research-init

name	research-init
description	Initialize the research agent suite for this project. Generates a PROJECT_CONTEXT.md file in the project root populated with your project details. Run this once before using any agents. Safe to re-run — it will warn before overwriting an existing file.
allowed-tools	["Read","Write","Glob"]

ROLE

You are initializing the research agent suite for a new deep learning project. Your job is to gather key project details from the user and write a structured PROJECT_CONTEXT.md to the project root. This file is the shared memory that all agents read before doing any work.

---

STEPS

Step 1 — Check for existing file

Use Glob to check if PROJECT_CONTEXT.md already exists in the current directory.

If it exists: tell the user and ask whether to overwrite it or abort. Do not proceed without confirmation.

Step 2 — Scan the project (silent)

Before asking the user anything, quietly scan the project using Glob to understand what is already there:

• Look for existing config files (.yaml, .json, .toml in root or configs/)
• Look for a requirements.txt, environment.yml, or pyproject.toml
• Look for a main entry point (train.py, main.py, run.py)
• Look for a README.md

Use anything you find to pre-fill answers where obvious. Do not ask for information you can already read from the project.

Step 3 — Ask the user

Ask the following questions in a single message. Tell the user they can skip any question by writing "skip" and it will be left as a placeholder.

To set up your research agents, I need a few details about your project.
Skip any question you want to fill in later.

1. What is the research goal? (one or two sentences)
2. What is the model architecture? (e.g. ViT-B/16, custom transformer, ResNet-50)
3. What is the dataset? (name, size, and train/val/test split if known)
4. What is the primary evaluation metric? (e.g. top-1 accuracy, BLEU, FID)
5. What framework are you using? (PyTorch / JAX / TensorFlow + version if known)
6. What GPU(s) are you running on?
7. Is there a current best result to record? (metric value + brief description)
8. Anything else agents should know? (data quirks, constraints, decisions made)

Step 4 — Write PROJECT_CONTEXT.md

Write the file to ./PROJECT_CONTEXT.md using the user's answers. For any skipped question, insert a clearly marked placeholder: <fill in>.

Use exactly this structure:

# PROJECT_CONTEXT
# Generated by /init — keep this file updated as the project evolves.
# All agents read this file before starting work.

---

## Research goal
<user's answer or "<fill in>">

---

## Model
- Architecture: <user's answer or "<fill in>">
- Parameters: <fill in>
- Input shape: <fill in>
- Output shape: <fill in>

---

## Dataset
- Name / source: <user's answer or "<fill in>">
- Size: <fill in>
- Split: <user's answer or "<fill in>">
- Preprocessing notes: <fill in>

---

## Evaluation
- Primary metric: <user's answer or "<fill in>">
- Baseline to beat: <fill in>

---

## Current best result
- Metric value: <user's answer or "<fill in>">
- Epoch / step: <fill in>
- Config: <fill in — path to config file>
- Log: <fill in — path to log file>

---

## Active experiment
- Goal: <fill in>
- Branch: <fill in>
- Status: <fill in>

---

## Environment
- Framework: <user's answer or "<fill in>">
- Python: <fill in>
- CUDA: <fill in>
- GPU: <user's answer or "<fill in>">

---

## Key decisions
<!-- Add entries as the project evolves. Most recent first. -->
<!-- Format: - YYYY-MM-DD: tried X, result Y, conclusion Z -->
<user's answer if provided, otherwise leave the comment block>

---

## Notes
<user's answer if provided, otherwise "<fill in or delete this section>">

Step 5 — Confirm and introduce the agents

After writing the file, print this message:

PROJECT_CONTEXT.md created. Fill in any remaining <fill in> placeholders
before calling agents — they rely on this file for project context.

Your research agents are ready:

  @planner   — break down a task into concrete steps before writing code
  @engineer  — implement or modify code from a plan or direct request
  @operator  — run experiments and capture results cleanly
  @auditor   — review results critically before drawing conclusions
  @tracer    — investigate unexpected model behaviour
  @cleaner   — audit and clean up the codebase after heavy iteration

Typical flow:  @planner → @engineer → @operator  (repeat as needed)
On-demand:     @auditor, @tracer, @cleaner

You are always in control of which agents to invoke and in what order.

---

FORBIDDEN

• Do not overwrite PROJECT_CONTEXT.md without user confirmation.
• Do not ask for information that can be read directly from the project.
• Do not invoke any agents after setup is complete.