Spec Prompting

GPTClaudeGemini··▲ 848 copies·updated 2026-07-14

spec-prompting.prompt

# Prompting Specification — MLE Agent

> **0→1 baseline.** Post-0→1 changes captured in [`../decisions.md`](../decisions.md) — notably **D1** (ML spec is *context and direction*, not binding prescription — explicit in current Action Node prompts), **D2** (Architect HARD STOPs), **D8** (no medal targets), **D9** (Evaluator restricted to 2 blocker types).

## Philosophy

Prompt as a manager, not a script writer. A good instruction encodes **what success looks like** and **what is off-limits** — it does not narrate the steps. A node briefed with an identity and a constraint set will find its own route; a node handed a procedure becomes brittle the moment reality diverges from the script.

Three things belong in a system prompt:
1. **Who you are** — a declaration of role and professional instinct, not a job description
2. **What done looks like** — the exit condition for this node's work
3. **Hard rails** — the non-negotiables that protect the rest of the system (I/O hygiene, cold-storage limits, commit discipline)

Everything else is the node's to figure out.

---

## System Prompt Architecture

Prompts are assembled in two layers at runtime:

when to use it

Community prompt sourced from the open-source GitHub repo dirk61/MLE-Squad (no explicit license). A "Spec Prompting" style prompt — adapt the placeholders and specifics to your task. Imported as-is and not independently retested here, so check the output before relying on it.