what you get
how it is built
what it believes
what you getTop ↑
B.O.S.S.

It speaks once, then gets out of the way.

This is the part of BOSS people either love or want to switch off, so it's worth being precise about what it does, what it can't do, and every way you can shut it up.

▸a real moment
you've captured four ideas and tested none 
… you submit a prompt — anything — and it speaks once:
 
▸ That's the fourth thing you've added here, and none of it's
  been tested. Who would you talk to first to find out if any
  of them are real? /canvas is one way to pressure-test it —
  but a 15-minute call with the right person beats it.
 
Then it goes quiet. You decide.

Note what it didn't do: it didn't block the prompt, didn't repeat itself, didn't grade you, and it pointed at something better than its own feature.

▸mechanism

how it works

project state files on disk hook a mechanical condition silence the usual path model composes the words only now is a model involved read not met met one message, then quiet
The model never decides whether to speak. A hook evaluates a named condition against files on disk; only once it fires is a model asked to compose anything. Nothing ever asks a model “is this person drifting?” — which is why the judgment is inspectable, arguable, and switchable off.

Detection is mechanical

A lightweight hook watches the state of the work: how many ideas are captured, how many are tested, whether a spec traces to anything the canvas validated. When a named moment's condition is met, it fires. No model is asked “is this person drifting?”

Tact is the model's job

Only once a moment is earned does a model compose the actual words, cohort-aware, so the same moment reads differently for a first-time builder and someone who's shipped three products, and aware of why you said you're building it: a community project is asked whether anyone came back, not whether anyone paid. The judgment about whether to speak is never the model's.

That split is the whole design. A model deciding when to interrupt you would interrupt you constantly, and it would be unfalsifiable. A hook's condition can be read, argued with, and turned off.

▸what it notices

The moments it can speak on

Each one is a named condition with an entry and an exit, not a mood. A few of them, so the shape is concrete:

you've captured, and never tested Several ideas in, none pressure-tested. It asks who you'd talk to first, and points at a 15-minute call ahead of its own feature.
you've built a lot, and nobody outside has seen it It reads the code, not your records, so a founder who just keeps asking for the next feature is seen too: enough code changed since the last time someone was heard from. It asks who has seen the last few things, and stays silent when the work is one thing getting deeper.
you shipped, and nothing tests it Entry is deliberately the exit of the focus loop: at least one feature actually shipped. Exit is one file (two, if the feature named a negative path). A low bar on purpose: it asks whether anything guards the thing you just put in front of people, not whether your coverage is good.
you're building what nothing validated A spec tracing back to no evidence and no canvas cell. Surfaced as restraint, once.
evidence landed, and your user never moved Two pieces of real evidence recorded since your persona was last touched. One of the moments that watch for something going stale rather than something never made. It goes quiet the moment you fold the evidence in.
two of you, and no shared decision Real work has happened and the decision log is empty. It asks whether you're deciding together or in parallel, never whose fault that is.

Refusals are stated in the loop rather than left to drift: it won't quietly self-heal an assertion to make a test pass, and it says so. And when you'd rather ask than wait to be asked, /drift-deep is the same question run deliberately across the whole project at once (canvas, specs, devlog and the actual code): am I fooling myself about any of this?

▸control

Every way to make it stop

None of these require a reason, and none of them are grudging.

the off switches 
boss conscience pause --for 8h   # the whole thing, bounded
boss conscience mute drift       # ONE moment; the rest keep speaking
boss conscience resume           # bring it back
boss status --conscience         # what's open, what would close it,
                                  # and every override you've recorded
  • It never blocks. Not once, anywhere. It has no mechanism to stop you doing a thing, only to name a cost before you do it.
  • It remembers what it said. Once a moment has been voiced it stays quiet on that same condition for a week, across sessions, and speaks again only if the condition changes. Re-raising the unchanged is how care curdles into control.
  • Overrides are recorded, not scored. /decide captures what you chose and why. Nothing counts them against you.
▸the honest part

How it could be wrong

The failure mode isn't speaking harshly; it's speaking too often, which turns a conscience into a nag and trains you to ignore it. BOSS treats that as a measurable bug rather than a matter of taste: every firing is logged, so over-speaking shows up in a frequency ledger you can look at instead of a feeling you have to argue about.

The judgment itself is graded by evals that get re-run when the model curve moves. It is still a young instrument, and it has been tuned mostly against one person's projects. If it talks too much for you, that's worth reporting (/feedback), and it's worth muting in the meantime rather than enduring.

The line it holds itself to

▸attribution

The two practices behind how the conscience speaks — what it names, and the one thing it never does — with who they were learned from.

How the conscience speaks

What it names, and the one thing it never does.

conscience voicing (voice the tension, never filter the menu)

conscience-voicing, owned by mentor-founder, on a craft-moving curve; last checked 2026-06-21.

source BOSS’s own practice, extracted from its build, not from outside

where this came from

Written from BOSS's own rule for how a conscience speaks — voice the tension, never filter the menu — and from watching how a model can name a concern without blocking. It reshaped two of BOSS's mentors (name, never override). No outside source is claimed.

Harm taxonomy (name the axis, don't improvise)

harm-taxonomy, owned by mentor-founder, on a humane-moving curve; last checked 2026-08-20.

learned from Anthropic — Unified Harm Framework Ada Lovelace Institute Brooks, Huang, Kearney & Murray — PNAS 2014

where this came from

Distilled from Anthropic's Unified Harm Framework (five dimensions) and the Ada Lovelace Institute's work on advanced-AI-assistant harms (four relationship harms), with Brooks, Huang, Kearney and Murray (PNAS 2014) behind the dignity axis. Made a practice rather than an agent's private rule so every mentor and the conscience cite one taxonomy.

Deceptive patterns (the catalog is data; this is the judgment)

deceptive-patterns, owned by designer, on a humane-moving curve; last checked 2026-08-20.

learned from FTC — Operation AI Comply EDPB — Opinion 08/2024 (consent or pay) CDT — Dark Patterns in AI Chatbots (2026) Vaccaro et al., CHI'26

where this came from

The three-question test under The other limit is BOSS's own judgment — no research backs it, and it should not be read as though any did. The named patterns come from CDT's Dark Patterns in AI Chatbots (2026, CC-BY), Brignull's deceptive.design, Mathur et al.'s Dark Patterns at Scale, Gray et al.'s CHI-2024 ontology, and successive research sweeps covering accessibility, minors, agentic products, algorithmic management, and the patterns a model writes into generated code (Vaccaro et al., CHI 2026).