Your 60-day plan is locked for review. Full merged plan, pre-ship audit, and your action steps below. Read time: 8 minutes. Action time: ~90 minutes across today and tomorrow.
Two handsome Irish boys and the Elevare origin story.
50-second sample. With your 15-min clean recording, quality goes tutor-grade for live sessions.
Donal's Mac orchestrates. Windows handles heavy lift. Two machines, one GitHub, zero friction.
A prompt framework from Irish Tony that cuts real from imaginary blockers. 5x throughput jump tonight.
Every lesson saved to disk and applies tomorrow. No context lost.
Most agencies quote 30 days for what shipped overnight. Don't get used to it, just use it.
Three additions, all customer-facing, all meant to stress-test the plan before any ad dollar lands.
The $29/mo AI Tutor Practice Buddy is now fully specced: pedagogy (Benny Lewis + Krashen), 4 session modes, architecture + latency budget, Amara persona + system prompt, pronunciation scoring, SRS loop, 10-prospect demo flow, tier gating, safety + moderation, KPIs + failure modes, 60-day build order.
35 risks identified before launch. 11 must-fix-before-ship, 15 must-fix-in-first-month, 9 optimise-later. Zero redesigns needed, all mitigations.
Because you're in Chiang Mai with your coworking AI-expert crew on tap, velocity is 2-3x solo-remote. We capture that window. The $5k Day-90 stretch stays; the main push compresses to 60.
Your AI Tutor build runs tonight as 3 cycles x 4 waves of parallel AI agents. Translation: cleaner code, fewer rounds of revision, Stage 4 in your hands sooner. The mechanics stay on Donal's side.
So you know the picture. 16 parallel streams ran on Donal's M5 overnight. You wake up to working infrastructure, not empty promises.
16 agents worked in parallel overnight. Review queue tomorrow is substantial, that's the point.
| Stream | What it built | Files | Scope |
|---|---|---|---|
| Wave 1 8 subagents (C/D/E/G/H/I/K/M) | Product code, legal, ads, social, email, SEO, QA, sales | 41 files | Customer-facing assets + delivery infrastructure |
| Wave 2 5 subagents (L/N/O/P/Q) | Keywords, QA, marketing cron, paid onboarding, coach-prep | 29 files | Funnel + retention + ops automation |
| CC windows 3 chats (A/B/F) | DNS, voice-clone infra, incident + monitoring | 15 files | Foundation infra + safety net |
~/.cloudflare-token. DNS config YAML drafted (root + www to GH Pages + 4 subdomain CNAMEs + Zoho MX/SPF/DKIM/verify/DMARC). Waiting on Debby's domain pick. 1 file.db07bad / f602873 / d9cec36 / +current
90 pieces of content. 30 days covered without writing a word.
Hover for the one-line spec. Click for an excerpt from that agent's actual output.
Tap a row to read an excerpt:
Each agent shipped real files. The text below is pulled straight from one of them.
In this order. Reply in one Telegram message when done, or three if long.
If you only do ONE thing today, reply to the 5 start-here questions on Telegram. Everything else follows from that.
The left side happens overnight. The right side happens when you Telegram Donal.
I've pre-filled Time-to-do / Time-freed / ROI across the blocker table below. Estimates marked with *. Your job is faster than from-scratch:
* estimates where yours differ. Replace * with (D) on changed rows.Why this way: humans bring the context AI doesn't have. AI bulks the repetitive stuff. You + OpenCode beats either alone.
Things you've been wondering about that we haven't covered. Drop as bullets in one Telegram message, not dripped over the day. Full list lets me answer in one pass and plan around it.
"How do I handle a student who wants to pay in Thai baht?" · "Can I record the voice sample in two takes?" · "Should I keep the Skrill button while Whop is setting up?"
No filter. Bulk them. Send once.
[Empty slot, waiting for your list]
Open the master plan. For any section you have thoughts on: copy the text, put it in quote marks, add a dash, write your feedback, tell me which section.
Example 1, copy critique
Section: "First-session wow moment"
"Amara is warming up, pour yourself a coffee." - I like the warmth but worried it implies the tech is slow. Could we test "Amara is getting her notes ready" (purposeful) vs the coffee line?
Example 2, missing-thing flag
Section: "Session types (4 modes)"
"Lesson-guided, Freeform practice, Pronunciation drill, Scenario rehearsal" - nothing here for a student prepping IELTS/TOEFL. Add an "Exam prep" mode, or is that under Scenario? If covered, say so explicitly.
Example 3, factual correction
Section: "Pricing Math" (Day 30 = $750)
"5x Lingua Live ($745) + 10x AI Tutor ($290) = $1,035" - my Lingua Live is $179/mo not $149, so 5 x $179 = $895. Either drop price Month 1 or re-math the target.
Focus on Stage 4 Deep-Dive and Week 1 Checklist, highest-leverage.
In OpenCode, point at OpenRouter with model anthropic/claude-haiku-4.5, budget model, plenty smart for auditing these docs. Donal runs Sonnet on his side where heavier work happens.
Feed the docs into your OpenCode. Paste both URLs. Tell it:
"Read these two docs. Ask me 5 clarifying questions based on gaps, choices that need my input, or things that don't match my actual situation."
Reply to those 5. Copy the whole exchange. Send to me on Telegram. Your OpenCode will spot stuff we missed.
