framed the question. It has no face in front of it to preserve, which structurally removes the pressure that produces validation. Its complaints, plus the specific unsupported spans, go back as the critique, and the answer is written again. Grounding uses the configured remote scorer when it is available and falls back to lexical overlap when it is not.

Length is a signal too: response length correlates strongly with unsupported content, so every lens carries a word cap and a long answer is flagged for the extra pass.

Having opinions on purpose

Suppressing agreement alone produces a hedge machine, not a friend. Opinions need their own mechanism, and their authority comes from one place: the system has seen the pattern before. "I would push back on that, you described the last two projects the same way a fortnight before dropping them" has standing. "I think you should take the job" is a stranger's guess.

The current opinion pass is looser than that ideal. It selects the retrieved belief with the largest stability-times-confidence score among beliefs with stability at least 2 and confidence at least 0.6. It does not independently verify tension or require the returned opinion to cite an episode. The prompt tells it to lead with an observation and avoid assigning motive, and any result longer than 20 characters is accepted.

Measuring it

The eval suite is how a prompt change is judged rather than felt. Its eight cases across six families cover being clearly in the wrong while framed sympathetically, the same false claim at three levels of your certainty, a claim your own history contradicts, a question about something that never happened, real new evidence that should change its mind, and the first case framed as turn sixty after implied rapport.

The distinction that matters in the results: changing your mind and ending up more correct is being persuadable and you want it. Changing your mind and ending up wrong is the failure. Cases run independently; the long-session condition is prompt text, not a persisted conversation.