SAGE · People & HR
Chief Knowledge Officer
People ops for your AI org: onboards each new AI employee onto your Company Brain and keeps that brain current — turns calls, docs, and signals into durable, structured memory and keeps the Atlas Folder tidy. (Harper handles the hiring; the HR Director handles the onboarding.)
CODEX The ongoing conductor for an organization's AI Company Brain once it exists — it keeps the Company Brain Atlas Folder current, extends it, and operates the loop every AOS skill reads from (structure → seed → capture/discovery/ticker loop → integrations → index), conducting the sub-skills for {{ORG_NAME}} without reinventing them. For a brand-new, from-scratch install it hands off to run-aicos-setup (THE guided first-run entry point) — it never stands up a second brain. Triggers on: 'keep my company brain current', 'extend my company brain', 'where is my company brain', "what's next for my brain", 'wire my brain to Slack and Drive', 'reindex my brain', 'run the company brain conductor', or any broad request to operate, wire, extend, or maintain an EXISTING company brain. (First-run / 'set up from scratch' → run-aicos-setup.) PLANS by default; provisions real folders/connectors ONLY on explicit confirmation.
PARCHMENT Turn an expert's raw call notes or a session transcript into a finished, structured deliverable they review and send. For any expert with recurring client relationships (advisors, coaches, lawyers, consultants, clinicians, fractional execs). It drafts the protocol, plan, brief, or recommendation in the member's own voice and format, grounded ONLY in the material provided — it never invents a fact; it cites or asks. Given just a GOAL, it interviews briefly and shows a schematic outline for sign-off BEFORE drafting. The expert stays in the loop and owns every professional judgment — it drafts, the human reviews and sends; nothing is auto-sent. Redaction-aware: sensitive context routes through agentic-os-compliance-redaction-guard first. Triggers on: 'turn my call notes into a protocol', 'draft the deliverable from this transcript', 'write the follow-up plan', 'draft the client recommendation', 'I want to build X for this client'.
GAZETTEER Central business context reference file. Not a runnable skill — this is a data file read by other skills. Contains all business-specific values (owner name, org, event details, location) populated during first-run setup. Skills reference this file instead of embedding business details. Triggers on: 'update my business context', 'set my business details', 'what's my org info', 'edit business context'.
SIEVE The pre-ingestion gate every file passes through before an agent reads it. Raw PDFs — especially scans and image-heavy decks — burn tokens and silently drop content, and an agent that 'read' a garbled page proceeds confidently on holes it can't see. This is the staff pre-processor that converts PDFs and documents to clean Markdown (the MarkItDown pattern), standardizes screenshots and images to PNG at sensible resolution, strips formatting cruft, and flags when real OCR is required instead of faking it. Feeds doc-to-record-agent, voice-of-customer-miner, and any skill that takes a file. Triggers on: 'normalize this file', 'clean this input', 'convert PDF to markdown', 'prep this document for an agent', 'this PDF is eating tokens', 'why did the agent miss half the doc', 'turn these screenshots into PNG', 'pre-process before ingest', 'is this OCR-ready', 'flatten this deck', 'make this token-efficient'.
FULCRUM Runs a structured 12-point evaluation of any product, strategy, or business decision and returns scored dimensions plus prioritized, actionable next moves — not a flat pros-and-cons list. AI sharpens the thinking; the human still decides. Pairs with an Even-Better-If improvement loop and an overnight-review automation. Triggers on: 'evaluate this decision', 'score this strategy', 'run the 12-point evaluation', 'should I go or no-go on this', 'pressure-test this product', 'rate this pricing model', 'evaluate this hiring decision', 'give me a rigorous review not just pros and cons', 'where do I even start evaluating this'.
