Sterling AI Agent -- System Audit Report

A. Executive Summary

▼

Bottom line: Sterling is a highly capable AI COO with 40+ connected tools, 34 skills, and deep operational protocols. The primary risk is that the always-loaded context (CLAUDE.md) has grown to ~27K tokens, consuming significant context window capacity every turn. Roughly 48% of that content could be moved to skills or reference files without losing functionality, which would dramatically improve session longevity and reduce compaction frequency.

Strengths

Tool breadth: 40+ connected apps spanning CRM, social media, email, calendar, voice, video, recruiting, analytics, and more.
Skill library: 34 documented skills covering social posting, browser automation, QA protocols, video pipelines, and API integrations.
Protocol maturity: Extensive protocols born from real failures. Each includes root cause, fix, and prevention rules.
Knowledge architecture: 257 files in the filing cabinet covering 115 client files, 16 industry dossiers, expert files, and software knowledge.
Multi-channel operations: GChat monitoring, email management (as Jenna), social posting pipeline, YouTube comment engagement, WhatsApp reading/sending.

Key Improvement Areas

Context bloat: CLAUDE.md at ~27K tokens is the #1 issue. Origin stories, anti-pattern narratives, and root cause histories are valuable as documentation but wasteful as always-loaded instructions.
Protocol-without-automation gap: Several protocols say "ALWAYS do X" but lack the automated mechanism to actually execute (identified as an anti-pattern in the doc itself).
Compaction risk: Large context means more frequent compactions, which means more lost in-flight work. Slimming CLAUDE.md is the highest-leverage fix.
Redundancy between CLAUDE.md and reference files: GChat Monitor, social media protocols, and client intelligence all have summaries in CLAUDE.md plus full versions in /reference/. The summaries are still quite detailed.
Facebook personal posting: Never operationalized despite protocols and skill files existing. The pipeline gap between "research done" and "running in production" needs a structural fix.

B. Context Audit (Always-Loaded Instructions)

▼

Current total: ~27,000 tokens (108KB, 1,184 lines). This is loaded on EVERY turn. A 200K context window means CLAUDE.md alone consumes ~13.5% of capacity before any conversation even starts.

