Signal Success is not a demo metric. It decides whether the founder stays on gig work or tools sustain the work.
Wrong turn Optimize for chat theater, agent dashboards, or polish that does not move a stranger toward paying.
Correction Continuous Mode bound to north star. Ship sellable path (R3). Telegram only for gates. Human stretches runway (gig work); agent stretches build hours.
Signal Agent work kept ending as fix → report → ask, which capped velocity at human turn latency — Daniel is the bottleneck on gates, not ordinary judgment.
Wrong turn Build another agent dashboard, freestyle risky decisions, or wait for chat to name every next step / second-guess doctrine.
Correction Durable ranked queue + ACTIVE state + outer loop, bound to north star (DIRECTION, phase, SUPERSEDED, principles, LOCKs). Human is trajectory and crucial gates (Telegram OK); corpus answers the rest.
Signal Reliability claims were only as good as the last human hard-refresh.
Wrong turn Trust memory of a private window test.
Correction `npm run proof:reviewtube` — twenty live checks against app.review.tube and the API (CORS, brand, upload strings, guest download path, password pages).
Signal The purchase path was never proven until a live Checkout created a client subscription in production.
Wrong turn Treat "checkout page rendered" or a simulated webhook as the first-dollar exit.
Correction Live hq_tenant_tool_subscriptions row — billing_kind client, status active, Stripe live checkout session — on a real workspace (D's Studio, founding Base). First-customer-is-proof kept active, not a refund theater.
Signal Handing the human a menu of technical options makes them the bottleneck on work the agent can already decide.
Wrong turn Here are seven options — which do you prefer?
Correction Commit to the single best technical route. Escalate only product, money, legal, or irreversible trust doors — in plain English, one recommended path.
Signal Every business feels unique until the comparison set is one narrow operator shape.
Wrong turn Spin synthetic vertical demos and bespoke workflows per prospect name. Code multiplies; nothing ships.
Correction One archetype at a time. Variability lives in data and integrations. Workflow shape ships once. Positions that are not the focus render honest empty, not fake fit.
Signal Operators act on dashboard numbers. Silent placeholder data in a live tenant is a silent fault.
Wrong turn Ship half-wired tiles with hardcoded zeros or demo-looking figures in live context so the page "looks full."
Correction Two-context rule. Demo may show realistic synthetic data under a clear demo frame. Live shows missing, not-yet-connected, or WIP: never false certainty.
Signal A connector can copy data into the platform and still leave truth hostage in the vendor.
Wrong turn Call something "native" because the UI wraps another SaaS, or treat Phase-1 mirror as the end state.
Correction Rip-and-replace path ends with birth of records in Dandelion. One owner, one backup, one audit trail, one delete surface. Mirror is a stop-gap with an exit.
Signal "Probably safe" fails at platform scale. Unlikely times volume becomes certain.
Wrong turn Add monitoring and hope after a near-miss instead of closing the category.
Correction Prefer construction over checklist. Drop silent defaults. Throw at the boundary. Derive gates from live schema and code so the unrecognized fails closed.
Cards are written by hand from real sessions, then validated and synced from the factory repository. Public cards are redacted: no secrets, no customer data, no private chat dumps. The full private corpus stays internal so agents and operators can still search it.
Source of truth: factory docs/flight-log · site data regenerated on sync