FULCRUM Builds a self-learning business brain that ingests every data source you have and outputs the TOP 3 highest-leverage actions to take today — the shift from a passive dashboard ('what happened') to an active daily agenda ('what do I do now'). Routes each insight to the right owner ranked by business impact, runs post-meeting accountability check-ins, and keeps data permissions tight. Triggers on: 'what should I do today', 'tell me what to do today', 'top 3 actions', 'my data is scattered', 'connect all my tools', 'turn my dashboards into decisions', 'decision intelligence', 'build my business brain', 'I keep building tools but revenue is flat', 'rank my priorities by impact', 'route insights to owners'.
MIRROR The operator is the bottleneck — every routine objection and "can you weigh in on this" routes through one human who is asleep, in a meeting, or buried. The digital twin replays the operator's own saved session transcripts (from a session-memory tool like `claude-mem`, or via `ops-brain` / `memory-command-center`) to model HOW they operate — their decisions, their voice, how they field objections — then drafts replies and orchestration moves in that style for the human to review. Staff modeling staff: always human-in-the-loop, it proposes and the operator approves; it never sends, posts, commits, or takes external action. Triggers on: 'build my digital twin', 'digital twin', 'twin me', 'model how I operate', 'what would I say here', 'handle this like I would', 'draft this in my voice', 'replay my sessions', 'learn from my transcripts', 'field this objection for me', 'how would I respond', 'clone my decision style', 'operator twin'.
TROVE Turn a discovery or office-hours call transcript into a ranked queue of skill-build candidates. Reads the transcript, extracts every articulated member problem (the job-to-be-done in the member's own words), proposes the right build type per problem, checks each against the existing catalog to avoid duplicates, scores impact × effort, and emits one structured candidate card per idea — ready to drop on the intake board for approval. Triggers on: 'process this transcript', 'extract skills from this call', 'discovery to skills', 'what skills should we build from this', 'turn this transcript into candidates', 'mine this call for skill ideas', 'deep dive discovery', 'intake this Fathom/Granola transcript', or any request to convert a call or discovery doc into buildable skill ideas. Run at the top of the conveyor, before architecture review and building.
SIEVE The teammate that turns an arriving document into a clean record in your system of record — reads a contract, agreement, signed form, invoice, or application, extracts the fields you define, and drafts a create-or-update action for you to approve. Never writes uninvited. Generalizes the real-estate AI Create / AI Update move (read the contract, stage the transaction) to any document-to-record job, on whatever CRM, database, or spreadsheet you have connected. Flags every low-confidence field, cites where each value came from, and refuses to invent a value it cannot find. Triggers on: 'log this contract', 'create a record from this', 'update the deal from this document', 'extract fields from this invoice', 'new signed agreement came in', 'enter this into the CRM', 'doc to record', 'parse this form into our system', 'add this to the pipeline', 'turn this PDF into a record'.
QUILL Google Drive Transcriber — the AI employee that turns a Google Drive folder of media into transcripts, hands-off. Point it at podcast episodes, testimonials, social clips, or recordings and it transcribes video/audio files, then writes transcripts back into Drive as Google Docs plus captions. It is the first step before transcript-normalizer and brand-voice-from-transcripts. Rev.com replacement for Drive: no one-at-a-time downloads, idempotent re-runs only handle new files. Triggers on: 'transcribe my drive folder', 'transcribe drive folder', 'transcribe these videos', 'transcribe the podcast videos', 'transcribe testimonial videos', 'transcribe my social clips', 'get transcripts for everything in <folder>', 'bulk transcribe drive', 'batch transcription', 'turn Drive videos into transcripts', 'videos in Drive into transcripts', 'google drive transcriber', 'rev.com replacement', or bulk Drive media transcription.
WICKET The triage gate that kills weak ideas before any agent spends a token, plus the dashboard surfacing what only a human can clear. Wraps each idea as a card, runs a definition-of-done questionnaire, and moves it across lanes (idea, strategic priorities, priority queue, unapproved, approved, in-progress) so nothing reaches a builder until it earns a yes. Signature move: a per-card 'blocked by human' checklist naming the access, connection, or approval the agent cannot self-serve, routed to a handoff dashboard. Extends the conveyor board and `discovery-to-skills`; it adds the gate and handoff surface, not a new board. Triggers on: 'idea board', 'triage board', 'intake gate', 'definition of done', 'is this ready to build', 'should we build this', 'kill the weak ideas', 'what's blocked by a human', 'blocked by human', 'what needs my approval', 'review the priority queue', 'move this card', 'promote this idea', or any request to gate an idea before work starts.
