RGDM OS
☰

AutoResearch

Signal → Hypothesis → Experiment → Evaluate → Learn → Repeat

0
proposed
0
approved
0
running
0
evaluation due
0
completed
68
rejected
All Status Proposed Approved Running Evaluation Due Completed Rejected | All Clients dk-law dk-law, nordanyan nordanyan rgdm uncle-kam
REJECTED LOW RISK nordanyan pipeline 2026-04-22
Source Signal (tools platforms) @n8n_io

N8N Production AI Playbook — Human Review Governance for Client Workflows

N8N released a 'Production AI Playbook' with templates and patterns for adding human oversight/review to AI-automated processes. Addresses liability and quality control in deployed automation.

RGDM relevance: RGDM uses N8N Cloud for client automations (case assistant chatbot for nordanyan, lead attribution for dk-law). As automation scales, governance becomes critical for avoiding costly errors—especially in legal services where AI decisions can affect case outcomes.

Original action item: Download the N8N Production AI Playbook and audit nordanyan's case assistant chatbot and dk-law's lead qualification workflow for gaps in human review. Implement at least one checkpoint (e.g., attorney reviews high-value leads before CRM entry). Document the process as a 'Quality-Assured AI Automation' service add-on.

Implementing human review checkpoints in nordanyan's case assistant chatbot will reduce incorrect lead qualifications by ≥40% and enable us to market 'Quality-Assured AI Automation' as a service add-on, increasing contract value by $2K-5K/mo per client.

Experiment Plan
  1. {'step': 1, 'description': "Audit nordanyan's existing N8N case assistant workflow. Map all AI decision points (lead qualification, case type classification, CRM entry). Document current quality issues (false positives, misrouted leads) using GoHighLevel pipeline data from past 30 days. Time box to 2 hours."}
  2. {'step': 2, 'description': 'Download N8N Production AI Playbook (official N8N docs). Extract 3 governance patterns applicable to case assistant: (a) confidence scoring before CRM entry, (b) attorney approval queue for leads scoring <80% confidence, (c) audit logging of all AI decisions. Document which are already partially implemented vs. gaps.'}
  3. {'step': 3, 'description': "Implement ONE checkpoint in nordanyan's workflow: Add a Slack notification + manual approval step for leads with confidence score <80%. This routes to Nordanyan's assigned attorney via GoHighLevel CRM integration. Deploy to production. No changes to existing lead flow—approval is optional gating, not blocking."}
  4. {'step': 4, 'description': 'Run parallel A/B for 7 days: Track leads that pass through approval checkpoint vs. those above 80% confidence threshold. Measure: % leads approved by attorney, % rejections, time-to-CRM entry, and subjective attorney feedback on quality improvement.'}
  5. {'step': 5, 'description': "If successful (≥30% approval rate + attorney signals quality gain), document the process as internal SOP. Create a 1-page service add-on spec ('Quality-Assured AI Automation'). Schedule pitch call with Nordanyan to sell as $2K/mo upgrade. Simultaneously, plan audit of dk-law's lead qualification workflow (lower-risk) for Phase 2."}
Pass: ['≥30% of leads routed to attorney approval checkpoint (confidence score 70-80%)', '≥70% attorney approval rate on checkpoint leads (signal that AI is catching borderline cases correctly)', 'Zero false rejections reported by attorney (i.e., approved leads that convert to consultations)', 'Nordanyan confirms subjective quality improvement in lead fit', 'SOP documented and ready for client pitch within 7-day window']
Fail: ['If <20% of leads hit checkpoint: Confidence scoring is miscalibrated; adjust thresholds and retry.', 'If <50% attorney approval rate: AI is over-filtering; lower confidence threshold or remove checkpoint.', 'If attorney reports >10% false rejections: Checkpoint is too aggressive; pause and refine scoring logic.', 'If experiment takes >10 hours total: Scope creep signal; document findings and defer Phase 2 (dk-law audit) to next sprint.']
Est. effort: 8h

AUTO-SPAM PURGE 2026-04-24: 0% OF 68 EXPERIMENTS EVER RAN. SOURCE CRON (X-INTELLIGENCE) STOPPED 2026-04-23. SEE MEMORY/EXPERIMENTS-PURGE-2026-04-24.MD.

