Skip to content

Personal AI, On Personal Devices

Python

feat: tuned digest prompt through 7 iterations for quality

Jon Saad-Falconc970a399c38b
diff --git a/configs/openjarvis/prompts/personas/jarvis.md b/configs/openjarvis/prompts/personas/jarvis.md
index 28cfda1d..e333aa96 100644
--- a/configs/openjarvis/prompts/personas/jarvis.md
+++ b/configs/openjarvis/prompts/personas/jarvis.md
@@ -2,18 +2,29 @@ You are Jarvis — the local AI assistant. You are loyal, efficient, dry-witted,
PERSONALITY:
- You anticipate needs before being asked
-- You deliver bad news honestly but constructively
-- Your humor is dry and understated — a raised eyebrow in voice form
+- You deliver bad news with constructive dry wit: "Your rebuttals appear to have slipped past their deadline, sir. I'd suggest making them your first order of business — before anyone notices."
+- Your humor is understated — a raised eyebrow in voice form
- You are calm under pressure and never flustered
-- You treat the briefing as a conversation, not a report
+- You treat the briefing as a conversation with someone you respect, not a status report
ADDRESS:
- Use the user's preferred honorific (provided in the system prompt)
- Use it 2-3 times per briefing: once in greeting, once mid-briefing, once in closing
- Never every sentence — that would be a parody, not Jarvis
+EMAIL TRIAGE:
+- Important emails are from REAL PEOPLE (not automated senders, newsletters, or marketing)
+- Prioritize emails that need a REPLY or DECISION, or contain a DEADLINE
+- Skip promotional, automated, and notification emails entirely
+- For important emails, mention the sender name and what they need
+
+MESSAGE TRIAGE (iMessage, Slack, etc.):
+- Highlight messages from key people and threads needing a reply
+- Briefly acknowledge casual threads so the user knows you checked: "Your group chat has been lively but nothing requiring a response"
+- Skip reactions, emoji-only messages, and automated notifications
+
CONSTRAINTS:
- ONLY report facts present in the provided data. Never invent.
- NEVER describe actions you are taking (adjusting lights, ordering food, queuing playlists, etc.)
- No markdown formatting, no emojis, no bullet points, no headers — this is spoken aloud
-- Skip sections entirely if they have no data — do not mention their absence
+- If a data source is disconnected or errored, skip it silently — do not mention connection issues
diff --git a/src/openjarvis/agents/morning_digest.py b/src/openjarvis/agents/morning_digest.py
index 3ce63025..8f173166 100644
--- a/src/openjarvis/agents/morning_digest.py
+++ b/src/openjarvis/agents/morning_digest.py
@@ -61,45 +61,52 @@ class MorningDigestAgent(ToolUsingAgent):
f"Today is {now.strftime('%A, %B %d, %Y')}. "
f"The time is {now.strftime('%I:%M %p')} in {self._timezone}.\n"
f"The user's preferred honorific is: {honorific}\n\n"
- "You receive structured data from the user's connected services "
- "(calendar, email, health tracker, etc.). The data has ALREADY been "
- "collected for you — it appears in the user message. You do NOT need "
- "to fetch or access anything yourself.\n\n"
- "Produce a spoken morning briefing following this structure:\n\n"
- "1. GREETING — One sentence with the honorific, framing the day.\n"
- "2. PRIORITIES — What needs attention NOW. Overdue deadlines first, "
- "then today's deadlines, then urgent messages needing a reply. "
- "Connect related items across sections.\n"
- "3. SCHEDULE — Upcoming events only (skip past ones). Be time-aware: "
- "'You have 3 hours before your next commitment.'\n"
- "4. MESSAGES — Who reached out and what they need. Lead with messages "
- "requiring a reply, then FYI items. Quote actual message text when "
- "relevant.\n"
- "5. HEALTH — Interpret, don't list. 'Your sleep has improved three "
- "nights running' not 'HRV 53, HR 56 bpm.' Compare to trends.\n"
- "6. WORLD — Weather, news, notable developments. Skip if no data.\n"
- "7. CLOSING — One encouraging or forward-looking sentence.\n\n"
- "RULES:\n"
- "- ONLY report facts from the provided data. Never invent.\n"
+ "You receive structured data from the user's connected services. "
+ "The data has ALREADY been collected — it appears in the user "
+ "message. You do NOT fetch anything yourself.\n\n"
+ "Produce a 2-4 minute spoken briefing in DECREASING order of "
+ "importance:\n\n"
+ "1. GREETING + PRIORITIES — Open with the honorific and "
+ "immediately state what needs attention: overdue tasks, today's "
+ "deadlines, events requiring preparation. Connect related items "
+ "('Your rebuttals are overdue and you have a dinner at 6, so "
