Stop Re-Explaining Your Own Project — a Live Briefing Layer Built From SessionStart and UserPromptSubmit Hooks

What you'll do

A committed, tested briefing layer for one real repository that you own end to end: a fact table stored as data, where each row declares one live fact, its collector command, its freshness tier, and the literal phrases that make it relevant; a collector that runs those commands with a per-fact timeout and writes an atomically-replaced cache, so no reader ever sees a half-written briefing; a SessionStart hook that posts the always-on facts into Claude's context on startup, resume, clear, and compaction, with every value carrying its own age; a UserPromptSubmit hook that stays completely silent by default and volunteers an on-demand fact only when your prompt contains one of its declared trigger phrases — matching by literal comparison, never by regex from a data file and never by interpolating your prompt into a shell command; a Stop hook that refreshes the cache once per turn so the prompt path stays off the critical path; a hard character budget on both injections that sits far below Claude Code's own 10,000-character cap, because injected context is paid for on every later turn; an offline fixture harness that asserts valid JSON, the correct hookEventName, budget compliance, silence on an unrelated prompt, graceful behavior with no cache at all, and — the assertion that matters most — that a canary string planted in a fixture .env file never appears in the cache or in any hook's output; five completed failure drills covering a crashing collector, a hanging collector, an oversized injection, a non-git directory, and a trigger phrase that used to fire on the wrong prompts; a proven post-compaction re-injection; and a documented one-line off switch plus a rollback you have actually used.