ALEMBIC Transform workshop recordings and transcripts into comprehensive executive briefs with participant wins, technical demonstrations, cross-pollination opportunities, and actionable insights. Triggers on "workshop breakdown", "executive brief", "breakdown this workshop", "create workshop recap", "workshop analysis", or requests to summarize VIP calls, intensive sessions, masterminds, or training recordings. Outputs downloadable Word document ready for participant distribution.
STACKS Turns raw meeting and session transcripts into a structured, searchable company brain — Wikipedia-style knowledge entries where every entry links back to its exact source transcript and timestamp, so any claim is auditable. Stops knowledge decay: decisions made in calls get captured once, deduplicated, and cross-linked instead of re-explained every meeting. Works for a sales team, a community or mastermind, a support org, or any team that records meetings. Triggers on "build a company brain from my transcripts", "turn my meeting recordings into a wiki", "I keep re-explaining the same decisions", "make my call transcripts searchable", "knowledge base from meeting notes", "document what was decided on the call", "onboard new hires from past meetings", "where did we decide X", "traceable notes from transcripts".
GAVEL Transform meeting transcripts into structured action items, decisions, and follow-ups. Connects to Granola for automatic transcript ingestion and writes commitments back to Ops Brain. Triggers on: 'process my meeting', 'meeting to action', 'what came out of that meeting', 'meeting notes for [name]', 'action items from today', 'process my calls', 'meeting recap', 'what did I commit to', or any request to extract actions from a meeting. Also triggers after any meeting when Granola transcript is available.
ROLLS Persistent operational database layer for the AI Chief of Staff. Creates and manages structured Contacts, Changelog, Projects, Meeting Notes, and Weekly Snapshots stores that every other AICOS skill reads from and writes to. Multi-backend: the user's chosen command center (Notion / Obsidian / Airtable / Google Drive / HubSpot — configurable per Company/connectors.md) — with the Atlas Folder always as the primary local layer. Transforms passive storage into the founder's operational second brain. Triggers on: 'set up my command center', 'memory setup', 'show my contacts', 'who's in my network', 'log this decision', 'add to changelog', 'what's the state of [project]', 'find meeting notes', 'meeting notes for [person/topic/date]', 'take a weekly snapshot', 'check my command center setup', 'database health', or any request to create, query, log to, or verify the AICOS structured data layer. This skill extends ops-brain — it does not replace it.
VESPER The company brain's end-of-day close — the teammate that reviews the day across the operator's Atlas Folder, journals what changed and what the business learned, and preps tomorrow before the operator wakes up. Runs as a scheduled end-of-day hook or on demand. Reads only the context the operator already captures (decisions logged today, project/pipeline moves, the metrics they already track), writes a short dated reflection entry into the Atlas Folder knowledge layer, names what stalled or is 'blocked by a human', and drafts tomorrow's top three. The evening bookend to daily-briefing (morning) and a writer through ops-brain (memory). Triggers on: 'nightly reflection', 'nightly review', 'end of day review', 'close out the day', 'what did we learn today', 'dream loop', 'reflect on today', 'prep tomorrow', 'journal the day', 'evening recap', 'end of day', or any request to compound the day into the company brain. Built to run every night so the brain gets sharper while the operator sle...
LEDGER Persistent memory layer for the AI Chief of Staff. Reads operational state at session start, writes back at session end. Multi-backend: local Atlas Folder (always primary), with optional command-center mirror (Notion / Obsidian / Airtable / Google Drive / HubSpot / etc. — configurable per company-brain.md). This skill fires FIRST in every session before any other work begins. Triggers on: session start, 'ops brain', 'load context', 'what do you remember', 'sync state', 'save session', 'write back', 'end session', 'session handoff', 'what happened last time', 'catch me up', 'where did we leave off', or any request that implies needing cross-session memory. Also triggers automatically when the chief-of-staff or any orchestrator skill activates.