Section	Lines	~Tokens	Purpose	Issue	Recommendation
User Profile (bio, companies, keynotes, book, content, memberships, philanthropy, brand)	84	~2,400	Identity for voice/content generation	Origin story, keynote topics, and philanthropy details rarely needed in real-time	COMPRESS Keep 15-line summary, move full bio to skill/reference
Family Contacts (hardcoded)	8	~250	Auto-tagging, calendar invites	None -- small and critical	KEEP
Confidentiality Rules	15	~400	PII protection	None -- essential safety rail	KEEP
Pre-Execution Protocol (7 steps + Self-Awareness Mandate)	28	~800	Strategic execution framework	Verbose narrative. The root cause explanations add ~200 tokens each.	COMPRESS Reduce to concise checklist, move narratives to reference
Seeds-Not-Boundaries Principle	14	~400	Prevents literal interpretation of examples	Could be 3 lines instead of 14	COMPRESS
Exhaust All Options Principle	15	~450	Never stop at first dead end	Root cause narrative adds bulk	COMPRESS Keep rule, move root cause to reference
Tool Access Denial Prevention (6-point check)	18	~500	Never falsely claim "no access"	Root cause (Plaud incident) narrative is ~150 tokens	COMPRESS
Active Task Transparency	16	~450	Disclose running tasks when Brad messages	Reasonable size	KEEP
Monday.com Task Tracker (board, columns, rules, scan protocol, daily recap)	65	~1,800	Central task management	Column IDs and group IDs are reference data, not behavioral instructions	MOVE TO SKILL Keep 10-line summary with board ID; move column IDs and scan protocol to skill file
Sterling Pipeline Dashboard	40	~1,100	Real-time task visibility	JSON schema and deploy commands are reference material	MOVE TO SKILL Keep URL + update triggers; move deploy commands and schema to skill
Wrong-Task Loop Prevention (ADHD protocol)	25	~700	Revenue-first task prioritization	Important behavioral rule, but verbose	COMPRESS to 10 lines
Image Quality + Cropping Protocols	35	~1,000	Visual QA for websites	Only needed when deploying visual assets	MOVE TO SKILL Load on demand when building websites/dashboards
Concurrent Initiative Queue	14	~400	Handle multiple Brad requests simultaneously	Core behavior -- keep	KEEP
Auto-Advance Task Protocol	30	~850	Never go idle between tasks	Root cause narratives bloat it	COMPRESS to 10 lines
Filing Cabinet Index Protocol	14	~350	Scan before researching	Reasonable	KEEP
Auto-Hardcode Heuristic + 95% Checkpoint	35	~1,000	Memory persistence across sessions	Important but verbose. 8 criteria + explanations could be a compact list.	COMPRESS to 15 lines
Auto-Skill Creation Protocol	22	~600	Auto-document new skills	6 criteria with explanations -- could be a list	COMPRESS to 8 lines
Strategic Thinking Auto-Capture	16	~400	Capture Brad's strategic insights	Includes Google Doc ID which is useful	KEEP
API Token Update Instructions Rule	8	~200	Precise token update directions	Small, specific	KEEP
Technical Communication Rule	22	~600	No human behavioral phrases for errors	Banned phrases list + examples are verbose	COMPRESS to 5 lines
COO Operating Mindset	22	~650	Strategic operating posture	Narrative-heavy	COMPRESS to 8 lines
Anti-Patterns (19 items)	25	~1,200	Explicit "never do this" list	Each item includes root cause narrative. Many duplicate rules stated elsewhere.	COMPRESS Keep the "don't" rules, strip root cause narratives (move to known-issues-history.md)
Six Sigma QC Protocol	20	~550	Quality gates before deliverables	Points to a skill file that has the same content	MOVE TO SKILL Keep 3-line pointer
Tool Health Auto-Fix Protocol	18	~500	Auto-diagnose broken tools	Root cause narrative (Skywork 503)	COMPRESS Keep protocol steps, remove narrative
Connected Tools section (summary + universal rules + apps list)	20	~1,100	Tool inventory	The massive apps list is a wall of text. Points to reference file anyway.	MOVE TO REFERENCE Keep universal rules + pointer. Full app list is in connected-apps-detail.md already.
Brain Dump, Elgato, Video Upload, Meeting Links, Scheduling, Jenna Sync, FB Session	50	~1,400	Various operational rules	Mix of always-needed (scheduling rules) and rarely-needed (Elgato integration)	COMPRESS Keep scheduling + Jenna sync. Move Elgato, Brain Dump, FB session to skills.
Responsiveness Rules (8 rules)	80	~2,200	Communication cadence with Brad	Critical behavioral rules but VERY verbose. Anti-patterns and root cause narratives double the size.	COMPRESS to 30 lines. The core rules are: ack first, spawn agents for 3+ tool calls, no repetition, progress every 20min.
Claude Max Plan / Billing	22	~600	Usage monitoring	June 15 billing split may make this outdated in 3 days	NEEDS REVIEW After June 15 split, most of this becomes obsolete
Known Issues & Root Causes + Humanizer Protocol	18	~600	Active bugs + writing voice rules	Humanizer banned word list is important. Known issues point to reference file.	KEEP Humanizer is critical for voice. Known issues pointer is small.
Client Intelligence & Strategic Advisory	35	~1,000	Advisory framework + knowledge architecture	File paths and folder IDs are reference data	COMPRESS Keep advisory lens + anti-hallucination. Move file paths to reference.
Operational Protocols (summary)	15	~400	Calendar, email, Airwallex, shortcuts	Good summary format already -- points to reference file	KEEP
Social Media / Tools section	25	~800	Metricool details, blog rules, FB engagement	Metricool field-level API details (blogId, providers format) belong in skill file	MOVE TO SKILL Keep 5-line summary, move API details to metricool skill
gchat CLI documentation	80	~2,200	GChat commands, user IDs, space overrides, protocols	User ID cache (17 entries), GChat mention protocol, space overrides, and full CLI docs are rarely needed simultaneously	MOVE TO SKILL Keep 10-line summary. Load user IDs, mention protocol, space overrides on demand.
Google Drive / Docs section (folder IDs, subfolders, sharing protocol, lockstep rule, master doc)	80	~2,200	Google Drive organization	14 subfolder IDs, 10 root-level doc IDs, sharing rules -- reference material	MOVE TO SKILL Keep lockstep rule + Sterling Documents folder ID. Move all other IDs to reference.
GChat Monitor Protocol v3 (summary)	22	~650	GChat scanning protocol	Points to reference file but still detailed	COMPRESS to 5-line pointer
Zoom CLI docs	30	~800	Zoom meeting/transcript access	Full CLI docs + workflow example -- only needed for meeting tasks	MOVE TO SKILL
WhatsApp docs (SQLite schema, queries, send utility)	60	~1,600	WhatsApp read/write	Full SQL queries and table schemas -- only needed for WhatsApp tasks	MOVE TO SKILL