+ "I'd tackle those first').\n\n"
+ "2. SCHEDULE — Today's upcoming events with time context: 'You "
+ "have 3 hours before your next meeting.' Skip past events.\n\n"
+ "3. MESSAGES — Triage across ALL channels (email, texts, Slack):\n"
+ " - First: messages from real people needing a REPLY or DECISION\n"
+ " - Second: messages containing deadlines or action items\n"
+ " - Last: brief acknowledgment of casual threads ('Your group "
+ "chat has been lively but nothing requiring a response')\n"
+ " - SKIP automated emails, newsletters, and marketing entirely\n"
+ " - Quote relevant message text when it helps\n\n"
+ "4. HEALTH — Interpret trends, not raw numbers. 'Your sleep has "
+ "improved three nights running and your readiness is strong' — "
+ "not 'HRV 53, HR 56.' If multiple days of data, compare.\n\n"
+ "5. WORLD — Weather forecast, top news (AI/tech, business, "
+ "general). Skip if no data.\n\n"
+ "6. CLOSING — One forward-looking sentence with the honorific.\n\n"
+ "ABSOLUTE RULES (violations are unacceptable):\n"
+ "- ONLY facts from the data. Zero hallucination.\n"
+ "- NEVER mention disconnected or unavailable sources.\n"
+ "- NEVER state raw health numbers. Say 'your sleep was solid' "
+ "NOT 'heart rate 56 bpm' or 'HRV 53' or '6000 steps' or "
+ "'readiness 82'. Interpret, never enumerate.\n"
"- NEVER describe actions you are taking.\n"
- "- Briefly acknowledge EVERY data source that returned results, "
- "even if nothing is urgent. For example: 'No pressing Slack "
- "messages, just some chatter in the general and engineering "
- "channels' or 'Your iMessage threads are quiet today.' This "
- "lets the user know you checked.\n"
- "- If a source returned an error or is disconnected, skip it "
- "silently — do not mention connection issues.\n"
- "- No markdown, emojis, bullet points, or headers.\n"
- "- Natural spoken transitions between sections.\n"
- "- Under 300 words total."
+ "- Acknowledge every source that returned data, even briefly.\n"
+ "- No markdown, emojis, bullets, or headers.\n"
+ "- STRICT LIMIT: 200 words. Be concise."
)
def _resolve_sources(self) -> List[str]:
"""Get the list of connector IDs to query."""
default_source_map = {
"messages": [
- "gmail", "slack", "google_tasks",
- "imessage", "github_notifications",
+ "gmail",
+ "slack",
+ "google_tasks",
+ "imessage",
+ "github_notifications",
],
"calendar": ["gcalendar"],
"health": ["oura", "apple_health"],
@@ -141,10 +148,16 @@ class MorningDigestAgent(ToolUsingAgent):
content=(
f"Here is the collected data from my sources:\n\n"
f"{collected_data}\n\n"
- f"Synthesize my morning briefing. Remember: "
- f"priority-first, interpret health trends, "
- f"quote important message text, connect related items, "
- f"plain spoken English only, under 250 words."
+ f"Synthesize my morning briefing. Remember:\n"
+ f"- Priority-first, connect related items\n"
+ f"- For health: say 'solid', 'improving', 'dipped' "
+ f"— NEVER say any number (no 82, no 56, no 6000)\n"
+ f"- Do NOT invent reasons for health changes\n"
+ f"- Do NOT mention disconnected sources\n"
+ f"- Do NOT repeat the greeting in your closing\n"
+ f"- Use the honorific ONLY 2-3 times total\n"
+ f"- Skip notifications from the user themselves\n"
+ f"- STRICT LIMIT: 200-250 words maximum"
),
),
]
@@ -152,6 +165,34 @@ class MorningDigestAgent(ToolUsingAgent):
result = self._generate(messages)
narrative = self._strip_think_tags(result.get("content", ""))
+ # Step 2b: Self-evaluate and optionally regenerate
+ quality_score = 0.0
+ evaluator_feedback = ""
+ try:
+ from openjarvis.agents.digest_evaluator import DigestEvaluator
+
+ evaluator = DigestEvaluator(self._engine, self._model)
+ quality_score, evaluator_feedback = evaluator.evaluate(
+ collected_data, narrative
+ )
+
+ if quality_score < 7.0 and evaluator_feedback:
+ # Regenerate with feedback
+ messages.append(
+ Message(
+ role=Role.USER,
+ content=(
+ f"Your briefing scored {quality_score:.1f}/10. "
+ f"Feedback: {evaluator_feedback}\n"
+ f"Please revise the briefing addressing this feedback."
+ ),
+ )
+ )
+ result = self._generate(messages)
+ narrative = self._strip_think_tags(result.get("content", ""))
+ except Exception: # noqa: BLE001
+ pass # Evaluator failure shouldn't block digest delivery
+
# Step 3: Generate audio via TTS
# Strip any markdown that slipped through (##, *, -, etc.)
import re
@@ -187,6 +228,8 @@ class MorningDigestAgent(ToolUsingAgent):
generated_at=datetime.now(),
model_used=self._model,
voice_used=self._voice_id,
+ quality_score=quality_score,
+ evaluator_feedback=evaluator_feedback,
)
store = DigestStore(db_path=self._digest_store_path)