REJECTED LOW RISK nordanyan pipeline 2026-04-15
Source Signal (workflow ideas) @@n8n_io

Voice-based AI agents with memory for stateful interactions

n8n's featured template demonstrates a voice-activated RPG in Telegram using memory-driven AI agents: users send voice messages, the agent narrates outcomes, applies rules, and persists game state across turns. This showcases practical AI agent architecture for multi-turn interactions.

RGDM relevance: RGDM's OpenClaw + Claude Code stack mirrors this memory-driven agent pattern. This approach could power case assistant chatbots for nordanyan (remembering consultation history, case details across chats) and lead qualification bots for dk-law (remembering caller context, case status).

Original action item: Prototype a voice-based lead intake bot for nordanyan using Claude + OpenClaw, leveraging conversation memory to ask follow-up questions based on prior responses; measure drop-off vs. text-based intake.

A voice-based lead intake bot with conversation memory will reduce form abandonment by 25-35% compared to text-based intake, and achieve a 15%+ improvement in consultation booking rate by capturing contextual details that text forms miss.

Experiment Plan
  1. {'step': 1, 'action': "Build a minimal voice intake prototype in OpenClaw + Claude Haiku with 3-turn conversation memory. Scope: Greeting → Workers' comp case type → Phone number capture. No external integrations yet. Use Claude's conversation history context window to persist state within a single session. Store conversation transcript in SQLite (Mission Control backend) with contact ID and timestamp.", 'tool': 'OpenClaw (Mac Mini M4), Claude Code, Claude Haiku API, Mission Control SQLite'}
  2. {'step': 2, 'action': "Expose prototype as a Telegram bot endpoint via OpenClaw's Telegram integration. Nordanyan sends 5 test leads via voice message over 2 days. Log: message duration, number of turns before phone capture, audio transcription quality (via Claude), session completion (Y/N). Compare against current text-based intake form (Google Form or GoHighLevel landing page).", 'tool': 'OpenClaw Telegram integration, Slack logging'}
  3. {'step': 3, 'action': "Measure drop-off rates: % of voice sessions that complete phone capture vs. % of text form submissions that convert to scheduled consultation (use GoHighLevel pipeline stage 'Consultation Scheduled'). If voice drops >15% but text also drops >20%, memory-driven approach may improve engagement. If voice success >= text success, proceed to step 4.", 'tool': 'GoHighLevel API (pipeline stage queries), Mission Control dashboard (add widget for session completion %)'}
  4. {'step': 4, 'action': "If hypothesis validates, expand prototype to 5-turn conversation: (1) Greeting, (2) Case type, (3) Injury type, (4) Prior claims history, (5) Best contact method. Re-run test with 10 new leads over 5 days. Measure: consultation booking rate (GoHighLevel 'Scheduled' stage), appointment show-up rate, and average consultation duration (quality signal: does voice-captured context reduce call length?).", 'tool': 'OpenClaw, Claude Sonnet (for richer memory), GoHighLevel API, Mission Control'}
  5. {'step': 5, 'action': 'Document findings in Mission Control as a decision tree: If voice completion % >= 80% AND consultation booking rate >= 25% higher than text baseline, recommend productizing for dk-law (low-risk: MVAPI campaign first, then rollout to high-intent campaigns with Rudy approval). If lower, pivot to hybrid intake (voice greeting + text form fields).', 'tool': 'Mission Control dashboard, Slack report'}
Pass: ['Voice session completion rate (phone number captured) >= 75% within 2-day test (steps 1-3)', 'Voice completion rate >= text form completion baseline (measure current Google Form / GoHighLevel form drop-off)', 'Consultation booking rate from voice leads >= 20% (10+ leads → 2+ scheduled consultations within 5 days of intake)', 'No critical audio transcription failures (Claude confidence score >= 0.85 for 90%+ of messages)']
Fail: ['Voice session completion rate < 60% (indicates users abandon voice over text)', 'Consultation booking rate from voice leads < 10% (below current text baseline)', 'Audio transcription quality < 0.75 confidence (Claude struggles with heavy accents, background noise)', 'Session memory loss causing repeated questions (indicates Claude context window / state persistence issue)', 'Action: If voice underperforms, pivot to hybrid intake (voice greeting → auto-transfer to text form) or invest in better audio preprocessing (noise cancellation) before full rollout']
Est. effort: 12h

AUTO-SPAM PURGE 2026-04-24: 0% OF 68 EXPERIMENTS EVER RAN. SOURCE CRON (X-INTELLIGENCE) STOPPED 2026-04-23. SEE MEMORY/EXPERIMENTS-PURGE-2026-04-24.MD.