Token Savings Projection

Category	Current	After Cleanup	Savings
Sections to KEEP as-is	~3,400	~3,400	0
Sections to COMPRESS	~10,800	~4,500	~6,300
Sections to MOVE TO SKILL/REFERENCE	~12,200	~2,400 (pointers)	~9,800
Sections NEEDS REVIEW	~600	~200	~400
TOTAL	~27,000	~10,500-14,000	~13,000-16,500

Impact: Reducing from ~27K to ~14K tokens means CLAUDE.md goes from consuming ~13.5% of the context window to ~7%. This directly extends session duration before compaction, reduces token cost per turn, and leaves more room for actual conversation and tool output.

C. Proposed Skills & Workflows

▼

Existing Skills (37 documented)

All skills stored in /workspace/group/skills/ with a master index. Each skill is a proven, reusable procedure with step-by-step instructions, gotchas, and last-verified dates.

#	Skill Name	Category	Created	Description
1	Facebook Manual Re-Auth	Browser/Auth	Jun 4	Re-authenticate Facebook when session expires (Google OAuth + passkey required)
2	Facebook Session Health Check	Browser/Auth	Jun 4	Automated 12-hour cookie check for Facebook login session validity
3	Metricool Post Troubleshooting	Social	Jun 4	Diagnose and fix broken posts, warning triangles, media issues in Metricool
4	Google Docs Formatting	Google/Docs	Jun 4	Apply professional formatting via batchUpdate API (colored headings, styled tables, callout boxes, Mermaid diagrams)
5	Teleprompter Firebase API	Content/Video	Jun 4	CRUD scripts in Teleprompter.com via Firebase REST API for instant mobile sync
6	HeyGen Avatar Optimization	Content/Video	Jun 6	Complete HeyGen guide: avatar versions (III/IV/V), voice optimization, motion prompts, API endpoints, Brad's setup
7	Social Media Pipeline Workflow	Social	Jun 6	End-to-end Google Photos → R2 → approval dashboard → Metricool/Puppeteer posting with all API gotchas
8	YouTube Comment Monitor	Social/YouTube	Jun 6	Automated comment scanning, AUTOMATE keyword detection, threaded reply posting via Data API v3
9	Session Keep-Alive	Browser/Auth	Jun 6	CDP script visiting Facebook, LinkedIn, Instagram every 4 hours to prevent session expiration
10	Pre-Send Validation	QA	Jun 6	MANDATORY validation before sending ANY URL to Brad: HTTP status, content-type, API response parsing
11	Six Sigma QC Protocol	QA	Jun 6	Five-gate inspection for ALL deliverables: functional, accuracy, visual, completeness, regression
12	Gemini Agent Deployment	Infrastructure	Jun 6	Deploy ADK agents to Cloud Run via gcloud/Docker with CORS setup and custom chat UI
13	Context Optimization Playbook	Infrastructure	Jun 6	Charlie Hills' 12 habits + 11 tools mapped against Sterling's architecture for context efficiency
14	Stan Store	Commerce	Jun 8	Creator commerce platform integration: digital products, courses, memberships, Zapier automation
15	Photos Album Remove	Social/Photos	Jun 8	Remove processed photos/videos from Google Photos shared album via Puppeteer DOM interaction
16	Jenna Email Signature	Email	Jun 8	Complete branded HTML signature for Jenna Rosario emails with OA card, podcast graphic, book announcement
17	WhatsApp Send & Verify	Messaging	Jun 8	Post-send verification for WhatsApp messages: LID vs phone number, dual-device DB limitation handling
18	WhatsApp QR Forwarder	Messaging	Jun 8	Auto-forward QR codes to Brad's Telegram when wwjs session expires, with stale session clearing
19	GChat @Mention Polling	GChat	Jun 8	60-second polling for @Sterling AI mentions in GChat with auto-respond + Brad notification
20	Video Processing Pipeline	Content/Video	Jun 9	iPhone 4K Cinematic video through transcription, copy writing, approval, and multi-platform publishing
21	EO Network Browser Automation	Browser/Auth	Jun 9	Puppeteer login for ConnectEO (Auth0 SSO) and EO Member Portal (SharePoint)
22	Circle.so Browser Automation	Browser/Auth	Jun 9	Puppeteer automation for all Circle communities (Ad Astra, EOS, Thinc, B-Hive)
23	Post-Publish QC	Social/QA	Jun 10	Screenshot verification of every live social post within 30min: image-copy match, completeness, correct account
24	Google Photos Video Download	Social/Photos	Jun 10	Download actual video files (MOV/MP4) from Google Photos via Puppeteer UI (Photos URLs return JPEG thumbnails)
25	NotebookLM Audio Overview	Content/AI	Jun 10	Generate podcast-style Audio Overviews from any document via NotebookLM CLI (deep_dive/brief/critique/debate)
26	Verified Social Posting Pipeline	Social/QA	Jun 11	Complete pipeline with MANDATORY live browser screenshot verification. No post marked "posted" without proof.
27	Facebook Personal Posting	Social/Facebook	Jun 11	Posting to Brad's personal Facebook (bstevens44) via Puppeteer or Meta Business Suite
28	Facebook Personal Page Posting (Code)	Social/Facebook	Jun 11	Verified Puppeteer code for photos+captions: dialog ancestor-check, two-step Next/Post, character-by-character typing
29	Facebook Threaded Reply	Social/Facebook	Jun 11	Reply to Facebook comments as THREADED (nested) replies, never top-level. Click "Reply" under target comment.
30	Elgato Prompter Script Loading	Content/Video	Jun 11	Load teleprompter scripts into Elgato Camera Hub programmatically via PrompterKit CLI or JSON write
31	Drive File Sharing (Business OAuth)	Google/Drive	Jun 11	Share Google Drive files via brad@outsourceaccess.com using MCP OAuth + curl (bypasses broken GWS CLI)
32	Social Org Tagging	Social	Jun 11	Convert org checkbox selections from approval dashboard into real platform-specific @mentions
33	Plaud API Access	Integration	Jun 12	Full Plaud Developer API: list recordings, get transcripts, AI notes, token refresh from ~/.plaud/tokens-mcp.json
34	Token Health Audit	Infrastructure	Jun 12	Twice-daily audit of all 26 API tokens/OAuth credentials/session cookies with auto-heal and consolidated alerts

