Prompt Templates & Checklists
Copy-paste scaffolds for the highest-leverage moves, plus checklists you can run before, during, and in production.
Cache-friendly system scaffold
Use when: Any repeated or multi-turn API workload.
[STABLE — cache this prefix, keep byte-identical]
Role: <assistant role>
Policies: <rules, tone, refusals>
Tools: <tool/function definitions>
Reference: <large doc / schema / examples>
[VOLATILE — changes per request, place last]
User request: {{input}}Model-routing decision
Use when: Choosing the cheapest model that will still pass.
Classify this task's difficulty as SIMPLE | MODERATE | HARD.
SIMPLE = extraction/classification/format → cheapest tier.
MODERATE = standard coding/writing → mid tier.
HARD = multi-step reasoning/novel design → flagship.
Return only the label and the chosen model.
Task: {{task}}Output control snippet
Use when: Chatty tasks where output tokens dominate.
Answer only. No preamble, no restating the question, no summary.
Format: {{format e.g. JSON matching schema X}}.
Max length: {{n}} tokens. Stop when complete.Plan-before-build (draft in Claude/ChatGPT, paste to Lovable)
Use when: Before spending Lovable credits on a feature.
You are my app architect. Produce a single Lovable build prompt for:
<feature description>.
Include: data model, components, states/edge cases, acceptance criteria,
and explicit 'do NOT touch' areas. Keep it specific enough to build in one pass.Compaction / handoff summary
Use when: A long session is approaching the context limit.
Summarize this session into a compact handoff:
- Goal & constraints
- Decisions made (with rationale)
- Current state (files/data touched)
- Open questions / next step
Drop raw logs and resolved detours. Be terse.RAG context-trim instruction
Use when: Long retrieved chunks inflating the prompt.
From the context below, keep only sentences needed to answer the question.
Discard background, repetition, and unrelated detail. Then answer.
Question: {{q}}
Context: {{chunks}}CLAUDE.md / AGENTS.md starter
Use when: Setting up a coding-agent project memory.
# Project conventions
Stack: <framework, language, key libs>
Structure: <where things live>
Naming: <conventions>
Do: <patterns to follow>
Don't: <anti-patterns, files to avoid>
Testing: <how to run/verify>
Keep this file lean — it rides in context on every call..cursor/rules starter
Use when: Standing guidance for Cursor.
- Framework: <X>. Prefer <patterns>.
- Naming: <conventions>.
- Always: <e.g. typed props, error handling>.
- Never: <e.g. inline styles, any-types>.
- Scope context with @file/@folder; avoid @codebase unless necessary.Checklists
Before a session
- Pick the smallest model that can plausibly do the job.
- Draft and refine the prompt in a free chat before spending paid credits.
- Load standing rules via instruction file (CLAUDE.md / .cursorrules), not the prompt.
- Decide what context is actually needed — nothing more.
Every prompt
- Stable content first, volatile content last (cache-friendly order).
- Ask for terse output, a fixed format, and no preamble.
- Cap output length / set max_tokens.
- Reference or diff instead of re-pasting whole files.
Production / API
- Enable prompt caching and structure prompts to maximize hit rate.
- Route by task complexity; escalate only on low confidence.
- Move non-urgent jobs to Batch.
- Add a semantic cache for repetitive query shapes.
- Set agent step/tool/token budgets and early-exit on success.
FinOps & governance
- Tag every call by feature/team/use-case (target: high tagging coverage %).
- Track cost per 1K calls and budget utilization %.
- Set budgets and anomaly alerts.
- Review token economics in a regular cadence; fold it into architecture decisions.