Every framework term, A→Z. If you saw it in the wizard and weren't sure what it meant, it's defined here.
Gitignored folder in the vault for testing the agent against PHI/PII without committing it.
Never leaves the laptop. Useful for 'what would the agent do with Mrs. Sharma's record?' tests. The .gitignore guarantee is critical for regulated industries.
Hospital management system on Supabase + Next.js — the live data layer for Hope Hospital's pharmacy reorder agent.
Adamrit is the operational HMS the agent reads from (medication catalog, suppliers, GRN deliveries, purchase orders). The framework treats it as the 'live data brain'; the vault holds the rules, Adamrit holds the data. They're combined only at agent runtime.
Append-only table that records every agent decision with version + rationale. DPO-readable.
Mandatory for every shipped agent. Captures who/what/when/why. Becomes the source of truth for closed-loop quality reviews and compliance audits. Hashes content rather than storing PII.
1-2 curious (not necessarily technical) people per department who run daily Claude practice for 2 weeks.
Champions don't have to be engineers. They have to be curious and willing to look dumb in front of an LLM for an hour a day. Their Capability Map at end of Week 2 drives the agent priorities for Step 7.
Replaces the C-suite reporting layer. Sets vision, reads the daily intelligence brief, makes capital allocation decisions.
One of the three new roles that emerge when middle management gets replaced by the intelligence layer. AI Founders don't synthesize reports — they consume the AI-generated synthesis and decide.
Tasks completed per week ÷ hours of human input. The core metric.
Target trajectory: 2× → 5× → 10× over four quarters. If it's flat, something's broken in your closed loop. Tracked on the /dashboard/kpis page.
Owns the Business Brain, the agent library, and the quality bar.
Replaces the Operations Manager / Chief of Staff role in an AI-First org. Reviews agent quality scores monthly, retires under-performing agents, prioritises the next 10 processes.
All company knowledge (SOPs, pricing, processes, escalations) in a format AI agents can query.
Markdown with frontmatter, version-controlled in Git, indexed in Supabase. Editable in plain English by anyone — the AI inherits the rule on next invocation. Distinct from documentation: docs are for humans, the Business Brain is for both humans and machines.
Two-column doc by end of Week 2: what AI handles well vs what AI needs a human for.
The deliverable from Step 4. Drives which workflow becomes Step 7's first agent. Without it, agent priority becomes guess-work.
Workflow that monitors its own output, scores it, and feeds insights back to improve next time.
Opposite of open-loop (run once, no feedback). Every agent in the framework is closed-loop: scoring checklist → if below threshold, loop back; if above, ship. Test harnesses + AI Retros are the human-side closed loops.
1-2 page brief the AI reads before EVERY task — org chart, pricing tiers, top-10 client questions, escalation rules.
The single most important file in your Business Brain. Lives in `corpus/`. Version-controlled, leadership-approved, updated quarterly.
Per-record boolean flag, default OFF. No agent runs on a record until staff explicitly consents.
Critical for regulated industries. Implemented as a column on the relevant table (e.g., patients.consent_for_ai). Without consent the agent silently does nothing — no error, no exception, just no action.
Directly Responsible Individual — owns ONE outcome end-to-end.
Replaces the Department Lead role. DRIs review agent output, approve drafts, escalate when needed. Authority to ship; accountability for the result.
First 2 weeks of every new agent — Slack alerts fire, but no real writes (POs / emails / actions).
Lets the human reviewer score every output against the scoring checklist before anything ships. Flip DRY_RUN=false only after a clean stretch where ≥90% of dry-runs would have been approved unchanged.
Deno-based serverless function on Supabase. Where agents live.
Auto-injected SUPABASE_URL + SERVICE_ROLE_KEY mean no cross-project key juggling when the function runs in the same Supabase as the data. Triggered by cron, webhook, or manual invoke.
YAML header at the top of every Markdown file in the vault: department, owner, status, last_reviewed, tags.
The GitHub Action's sync skips files without `department`. `status: retired` removes the file's chunks from the vector index on next sync. Keep these honest — agents inherit them.
Individual Contributor — builds prototypes, drafts, first cuts. Paired with an AI agent.
Replaces the Junior staffer role. ICs do the actual building; agents accelerate them 5-10×. Promotion path is now horizontal (more agents to pair with) rather than vertical (managing more people).
AI-generated daily brief that replaces middle-management reporting (status emails, weekly summaries, daily standups).
Goes live in Step 9. Reads every agent's audit log nightly, formats as Markdown, posts at 8am. Leadership reads it; nobody schedules a meeting to discuss it.
Bettroi's open-source skill harvester. Captures every Claude Code session into reusable Markdown skills.
github.com/chatgptnotes/ironbark. Auto-syncs with the community repo every 30 min. Becomes the seed for your AI Skills Library by Step 9.
Two independent ways to stop all AI activity instantly: Vercel feature flag + Supabase API key removal.
Either one alone stops everything. Independence is the point — no coordination needed in an emergency. Required for regulated industries.
macOS-native scheduled task. Backup for the Obsidian Git plugin auto-sync.
Runs every 10 min even when Obsidian is closed. Pulls/commits/pushes the vault. Pings Slack on push failure. The plist sits at ~/Library/LaunchAgents/.
Obsidian community plugin (v2.38.2 used) that auto-commits and pushes the vault to GitHub every 10 minutes.
The first step of the edit-to-agent pipeline. Configured to commit + pull + push at 10-min intervals. Frontend for non-technical users who don't want to touch the terminal.
5-10 checks every agent output must pass BEFORE a human reviews it.
The agent's 'definition of done'. Below threshold = loops back automatically. Above threshold = ships to human reviewer. Each agent has its own checklist in `agents/<dept>/<name>/scoring.md`.
Predefined criteria the AI uses to self-check its output before a human reviews it.
The machine-side closed loop. Pure functions exported from each agent's index.ts that you can run against fixtures in CI. Catches regressions before they hit Slack.
Scale output by increasing tokens-per-dollar (cheaper inference) and tokens-per-task (better prompts), not by adding headcount.
The unit economics shift. Without it, you're just adding AI tools to a traditional cost structure — same headcount, more SaaS bills. With it, marginal output goes up while marginal cost stays flat. claude-code-deepseek-backend cuts inference cost 60-80%.
The Obsidian folder of Markdown files that holds your Business Brain.
Lives at ~/<vault-path>. Synced to a private GitHub repo by the Obsidian Git plugin. Indexed into the agent-corpus Supabase Storage bucket by GitHub Actions on every push. The single source of truth for company knowledge.