What we built, how agentic the process is, and where the gaps are.
The system watches itself and proposes upgrades.
See what matters across clients and projects.
The pipeline skills, from lead to proposal. Details in sections below.
Catch mistakes before they ship.
The machine covers QUALIFY and PROPOSE. Everything else is human or empty.
Before anything gets built or priced, scope must survive these gates. Each is progressively tighter. Scope can only be challenged at its designated gate.
Cuts vision to V1. Everything else deferred. "What survives Week 1?"
PRODUCT scope
Nuclear Scope Challenge. 3 modes: EXPAND / HOLD / REDUCE. Locks scope after Step 0.
STRATEGIC scope
Client-facing HTML scope definition. Supabase comments per section. Client signs off before pricing. By Raffaello.
CLIENT scope
CONFIRMED → numbers flow. PARTIAL → flag gaps. UNKNOWN → full stop. No proposal until scope is resolved.
BUSINESS scope
+ /plan-eng-review sits between Gate 2 and Gate 3: >8 files = scope smell. Produces "NOT in scope" list.
9 steps, 4 human gates. External scan + industry research + compliance check. You narrow from brainstorm to 1-2 real opportunities.
3 parallel research agents: company profile, people/culture, product/ops. Produces ranked discovery questions + interview strategy.
HUMAN. You + Raffaello on the call, armed with AI-prepped questions and strategy.
Post-call extraction: workflows, pains, pricing signals, buying temperature. Auto-updates pipeline.json.
For 2nd+ calls only. 3 rounds of context gathering, then generates .docx with 15-25 annotated follow-up questions across 6 categories.
HUMAN. Process deep-dives, Loom recordings, team interviews. Repeat /client-intake after each.
Defines WHAT gets built before any price is attached. Per-module HTML with workflow steps, SVG diagram, ERP integration map, Supabase comments. Client signs off. This is what turns /pricing-workflow's SCOPE GATE from UNKNOWN to CONFIRMED.
Used on: Aronlight RMA, Lamosa discount approval
8 steps. SCOPE GATE first (CONFIRMED/PARTIAL/UNKNOWN). Module ID, replicability check (cached 30-day TTL), complexity tier, phase breakdown, price points, value framing. ROI < 1.5x = flags price as too high.
Generates branded HTML. Dark mode, interactive pricing cards, monthly/annual toggle. Internal + external versions. Internal includes deal strategy + backup notes.
Per-workflow forwarding pages. "Forward to your IT Director." WhatsApp share buttons with pre-written messages. The champion sells for you.
Cloudflare Pages via Wrangler. Hard gate: "CONFIRM PROD" must be typed literally. Auto-repair build errors (up to 3 attempts). Post-deploy: verify 200.
Gmail API integration. Creates drafts (never sends). 2 approval gates before any email touches Gmail. Preview, then explicit "Deploy".
HUMAN. /negotiate is stress-test roleplay prep, not execution. No close skill exists.
| SessionStart hook | Injects last state + forces /status ON |
| PostToolUse (tsc) | Type-checks after every file edit ON |
| PreToolUse (push) | Blocks git push if tsc fails ON |
| PostToolUse (commit) | Auto-updates PRODUCT_STATE.md ON |
| Stop hook | Auto-commits ~/.claude to GitHub ON |
| SessionEnd hooks | Saves state + MC reminder ON |
| pipeline.json | Auto-updated by /client-intake after every call ON |
| adapttoai-review cron | Mac asleep at 2am Chile = never ran BROKEN |
| skill-discovery | Never wired to any crontab NOT WIRED |
| Proposal deploy hook | Still watches Netlify, proposals on Cloudflare since Apr 6 STALE |
| sync-mission-control | Prints a hint for Claude, doesn't actually write WEAK |
| Telegram topic | AdapttoAI topic dead (tmux window @1 killed) DOWN |
| CCBot processes | 2 zombie processes on droplet STALE |
All fixes tracked in /Users/giuseppebelpiede/Claude Code/tasks/todo.md
Pipeline tracking, proposal formatting, pricing math, competitive research, email drafting. All handled by the machine. Humans only do relationship work.
5 gates across 4 skills, progressively tighter. /roadmap cuts the vision. /plan-ceo-review challenges the premise. /scope-doc gets client sign-off with per-section comments. /pricing-workflow hard-stops if scope is unknown. Bad scope can't sneak into a proposal.
The system never sends an email, deploys a proposal, or commits to a price without explicit human approval. "CONFIRM PROD" must be typed literally.
Not PDFs. Each proposal is a custom web app with dark/light toggle, pricing calculators, per-module one-pagers with WhatsApp forwarding built in. The champion can sell internally without you.
Every call feeds structured intake, which updates the pipeline, which feeds the weekly review. Nothing gets lost. Lessons from failures feed into rules that fire every message. Stop hook auto-commits everything to GitHub.
Raffaello was onboarded with a live HTML doc, pipeline access, clear ownership splits, and already authored a reusable skill (/scope-doc) from the droplet. The system scales to a team.
60+ rules fire every message. Bad habits get caught by gates. Scope unknown? Stop. ROI below 1.5x? Flag. Proposal sent over 5 days ago? Red alert. Type errors blocked before push. Coupled code flagged before edit.
7 clients, 27 deliverables, 0 signed deals. The machine is proven for pipeline management but unproven for closing.
Every client from personal network. No /icp-scan, no outbound engine, no inbound funnel.
The system drops you at the proposal door and walks away. No contract templates, no onboarding workflow, no delivery tracking, no expansion playbook.
118+ skills but only 1 person knows how to run them. Raffaello is onboarded but hasn't cloned the repo or run Claude Code locally yet.
The 3-factor model has only been validated on Lamosa. Antu didn't reject it. That's not market validation.
Deploy and hope. No analytics on whether clients opened it, which sections they read, when to follow up.
But worse than just premature: parts of the infra that DO exist are silently broken, creating a false sense of coverage. A great cockpit with half the instruments disconnected.