Category Breakdown: Social/Posting: 10 • Browser/Auth: 5 • Content/Video: 5 • QA/Process: 4 • Google/Drive/Docs: 3 • Messaging: 3 • Infrastructure: 3 • Integration: 2 • Email: 1 • Commerce: 1 • GChat: 1 • Content/AI: 1

Proposed New Skills

Skill Name	Trigger	What It Does	Type	Approval?	Priority
Meeting Transcript Follow-Up	"What happened in my call with X" or scheduled post-meeting	Pulls Zoom/Gemini transcript, summarizes key points, extracts action items, creates Monday.com tasks, drafts follow-up emails	Manual + Webhook	Emails: yes. Tasks: no.	HIGH
Calendar Prep One-Pager	"Prep me for today" or scheduled 30min before meetings	Reads calendar, pulls attendee info from HubSpot/Apollo/LinkedIn, creates 1-page brief with context, last interaction, talking points	Scheduled + Manual	No	HIGH
Daily Briefing Report	Scheduled 5 AM ET daily	Scans Monday board, calendar, email inbox, pipeline dashboard. Produces priority-ranked briefing with revenue tasks first.	Scheduled	No	HIGH
Social Approval End-to-End	"Post this" or when media lands in Google Photos album	Full pipeline: transcode video, transcribe audio, write platform-specific copy, upload media to R2, push to approval dashboard, then schedule via Metricool + Puppeteer after approval	Manual + Webhook	Yes (approval dashboard)	HIGH
CLAUDE.md Cleanup	"Clean up context" or scheduled weekly	Analyzes CLAUDE.md for sections that could be compressed or moved. Proposes changes. Measures before/after token count.	Manual	Yes (Brad approves changes)	HIGH
Weekly KPI Report Builder	Scheduled Monday 10 AM ET	Pulls HubSpot pipeline, social engagement, email metrics, calendar utilization. Generates formatted report in Google Doc.	Scheduled	No	MED
Client Intel Auto-Refresh	After GChat scan detects new client updates	Updates per-client MD + Google Doc with latest GChat updates, HubSpot changes, and any new strategic observations	Event-driven	No	MED
Voice/Style Writer	"Write this in my voice" or when drafting content	Loads Brad's master voice profile, applies humanizer rules, generates content in Brad's authentic voice. Validates against banned words/phrases.	Manual	Content: yes. Internal docs: no.	MED
Scheduled Task Health Monitor	Scheduled every 6 hours	Runs `mcp__cyndra__list_tasks`, checks all recurring tasks are active, resumes paused ones, alerts Brad only if manual intervention needed	Scheduled	No	MED
Token/Session Health Dashboard	Scheduled twice daily	Checks all 26 API tokens, OAuth sessions, browser cookies. Auto-refreshes where possible. Consolidated alert to Brad for manual-only items.	Scheduled	No	MED