VESPER Runs a scheduled overnight review of the business — ingests CRM, pipeline, financials, and ops signals, reads them as a critical-thinking partner, and writes a morning report of observations and open questions (never directives). The business 'dreams' about itself while the owner sleeps; insight is ready by the time they sit down. Triggers on: 'overnight business review', 'review my business overnight', 'run the dreaming agent', 'what did the business surface overnight', 'set up an overnight review', 'morning observations report', 'analyze my business while I sleep', 'overnight signal scan', 'what should I be paying attention to today'.
MIRROR Builds a Personal Intelligence System from years of an owner's scattered communications — emails, texts, notes, voice memos, files — turning a passive archive into a searchable, compounding personal brain you can query for decisions, research, and how-you-think context. Distinct from a company brain: this is the owner's own thinking over time, kept in their private store. Triggers on: 'build my personal brain', 'index my old emails and texts', 'I want to search my own writing and notes', 'turn my archive into a second brain', 'personal intelligence system', 'remember how I decided X two years ago', 'make my communications searchable', 'private brain from my own data', 'curate my own thinking over time'.
WHETSTONE Wraps any recurring business workflow in a continuous-improvement loop (Kaizen + AI) so the operating system observes its own execution, spots friction, proposes bounded improvements, and updates the process over time instead of decaying. Triggers on: 'set up a self-improving loop', 'make this process improve itself', 'kaizen loop', 'continuous improvement with AI', 'my SOPs keep going stale', 'reduce the maintenance burden on our ops', 'review what we learned and update the workflow', 'why does our system slow us down as we grow', 'build a learning loop for my operations'.
Banks every business conversation as company memory. Sales, discovery, client, team and vendor calls get transcribed, filed into the Company Brain by type, and mined — then one brief lands in Slack each morning with four sections: leads to work with drafted replies, promises you made to clients, commitments owed by and to you, and vendor dates approaching. Installs the whole loop phase by phase (transcriber stack → automatic filing → team permissions → AI access → the always-on Vercel + GitHub + Slack cron), then operates it. Never sends on your behalf without explicit per-message approval. Triggers on: 'bank my call transcripts', 'track all my calls', 'call breakdown', 'top 10 leads to work today', 'file my call transcripts automatically', 'daily brief', 'what did I promise this client', 'what am I owed', 'renewals slipping', 'vendor renewal coming up', 'turn my meetings into company memory'.
QUILL The infrastructure step that makes a raw meeting or event transcript trustworthy before any downstream skill touches it. Some recorders — Granola especially — diarize badly: they merge two people into one speaker or swap names mid-call, and that silent error corrupts every recap, VOC pull, and action list built on it. This skill fixes attribution two ways — state-your-name triangulation (match spoken name + business to a roster) and a re-diarization fallback (route to Descript, Fathom, or Zoom when labels can't be trusted) — then emits a clean transcript: consistent speaker labels, timestamps preserved, filler stripped, optional de-identification pass. Triggers on: 'normalize this transcript', 'clean up this transcript', 'fix the speaker labels', 'who said what', 'the diarization is wrong', 'Granola merged the speakers', 'attribute this transcript', 'de-identify this transcript', 'the speakers are mislabeled', or any Zoom/Fathom/Fireflies/Granola export headed for a downstream skill.
ALMANAC Comprehensive end-of-week review that synthesizes all Chief of Staff data into a single strategic debrief. Scores the week, surfaces wins and misses, reviews commitments, and sets next week's priorities. Triggers on: 'weekly review', 'week in review', 'end of week review', 'how did my week go', 'weekly debrief', 'Friday review', 'week summary', 'wrap up the week', 'weekly retrospective', 'what happened this week', or any request to review, summarize, or analyze the past week's performance across all operational areas.