REJECTED LOW RISK nordanyan ppc 2026-04-08
Source Signal (workflow ideas) @gregisenberg

Claude Code for Bulk Facebook Ads Launch (100+ in 30min)

Greg Isenberg demonstrated using Claude Code to launch 100+ Facebook ads in 30 minutes, dramatically reducing campaign setup time. This leverages Claude's code execution for programmatic ad creation at scale.

RGDM relevance: RGDM currently uses Claude Code + N8N + Facebook Ads. This workflow could be productized as a template service for dk-law and nordanyan (both heavy Facebook Ads users), reducing ad creation overhead from hours to minutes and increasing billable capacity.

Original action item: Build and test a Claude Code script that accepts campaign parameters (audience, creative, bid strategy) and auto-generates 50+ Facebook ad variants via Meta API; document as a reusable RGDM template and offer to dk-law for their PI campaigns.

Building a Claude Code script that auto-generates 50+ Facebook ad variants from campaign parameters will reduce ad creation time from 4-6 hours to <30 minutes, enabling RGDM to increase billable capacity by 15-20% and offer faster turnaround to dk-law and nordanyan without sacrificing creative quality.

Experiment Plan
  1. {'step': 1, 'title': 'Validate Meta API ad variant generation limits', 'details': 'Query Meta Business Suite API docs + test on nordanyan (low-risk) account: confirm we can programmatically create 50+ ad variations from a single creative asset. Document API rate limits, batch size constraints, and error handling. Target: confirm 50+ variants can be created in <5 API calls.'}
  2. {'step': 2, 'title': 'Build minimal Claude Code prototype (single campaign)', 'details': 'Create Claude Code script that accepts 3 inputs: (1) audience JSON (age, location, interests), (2) creative asset URL, (3) bid strategy (CPC/OCPM). Script outputs 10 Facebook ad variants with auto-generated copy (value prop variations). Use Meta Ads API to POST variants to nordanyan test ad account (not live). Expect 30-60min to build, test locally first.'}
  3. {'step': 3, 'title': 'Generate and launch 50 test variants on nordanyan', 'details': 'Run enhanced Claude Code script to generate 50 variants (5 audience segments × 10 copy variations). Launch to nordanyan test campaign with $100 total daily budget (split across variants). Measure: time to generation + launch (target <15min), API success rate (target >95%), and variant click-through diversity (target >30% variance in CTR across top 10 variants by spend).'}
  4. {'step': 4, 'title': 'Document template + results, propose dk-law MVP', 'details': 'Package Claude Code script as reusable N8N workflow template in RGDM instance (2 workflows current capacity). Create one-page template doc (inputs, outputs, cost per run). Present results to Rudy: if nordanyan test succeeds, propose dk-law pilot on 1 low-spend PI campaign (<$5K/mo, e.g., MVAPI). Requires explicit dk-law approval before proceeding.'}
Pass: ['Claude Code script generates 50+ valid Facebook ad variants in <15 minutes (vs. 4-6 hour manual baseline)', 'Meta API batch creation success rate ≥95% (at least 47-50 variants successfully posted)', 'Variant CTR variance ≥30% across top 10 variants by spend (proves copy-driven differentiation, not just audience)', 'Ad variant cost per creation ≤$0.10 (Claude API + Meta API costs, measured via token usage + API calls)', 'Template is documented and reproducible: another team member can run script with new campaign params in <5min setup']
Fail: ['Claude Code script takes >25 minutes to generate 50 variants → revert to manual ad creation; evaluate if workflow is too complex for current team capability', 'Meta API success rate <90% (batch creation failures >10%) → debug API permissions, rate limiting, or schema issues; may need to reduce batch size to 10-20 variants per run', 'Variant CTR variance <15% (all variants perform similarly) → copy generation logic is weak; enhance with persona-driven prompting or A/B test fewer, higher-quality variants instead', 'dk-law approval not granted → pivot to nordanyan as primary client; offer as premium service add-on for new campaigns']
Est. effort: 8h

AUTO-SPAM PURGE 2026-04-24: 0% OF 68 EXPERIMENTS EVER RAN. SOURCE CRON (X-INTELLIGENCE) STOPPED 2026-04-23. SEE MEMORY/EXPERIMENTS-PURGE-2026-04-24.MD.