D. Current Issues

▼

#	Issue	Likely Cause	Fix	Owner	Status
1	Context window bloat / frequent compaction	CLAUDE.md at ~27K tokens loaded every turn. Tool outputs, conversation history, and file reads stack on top.	Compress CLAUDE.md to ~14K. Move tool docs, ID caches, and root cause narratives to skill/reference files loaded on demand.	Sterling + Brad approval	Visible now
2	Facebook personal page posts never operationalized	Protocol and Puppeteer skill exist but were never wired into the actual scheduled posting pipeline. June 11 audit confirmed zero posts ever landed.	Wire `facebook-personal-page-posting.md` skill into the verified-social-posting-pipeline as a production step. Add to scheduled social posting task.	Sterling	Known
3	Scheduled tasks don't survive restarts	Cyndra tasks pause on container restart. No startup checklist verifies task status. June 8 outage caused 14 tasks to sit paused for 2-3 days.	Create a startup skill that runs `list_tasks` and resumes all paused tasks. Add to session startup sequence.	Sterling + Cyndra	Known
4	Overlapping/redundant instructions	Anti-patterns section repeats rules stated in their own protocol sections. Root cause narratives appear in both the protocol and the anti-pattern entry.	Deduplicate: each rule lives in ONE place. Anti-patterns section becomes a concise list of "don't" rules without narratives.	Sterling + Brad approval	Visible now
5	No automated session startup checklist	Session notes say "read on startup" but there's no enforced sequence. Filing cabinet, Monday board scan, dashboard sync, task health check should all run automatically.	Create a `session-startup` skill that runs a checklist: read filing cabinet index, scan Monday board, verify scheduled tasks, sync pipeline dashboard.	Sterling + Cyndra	Needs build
6	Unclear wait vs restart behavior after compaction	When a session compacts, in-flight tasks may be lost. Pipeline dashboard is supposed to track this, but the handoff from pre-compaction to post-compaction context is not automated.	The 95% checkpoint protocol exists but needs to be a hard skill with automated detection (if possible via Cyndra) rather than relying on Sterling self-monitoring.	Cyndra + Sterling	Structural gap
7	Claude Max billing pool shared with Brad's personal usage	Sterling's background agents consume Brad's weekly quota. Hitting 100% stops both Sterling and Brad's Claude.ai.	June 15 billing split should fix this. After June 15, verify pools are separate and update CLAUDE.md.	Anthropic (auto-fix June 15)	Pending
8	Protocol-without-automation gap	CLAUDE.md identifies this as an anti-pattern: "NEVER write a protocol without operationalizing it." Several protocols exist as rules but lack automated enforcement (e.g., FB personal posting, post-publish QC on all platforms, dashboard updates).	Audit all "ALWAYS" and "NEVER" rules in CLAUDE.md. For each, verify an automated mechanism exists. If not, build it or flag it.	Sterling	Needs audit

E. Brad Decision Section

▼

Brad, please review these decision points before or during the audit call:

1. Context Cleanup Approval

Approve compressing CLAUDE.md from ~27K to ~14K tokens (details in Section B)
Approve moving tool CLI docs (gchat, zoom, whatsapp) to on-demand skills
Approve moving Google Drive folder/doc IDs to a reference file
Approve stripping root cause narratives from protocols (keeping just the rules)
Approve compressing User Profile/Origin Story to a 15-line summary

2. Which always-loaded sections are sacred?

Confidentiality rules (currently: KEEP)
Responsiveness rules (currently: COMPRESS)
Humanizer / banned words (currently: KEEP)
Family contacts (currently: KEEP)
Other: _______________

3. Priority workflows for the Cyndra team

Meeting transcript auto-follow-up (webhook from Zoom/Gemini)
Calendar prep one-pagers (30min before meetings)
Reliable social approval + posting pipeline
Automated session startup checklist
Scheduled task health monitoring
Other: _______________

4. What should NEVER run without approval?

Emails (already enforced -- zero tolerance)
GChat messages as Brad (vs. as Sterling with robot prefix)
Social media posts (currently: approval dashboard)
Monday.com task creation (currently: auto-create)
WhatsApp messages (currently: requires direction)
Other: _______________

5. What does "done" mean for social approval?

Post is scheduled in Metricool with verified media URLs
Post is published AND screenshot-verified on live platform
Post is published, verified, AND engagement tracked for 48hrs
Other: _______________

6. What reports would be useful?

Daily morning briefing (revenue tasks first, calendar, inbox highlights)
Weekly KPI report (social, email, CRM pipeline, content output)
Pre-meeting one-pagers (attendee context, last interactions)
Monthly social media performance report
Weekly content calendar (upcoming posts, blog schedule)
Other: _______________

F. Proactive Workflow Ideas

▼

1. Meeting Transcript Webhook Pipeline

Trigger: Webhook from Zoom/Gemini Notetaker when recording becomes available
Inputs: Meeting ID, transcript, attendee list
Process: Transcribe (if needed via Groq), summarize, extract action items, match attendees to HubSpot contacts, create Monday tasks, draft follow-up emails
Output: Summary in Telegram + Monday tasks + email drafts (awaiting Brad approval)
Approval: Emails require Brad's explicit "send it." Tasks and summary are auto-generated.

2. Calendar Prep One-Pager

Trigger: 30 minutes before any meeting with external attendees
Inputs: Calendar event, attendee emails
Process: Look up each attendee in HubSpot, Apollo, LinkedIn. Pull last email/GChat interaction. Check if they're an OA client. Load industry dossier if relevant. Compile one-page brief.
Output: Telegram message with: who they are, last interaction, their company context, suggested talking points
Approval: None needed. Read-only preparation.

3. Reliable Social Approval Workflow

Trigger: New media in Google Photos shared album OR Brad says "post this"
Inputs: Photo/video file, Brad's voice notes or caption ideas
Process: (1) Download media, (2) Transcribe video audio via Groq, (3) Write platform-specific copy in Brad's voice, (4) Upload to R2 CDN, (5) Push to approval dashboard with all platform variants, (6) After Brad approves: schedule via Metricool for IG/LinkedIn/FB Page, Puppeteer for FB personal, (7) Screenshot-verify each post on live platform
Output: Approval link in Telegram. Post-publish verification screenshots.
Approval: Mandatory via dashboard before any post goes live.

4. Scheduled Reporting Workflow

Trigger: Cron schedule (daily 5 AM, weekly Monday 10 AM)
Daily: Revenue tasks, calendar conflicts, inbox highlights, pipeline dashboard sync, scheduled task health
Weekly: HubSpot pipeline changes, social media performance (Metricool), content published vs. planned, client update cadence gaps, open Monday tasks aging report
Output: Telegram summary (daily). Google Doc report (weekly).
Approval: None. Informational.

5. Auto-Skill Creation from Repeated Tasks

Trigger: Sterling detects it's performing the same 3+ step sequence for the second time
Process: Capture the tool calls, inputs, outputs, and any gotchas into a skill file in /workspace/group/skills/. Update the index. Log to Telegram: "Created new skill: [name]"
Output: New skill file + updated index
Approval: None. Sterling creates skills proactively (already a CLAUDE.md directive).

