Writings ·
Model selection strategy
Which model to use for which kind of session based on the type of task, teaching value, and complexity
The one-line principle§
I’m using the agent as a teacher, not a bot. In teaching mode I write the code, so sessions are light on tokens and explanation quality matters more than cost. The moment a phase has no teaching value for me, I switch to auto-pilot (agent writes it) — and that’s where cheap/fast models earn their place.
Model roster§
| Model | Where | Cost (in / out per 1M) | Role |
|---|---|---|---|
| Composer 2.5 | Cursor (included pool) | ~free on plan ($0.5 / $2.5 metered) | Auto-pilot implementer for low-teaching-value work |
| GPT-5.6 Terra | Cursor (Other Models) | $2 / $12 | Daily teaching driver |
| GPT-5.6 Sol | Cursor (Other Models) | $4 / $20 | Hard 🔴 backend/security/RAG lessons |
| Opus 5 | Cursor (Other Models) | $5 / $25 | Final call on high-stakes architecture; deep debugging |
| Fable 5 | Cursor (Other Models) | $10 / $50 | Rare, phase-reshaping architecture decisions only |
| Gemini 3 Pro | Gemini app / AI Studio (FREE, outside Cursor) | free tier | Brainstorming + architecture debate + PRD drafting |
Core rules§
- Teaching sessions → Terra by default, Sol for the 🔴 tasks. A wrong explanation on security-critical work teaches me a bad mental model, so pay for reasoning where correctness compounds.
- Auto-pilot sessions → Composer 2.5. It’s a strong, near-free implementer. Use it only where I’ve judged there’s nothing to learn.
- Brainstorm & debate architecture in FREE Gemini, outside Cursor. These sessions are conversational, not in-editor — so they cost $0 and never touch the $20 pool. Then bring the plan to Opus for the final high-stakes call.
- Reserve Opus for the rare closer; Fable only for decisions I’ll live with for months.
- Privacy hard rule: never feed Core Drive or Alter Ego personal data to any free tier (free Gemini may train on inputs). Public-repo coding chat is fine; my principles/private affairs are not. Use the paid API or keep them out.
- Effort level: medium on Terra for routine teaching; bump to high on Sol for 🔴 concepts. Standard thinking on free Gemini for brainstorming; extended only when a design genuinely needs it (it burns the free quota faster).
Planning layer (before every phase)§
Phases 4, 5, 6, and 7 require a per-feature PRD first. Draft and debate that PRD in free Gemini 3 Pro (it holds the whole architecture doc + living plan in one context window), then have Opus 5 pressure-test the final design. Keeps all planning off the $20.
Phase-by-phase model map§
Legend — Teaching value: ⭐ low (I know this) · ⭐⭐ medium · ⭐⭐⭐ high (why I’m doing this project)
| Phase | Teaching value | Teaching session | Auto-pilot session | Notes |
|---|---|---|---|---|
| 0 — Foundation & tooling ✅ | ⭐⭐ | — | — | Done |
| 1 — Supabase & auth shell ✅ | ⭐⭐⭐ | — | — | Done |
| 1.1 — DB-level RLS ✅ | ⭐⭐⭐ | — | — | Done |
| 2 — API guard & feature system | ⭐⭐⭐ | Sol | — | Security backbone — teach every line; don’t auto-pilot |
| 3 — Graph store & metadata | ⭐⭐⭐ | Sol (recursive CTEs) → Opus for the design | — | New backend modeling; core learning |
| 4 — Taxila v1 | ⭐⭐⭐ backend / ⭐ palette UI | Terra, Sol for RLS+repo | Composer for the command-palette UI | Split the phase: learn the schema→migration→API→repo loop, auto-pilot the frontend |
| 5 — Zinsser v1 (Tiptap editor) | ⭐ | Terra (only if I want Tiptap internals) | Composer | Frontend — I already know this; auto-pilot most of it |
| 6 — Core Drive + AI layer + Alter Ego | ⭐⭐⭐ | Sol (pgvector/HNSW/chunking) → Opus for Context Assembler design | — | The RAG phase I’m here for — never auto-pilot. Privacy rule applies to Core Drive/Alter Ego data |
| 7 — War Room & Factory | ⭐⭐ | Terra | Composer once the pattern repeats | Established patterns; teach once, then auto-pilot |
| 8 — PWA & iOS shell | ⭐⭐ | Terra (service workers, Capacitor are new) | Composer for config/boilerplate | Mostly config |
| 9 — API keys & CLI | ⭐⭐⭐ security / ⭐⭐ CLI | Sol for key crypto, Terra for CLI | Composer for CLI scaffolding | Key generation/hashing is get-it-right-first-time |
| 10 — Observability & hardening | ⭐⭐ | Terra, Sol for the pen-test pass | Composer for logging boilerplate | Teach the security audit, auto-pilot the plumbing |
| 11 — Dogfooding (optional) | ⭐⭐ | Terra | Composer | Invite flow + admin on known patterns |
| 12 — Billing & Stripe (optional) | ⭐⭐⭐ if I care about Stripe | Sol (webhooks) / Terra | Composer for UI | Webhook state sync is the tricky bit |
| Post-MVP | exploratory | Gemini (free) to explore → Terra/Sol to build | Composer | Low stakes; experiment cheaply |
Quick decision flow§
- New phase? → Brainstorm PRD in free Gemini → finalize with Opus.
- Does this concept have teaching value for me?
- Yes, and it’s 🔴 → Sol (teach mode)
- Yes, routine → Terra (teach mode)
- No → Composer (auto-pilot)
- Stuck on a nasty bug across many files? → one Opus call.
- Architecture-defining fork I’ll live with for months? → one Fable call.
Budget reality check§
Because teaching sessions don’t burn agentic tokens and brainstorming lives in free Gemini, $20 Pro should hold up fine. Revisit Pro+ ($60 = $70 of third-party usage) only if auto-pilot phases start dominating and Composer isn’t enough.