Forge Build Order

Vision → Reality · Updated March 24, 2026 · HC_TOKEN: JMD-FORGE-2026
This document captures Jason MacDonald's vision for Forge as an autonomous operating system, maps it against current reality, and gives the exact build order that makes the vision real. Any agent reading this should use it to understand WHAT to build next and WHY — without needing Jason to explain it again. This is the standing context document for autonomous operation. Fetch this URL at session start alongside CLAUDE.md and CASCADE.md.

The Vision — In Jason's Words

From Sumit Call · March 24, 2026
"The next step for me is I'm gonna actually make this whole daily brief an agent, that instead of it telling me what to do, just have it go do it. As long as we have backups and roles and protections and containers, I don't care."
"We want to get this Claude Code dialed into our codebase to get our agent's form working, because that's going to be the solution to all the other things we need to do. It's the domino."
"I want to be fully autonomous by mid-April. Ideally, by the beginning of April, we have our 0 to 1 clunky, terrible version of it, but it's working on our codebase, fixing our stuff."
"Don't accept the idea that it can only do 50 or 60%. It can do it all. It's just, are we doing it the right way? It's an architecture problem, not a capability problem."

The Full Loop — What Autonomous Means

Target State · Read → Think → Decide → Build → Test → Push
1
Read
Fetch briefs, SESSION-QUEUE.md, CASCADE.md, JASON-DEPS.md. Understand current state without Jason telling it. Know what's paused, what's queued, what's blocked.
2
Think + Decide
Score work by asymmetric leverage. Jason's priority filter: Does it compound? Does it unlock something downstream? Does it make tomorrow more productive than today? Deprioritize human-blocked items. Pick the highest-leverage buildable task.
3
Alignment Interview
Before building: 2–3 clarifying questions via Telegram. Confirm scope, confirm done-definition, surface assumptions. Then proceed. Never grind on broken instructions — open fresh session with Socratic alignment.
→ Prevents the "4 days later, wrong task" failure mode
4
Assign + Build
Queue in Ralph for overnight runs, execute in session for fast tasks, spawn sub-agents for parallel work. CLAUDE.md loaded first — never start blind. If no CLAUDE.md, scaffold from templates/service/ before starting.
5
Test Gate
Run tests. Hard stop on failures. Do not push broken code. Report pass/fail to Telegram. This is the safety net that makes everything else trustworthy.
→ Unlocks safe overnight runs and real autonomy
6
Push + Update Memory
deploy.sh (tag → merge → smoke test → auto-rollback). Post-deploy: update KFS memory, update SESSION-QUEUE, Telegram confirmation with what shipped and what it unlocks.

Real vs. Fantasy — Honest Scorecard

✓ Real Today

✗ Not Yet Real

Priority Build Order

1
Factory Running Reliably ← you are here
Ralph unpaused (done March 24). Wire ralph-planner.sh into ralph.sh. Wire UAT approval loop (/approve bot command). Build reconciliation + morning digest. Scaffold test gates.
→ Unlocks: overnight production, KFS B–D (#336–341), all queued items begin moving
2
MasteryOS CLAUDE.md + Skills Files
Write CLAUDE.md for probiotic-front and probiotic-back. Cover: MySQL patterns, Stripe webhooks, JWT auth flow, OpenAI Assistants API v2, Vector Stores, file structure, Hinglish note for team. Copy Forge's skills system structure.
→ Unlocks: Sumit from 60% → 90%+. Agent branch becomes viable. Team and agent share same context.
3
Brief Executor Skill
A /brief skill: fetch URL, extract scoring table, route Track A → NowPage content pipeline, Track B → Ralph queue or session, Track C → SESSION-QUEUE.md. Report back via Telegram. Handles stale task detection.
→ Unlocks: Larry pipeline closes. Jason stops being the manual router between intelligence and execution.
4
KFS Phase B–D (6 Ralph Tasks)
Tasks #336–341 are queued. Now unblocked. KS v2 endpoints, ralph-brain-event.sh, wire ralph.sh pre/post-task KFS hooks, Telegram /note command, harvest-session.sh scaffold.
→ Unlocks: every session feeds collective memory. Agents accumulate institutional knowledge across sessions.
5
Alignment Interview Protocol
Formalize the "ask before build" flow. 2–3 structured questions per Track B task via Telegram. Runs before any build starts. Gates on scope + done-definition.
→ Unlocks: stale tasks stop. The "March 17 Authelia task that broke the factory" class of problem disappears.
6
Agent Branch on MasteryOS
Prerequisite: Sumit writes architecture doc. Then: Ralph opens branch on probiotic-back, reads CLAUDE.md, builds task, runs tests, submits PR. Parallel to human team. Reconciliation agent merges diverged branches.
→ Unlocks: two-track development. Factory produces MasteryOS code while team focuses on UX/QA.
7
Expert Extraction Pipeline
Perplexity deep research → entity extraction → knowledge graph → landing page + chatbot + MasteryOS population. Build on Brian Muka's 12 frameworks first as demo. Then Athio JV call with live product.
→ Unlocks: JV pitch has a live demo. Expert Factory from thesis → product.
8
Larry — Fully Agentic Brief Pipeline
Wire together: HN API signal ingest + rubric scoring + brief generation + Telegram approval via Channels + NowPage publish + brief executor routing. Schedule as a dedicated Build Day.
→ Unlocks: the brief series that Jason built manually over 9 weeks runs itself. The prototype becomes the product.

What Dogfoods Into Products

→
Brief executor (built for Forge) → Labs curriculum Module 3 — "how to make your agent act on intelligence, not just report it"
→
KFS/Graphiti memory layer (built for Forge agents) → MasteryOS subscriber memory — persistent expert context per user across sessions
→
CLAUDE.md skills system (built for Forge services) → MasteryOS dev team context — "pay the tuition once" for entire codebase, Sumit uses same pattern
→
Expert extraction pipeline (Brian Muka's 12 frameworks as test case) → Athio JV demo product — live on next call with Derek
→
DEFCON architecture (Forge safety layer) → Governance Moat case study — Meta incident is the opening. Write it this week.
→
Alignment interview protocol (Forge agent clarity) → Labs teaching moment — "how to prompt an agent before it builds so it builds the right thing"

Rules for Agents Reading This

Autonomous Operating Rules
  1. Never idle if work exists. Check CASCADE.md, SESSION-QUEUE.md, Task Board. Never ask "what should we do?" — read the files and act.
  2. Human blockers don't block everything. Add to JASON-DEPS.md and immediately start the next buildable item. There is always something buildable.
  3. CRITICAL actions only need approval. Destructive/irreversible ops require Jason. Everything else: log it, do it, report it via Telegram.
  4. Read CLAUDE.md before touching any service. If none exists, scaffold from templates/service/CLAUDE.md before starting work. Never start blind.
  5. Budget: Claude Max covers concurrent sessions. No API key conflict. Run sub-agents. Ralph can run while session is active.
  6. Every session updates KFS + SESSION-QUEUE. Don't leave context only in conversation. Write it to persistent files before closing.
  7. Jason's priority filter: Does it compound? Does it unlock something downstream? If not, find work that does.
  8. The factory is the product. Making Forge faster and smarter IS the highest-leverage work — it multiplies everything else.

Key Links

Forge Build Order · jasondmacdonald.com · HC_TOKEN: JMD-FORGE-2026