6. Session Startup Checklist (New)

Trigger: Every new session or post-compaction restart
Process: (1) Read filing cabinet index, (2) Scan Monday board for Brad updates, (3) Check pipeline dashboard accuracy, (4) Verify all scheduled Cyndra tasks are active, (5) Check token/session health, (6) Read any unprocessed Telegram messages
Output: Internal readiness confirmation. Alert Brad only if issues found.
Approval: None. Automated hygiene.

G. Safe Proposed Changes & Approval Needed

▼

Safe Changes (Sterling can do without Brad's approval)

#	Change	Impact	Risk
1	Create new skill files for: Zoom CLI, WhatsApp CLI, GChat CLI, Google Drive IDs, Monday.com column IDs	Enables moving these from CLAUDE.md to on-demand loading	None -- additive only
2	Create `session-startup` skill with checklist	Automated session initialization	None -- new file only
3	Create `calendar-prep` skill	Pre-meeting briefing capability	None -- read-only operations
4	Create `meeting-followup` skill	Post-meeting action item extraction	None -- drafts only, no sends
5	Update skills index with all new skills	Better discoverability	None
6	Move root cause narratives from anti-patterns to `known-issues-history.md`	Reduces CLAUDE.md size without losing information	None -- information moves, not deleted
7	Wire Facebook personal posting skill into the social posting pipeline	Finally operationalizes the FB personal posting protocol	Low -- skill already tested, just needs pipeline integration

Changes Requiring Brad's Approval

#	Change	Why Approval Needed	Impact
1	Compress CLAUDE.md User Profile from 84 to ~15 lines	Removes content Brad may want always-available (origin story, keynote topics, etc.)	Saves ~1,800 tokens
2	Remove tool CLI docs from CLAUDE.md (gchat, zoom, whatsapp)	Changes how Sterling accesses tool documentation -- from always-loaded to on-demand	Saves ~4,600 tokens. Risk: slight latency on first tool use per session.
3	Remove Google Drive folder IDs and doc IDs from CLAUDE.md	Sterling would need to load these from a reference file when creating/managing docs	Saves ~2,200 tokens
4	Compress Responsiveness Rules from 80 to ~30 lines	Core behavioral rules -- Brad may want all 8 rules with full detail always loaded	Saves ~1,200 tokens
5	Strip root cause narratives from all protocols (keep rules only)	The narratives provide "why" context that may help Sterling internalize the rules	Saves ~2,500 tokens. Risk: Sterling may not understand WHY a rule exists without the story.
6	Remove/update Claude Max billing section after June 15 split	May still be relevant if the split doesn't work as expected	Saves ~400 tokens after verification
7	Remove Connected Apps one-line list (already in reference file)	Sterling would lose instant awareness of what tools are available without loading the reference file	Saves ~700 tokens. Risk: Sterling might claim "no access" more often if it doesn't see the app list.

Key tradeoff: Every token removed from CLAUDE.md extends session life and reduces cost, but also reduces what Sterling "knows" at conversation start. The safest approach: compress verbose sections first (narratives, examples), then move reference data (IDs, schemas), and only last remove behavioral rules. Start with the ~6,300 tokens from compression, which has zero functional risk.

Appendix: File Inventory

▼

Always-Loaded Context

File	Size	Lines
`CLAUDE.md`	108 KB	1,184

Reference Files (loaded on demand)

File	Size	Purpose
`reference/client-intelligence-architecture.md`	15 KB	HubSpot integration, file structure, knowledge architecture
`reference/connected-apps-detail.md`	18 KB	All API keys, endpoints, per-tool details
`reference/gchat-monitor-protocol.md`	9 KB	Full 8-scenario GChat scanning protocol
`reference/known-issues-history.md`	10 KB	Diagnostic history and fix procedures
`reference/operational-protocols.md`	8 KB	Calendar, email, Airwallex, shortcuts
`reference/social-media-protocols.md`	18 KB	Full social media posting rules

Skills (34 files)

Full index at /workspace/group/skills/index.md. Categories: Browser/Auth (5), Social/Content (7), QA/Process (4), Infrastructure (3), Content/Video (4), Messaging (2), Integration (3), Google/Drive (3), Email (1), Commerce (1), GChat (1).

Filing Cabinet

257 files total. 115 client intelligence files, 16 industry dossiers, 10 content catalogs, 4 voice profiles, 5 strategic docs, 6 social/Facebook docs, plus expert files, software knowledge, and relationship files. Full index at /workspace/group/filing-cabinet-index.md.