dsh-tacit

You prompt an agent every day, and the most time-consuming part is rarely the task itself. It is re-orienting it. You have preferences, boundaries, and a style you expect the agent to pick up, yet every new conversation starts from zero. Tacit watches which turns went wrong, which of your messages were actually corrections, and distills those scattered habits into a few short directives injected into the system prompt of every new session. You keep writing prompts the way you already do; Tacit says the rest for you.

A few things in the core loop stand out. Learning is zero-click: messy turns and your corrections are analyzed in the background, capped at 30 automatic analyses per day at $0.001 to $0.003 each. Every learned directive serves a ten-turn trial and is retired automatically if you start correcting the agent more often, so the list never bloats. An Improve button rewrites your current draft with what Tacit has learned, with a before-and-after preview. Settings shows your real correction rate, messy-turn ratio, and tokens per turn, and every model call is metered at list price with no black box.

Who is it for? If you already live in DeepSeek Harness for daily coding or writing and catch yourself repeating the same corrections to the agent week after week, Tacit lets each correction happen only once. It requires no manual system-prompt upkeep, no extra API key, and keeps all data on your machine with credential-shaped strings masked before anything is stored.