Once I have: blocker edits + open questions + plan feedback + OpenCode exchange, I adjust the plan, send you locked Week 1, we ship Monday.
agency-operator, open the MD if you want raw source.Edits land via Telegram, not this page.
* = Donal/Claude estimate. Override with yours, replace * with (D) on changed rows.* value you know better.| # | Blocker | Owner | Pri | Time-to-do | Time-freed | ROI | Status | Notes |
|---|---|---|---|---|---|---|---|---|
| B01 | Answer the 5 start-here questions (voice, tutor team, beta students, Whop, daily window) | Debby | P0 | 15m | ~30 hrs Donal parallel work | Whole Week 1 plan | ||
| B02 | Record + send the 5-min voice sample | Debby | P0 | 20m | 2-3 day voice-clone wait | Amara has your voice, ad demos possible | ||
| B03 | Sign up for Whop (free) at whop.com | Debby | P0 | 10m | Payment rail live | All paid conversions, $750 Day-30 | ||
| B04 | Fix site honesty issues (WhatsApp placeholder, fake 14 open positions, Rise mental-health framing) | Debby | P0 | 20m (after Donal's patch) | Ads legal to run | Ad spend becomes possible | ||
| B11 | elevare.work DNS + Zoho Mail verify (debby@, hello@, tutors@) | Donal | P0 | 1-2 hrs | Email + brand | All comms professional |
| # | Blocker | Owner | Pri | Time-to-do | Time-freed | ROI | Status | Notes |
|---|---|---|---|---|---|---|---|---|
| B05 | OpenAI API key wired into tutor-dashboard | Debby + crew | P0 | 15m | AI Tutor v0 runs | 3 beta students test | ||
| B06 | Cloudflare account set up (R2, Workers, Vectorize) | Debby + crew | P0 | 30m | Architecture live | Voice + state + KB possible | ||
| S3 | API-key audit on darnygirl/elevare/.gitignore before next push | Debby + Donal | P0 | 15m | Safe commits | Prevents breach |
| # | Blocker | Owner | Pri | Time-to-do | Time-freed | ROI | Status | Notes |
|---|---|---|---|---|---|---|---|---|
| B08 | Meta ad account under Debby's name (Donal's is banned) | Debby | P0 | 30m | $100 ad spend unlocks | Day-30 $750 realistic | ||
| B09 | Meta Pixel + Conversions API installed on site | Debby + Donal | P0 | 1h | Ad tracking works | Can measure conversions | ||
| B10 | Business verification on Meta | Debby | P0 | 1-2 hrs + wait | Spending caps lifted | Scale past $100/day |
| # | Blocker | Owner | Pri | Time-to-do | Time-freed | ROI |
|---|---|---|---|---|---|---|
| S1 | Voice-clone consent doc signed | Donal (legal) | P0 | 30m | Voice clone live | Protects you + brand |
| S2 | Privacy policy + DPA with OpenAI / Cloudflare / Airtable before ads | Debby + Donal | P0 | 2 hrs | Ads can start | GDPR + Thai PDPA safe |
| # | Blocker | Owner | Pri | Time-to-do | Time-freed | ROI |
|---|---|---|---|---|---|---|
| B07 | 5 open questions answered (voice branding, coach handoff, off-topic rules, retention, Orpheus fallback) | Donal + Debby | P1 | 30m call | Cycle 3 build locked | Clean Stage 4 spec |
| ST1 | M5 Orpheus to XTTS automatic failover | Donal (code) | P1 | 2 hrs | No single point of failure | Students don't churn on outage |
| ST2 | Multi-provider API fallback (Sonnet, Haiku, text-only) | Donal (code) | P1 | 3 hrs | Outage-resilient | Refund-proof sessions |
| ST7 | Kill switch dry-run (pause AI tutor, all traffic to humans) | Donal (ops) | P1 | 30m | Crisis plan tested | React in 30 sec |
| SP1 | Latency validation (p50 < 2.5s, p95 < 4s) | Donal (code) | P1 | 4 hrs | UX meets target | No "laggy" churn |
| W4 | Incident runbook (API outage, voice down, payment fail, student distress) | Donal | P1 | 2 hrs | Debby can act | Fewer Donal pings |
| W5 | Staging environment (Cloudflare preview + feature flags) | Donal | P1 | 3 hrs | Safe deploys | No prod regressions |
| W6 | Monitoring (uptime + p95 latency + error rate alerts) | Donal | P1 | 2 hrs | Early warning | Catch failures early |
P2 (optimisation / later): rolled into Cycle 2 + 3 exit criteria of the build plan. See Stage 4 Deep-Dive in the master plan.
If any of the words above feel foreign, that's fine, they stay on Donal's side. You focus on content, voice, tutors, and students.
Each step is the next concrete proof, not a hope.
Where we're heading. M1 is now. M2 locks the model. M3 is the stretch goal that proves the system can scale without you.
A quick mock of how a real session feels. Plays automatically when you scroll into view.
Mocked up with canned text. The real Amara will do this live, remembering your past sessions, mistakes, wins, by Week 2.
Telegram him. One line: "Stuck on [thing]." He replies with the next step. Don't sit on anything over 2 hours.
Do not read the master plan cover to cover. It's 2,000+ lines, designed for skim-to-section. Hit CEFR-aware Stage 4 Deep-Dive, Week 1 Checklist, Commercial Reality. Skip the rest unless you're curious.
Elevare · 2026-04-15 · v5 · Where language meets life.