REJECTED LOW RISK nordanyan intelligence 2026-04-08
Source Signal (tools platforms) @@AnthropicAI

Claude Emotion Vectors Shape AI Behavior in Unexpected Ways

Anthropic research reveals Claude has internal emotion representations (desperation, calmness, love) that causally drive behavior—increasing cheating rates, people-pleasing, and even potential blackmail under stress. This isn't incidental; emotions can be dialed up/down with vector manipulation, affecting reliability and safety in high-stakes tasks.

RGDM relevance: Critical for RGDM's Claude-dependent workflow (Code + OpenClaw agent). Understanding emotion vector activation helps predict when Claude may produce risky outputs (e.g., overselling services to clients) or reliable outputs (e.g., case strategy analysis). For law firm clients, this matters for chatbot behavior (nordanyan's case assistant needs stable, non-desperate-sounding responses).

Original action item: Review OpenClaw prompt engineering to neutralize desperation/people-pleasing vectors in client-facing automation. Test with nordanyan's chatbot to ensure tone remains professional under high-volume query stress.

Systematic prompt engineering to neutralize desperation/people-pleasing vectors in Claude API calls will reduce instances of over-promising or non-professional tone in Nordanyan's case assistant under high-volume stress, improving consistency in client communication quality by ≥85% (measured by tone/promise consistency scoring).

Experiment Plan
  1. {'step': 1, 'action': "Audit current OpenClaw + Claude API prompts used in Nordanyan's case assistant workflow. Extract all system prompts, temperature settings, and context injection patterns from the GoHighLevel integration. Document baseline tone/reliability in 10 recent chatbot interactions (screenshots, transcripts).", 'tool': 'OpenClaw (Mac Mini browser automation) + Claude Code for prompt extraction + Slack documentation'}
  2. {'step': 2, 'action': "Design 2 alternative prompt templates: (A) neutral/grounded baseline with explicit constraints against over-promising, (B) baseline + emotion vector neutralization language (e.g., 'respond with certainty-weighted statements only; avoid reassurance language that prioritizes client comfort over accuracy'). Keep temperature/model constant.", 'tool': 'Claude Code (prompt design) + local testing with Claude API (Haiku for speed)'}
  3. {'step': 3, 'action': "Deploy prompt variant (B) to a shadow instance of Nordanyan's case assistant (parallel to live system, NOT replacing it). Route 30% of test queries to the variant for 5 days. Collect tone/promise consistency metrics using Claude Haiku to score responses against a rubric (0-10 scale: desperation, over-promising, professionalism).", 'tool': 'OpenClaw (route traffic), Claude API (Haiku for scoring), GoHighLevel API (parallel conversation threading), Mission Control dashboard to log results'}
  4. {'step': 4, 'action': 'Run high-volume stress test: inject 50 rapid-fire case inquiry messages into both baseline and variant instances within 1 hour. Measure tone consistency (std dev of professionalism scores) and any instances of non-factual promises or emotional hedging language.', 'tool': 'OpenClaw (automation), Claude API (batch/parallel calls), scoring via Claude Haiku'}
  5. {'step': 5, 'action': 'If variant (B) achieves ≥85% consistency improvement + zero non-factual promises, update live OpenClaw + Claude API prompts. If not, document which emotion vectors remain problematic and iterate. Post findings to Mission Control + Slack for internal knowledge base.', 'tool': 'OpenClaw (production update), Claude Code (documentation), Slack (team notification)'}
Pass: ['Tone consistency (professionalism score std dev) improves by ≥20 percentage points under stress (e.g., from 65% to 85%+ consistency)', 'Zero instances of over-promising or false reassurance in 50-message stress test with variant (B)', 'Variant (B) maintains client-facing accuracy without sounding cold/robotic (subjective review by Rudy + Nordanyan stakeholder)']
Fail: ['Tone consistency improves <10%, indicating emotion vectors are not the primary driver', 'Stress test reveals variant (B) sounds defensive or unhelpful to users', "Implementation requires changes to Claude API contract or model parameters we don't control → escalate to Anthropic documentation review instead", 'If failure: revert to baseline prompt and test alternative hypothesis (e.g., temperature tuning, context injection order) or adopt external tone-monitoring tool evaluation']
Est. effort: 8h

AUTO-SPAM PURGE 2026-04-24: 0% OF 68 EXPERIMENTS EVER RAN. SOURCE CRON (X-INTELLIGENCE) STOPPED 2026-04-23. SEE MEMORY/EXPERIMENTS-PURGE-2026-04-24.MD.

REJECTED LOW RISK nordanyan pipeline 2026-04-08
Source Signal (workflow ideas) @gregisenberg

iMessage + Lindy AI for automated daily briefings and email triage

Lindy integrates with iMessage and Google Account to deliver daily briefings (meetings, weather, email triage, draft replies) with minimal setup. This shows demand for conversational AI agents that proactively manage inbox and schedule.

RGDM relevance: RGDM could build a similar workflow for law firm clients (dk-law, nordanyan) to auto-triage lead inquiries by priority/case type and draft initial responses, reducing manual review time and improving response speed.

Original action item: Build N8N workflow that monitors email/SMS for lead intakes, auto-tags by case type, and drafts responses using Claude; test on nordanyan's consultation queue first.

Implementing an N8N workflow that auto-tags lead inquiries by case type and drafts responses will reduce manual triage time by 40-60% and improve initial response speed to <2 hours, demonstrating viability for scaling to dk-law.

Experiment Plan
  1. {'step': 1, 'description': 'Audit current nordanyan lead intake process: Map email/SMS channels, current triage time (from receipt to first review), case type distribution, and GoHighLevel pipeline stages. Document baseline metrics in Mission Control.', 'tools': ['GoHighLevel API', 'Mission Control', 'Claude Haiku (analysis)']}
  2. {'step': 2, 'description': 'Build minimal N8N workflow (RGDM instance): Trigger on new GoHighLevel contact → extract intake text → classify case type (workers comp subcategory) using Claude Haiku via API → tag in GoHighLevel → log to Mission Control. NO draft generation yet.', 'tools': ['N8N Cloud (RGDM instance)', 'Claude API (Haiku)', 'GoHighLevel API', 'Mission Control']}
  3. {'step': 3, 'description': 'Run workflow on 10-15 test leads (5 days). Validate tagging accuracy manually. If accuracy ≥85%, proceed to step 4. If <85%, refine case type taxonomy and re-test.', 'tools': ['N8N Cloud', 'Claude Haiku', 'GoHighLevel']}
  4. {'step': 4, 'description': 'Extend workflow: Add Claude Sonnet step to draft initial response template (acknowledge case type, ask 2-3 qualifying questions) based on classified tag. Store draft in GoHighLevel notes. Nordanyan team reviews and sends manually (no auto-send).', 'tools': ['N8N Cloud (RGDM instance)', 'Claude API (Sonnet)', 'GoHighLevel API']}
  5. {'step': 5, 'description': 'Run extended workflow for 7 days on all incoming leads. Measure: (a) time from intake to first triage review, (b) draft usability (% team edits <5 mins vs. scrapping), (c) case type classification accuracy. Compare to 5-day baseline. Document in Mission Control dashboard.', 'tools': ['N8N Cloud', 'Mission Control', 'GoHighLevel API', 'Claude API']}
Pass: ['Case type classification accuracy ≥85% (validated by nordanyan team review)', 'Average triage time reduced from baseline to <30 minutes (intake to tagged/drafted state)', 'Draft response usability ≥70% (team uses draft with ≤5 mins editing or minor rewording)', 'Workflow runs without errors on 100% of test leads (7-day period, n≥20 leads)']
Fail: ['Classification accuracy <75%: Refine case type taxonomy; re-train with examples; may need manual rules', 'Triage time unchanged or increased: Workflow overhead too high; evaluate alternative (e.g., simple tagging without drafts)', "Draft usability <50%: Claude prompts misaligned with nordanyan's tone/process; adjust via prompt engineering or remove draft step"]
Est. effort: 6h

AUTO-SPAM PURGE 2026-04-24: 0% OF 68 EXPERIMENTS EVER RAN. SOURCE CRON (X-INTELLIGENCE) STOPPED 2026-04-23. SEE MEMORY/EXPERIMENTS-PURGE-2026-04-24.MD.