research-d-implementation

name	research[D]-implementation
description	ResearchPilot Research Assistant [Phase D]: Implementation design
version	2.0.0
license	LICENSE

user_requirements.md priority: All user constraints recorded in docs/user_requirements.md (direction preferences, implementation requirements, document format, etc.) take precedence over any default instruction in this skill. Always read that file before generating any output to ensure compliance with confirmed user constraints.

Phase D: Implementation Design

Generate a function-level coding guide — precise to every function signature, parameters, return values, and implementation logic — then validate and confirm before coding begins. Produces docs/implementation.md, then moves to Phase E.

Prerequisite: docs/idea_report.md contains Part 3 (Phase C complete).

Workflow Overview & Outputs

ResearchPilot-Skills splits a complete academic research project into seven independent stage skills. The current skill is one link in that chain.

Seven-Stage Chain

Skill	Phase	Main Output
/research[A]-exploration	Direction Exploration	docs/idea_report.md Part 1
/research[B]-idea	Idea Deepening	docs/idea_report.md Part 2
/research[C]-experiment	Experiment Design	docs/idea_report.md Part 3
/research[D]-implementation	Implementation Design	docs/implementation.md
/research[E]-coding	Coding	code/ + docs/dev_log.md
/research[F]-iteration	Code Iteration	dev_log.md iteration records
/research[G.0]-plan	Paper Planning	manuscript architecture + notebooks/figures.ipynb
/research[G.1]-method	Method	manuscript Method section
/research[G.2]-experiments	Experiments	manuscript Experiments section
/research[G.3]-abstract	Abstract	manuscript Abstract
/research[G.4]-introduction	Introduction	manuscript Introduction
/research[G.5]-related	Related Works	manuscript Related Works
/research[G.6]-conclusion	Conclusion + References	manuscript Conclusion
/research[G.7]-review	Full-paper Review	review report

Project Directory Structure

docs/
  idea_report.md        # Research report in three parts:
                        #   Part 1: Motivation, RQs, Key Works  (Phase A)
                        #   Part 2: Introduction, Related Works, Method  (Phase B)
                        #   Part 3: Datasets, Experiment Design, Resource Estimate  (Phase C)
  implementation.md     # Coding guide: file- and function-level implementation plan  (Phase D)
  dev_log.md            # Dev log: progress, decisions, "How to Run" section  (Phase E)
  user_requirements.md  # User constraints: collected and maintained by Claude
  papers/               # Downloaded paper PDFs or abstract TXTs
  manuscripts/          # Paper drafts, each revision archived separately
                        #   (e.g. v1.0-initial-draft.md, v1.1-revision.md)

code/
  src/                  # Core model and training code
  scripts/              # Run scripts (train.sh, evaluate.sh, ablation.sh)
  configs/              # Hyperparameter config files
  src/models/baseline/  # Baseline implementations (interface matches main model)
  notebooks/            # Visualization notebooks; paper figure/table generation
  data/                 # Datasets (gitignored)
  results/              # Experiment results (gitignored)
  logs/                 # Training logs (gitignored)
  README.md             # Environment setup and run commands
  requirements.txt      # Dependencies (library names only, no torch family)

---

Command

/research[D]-implementation

An optional natural-language instruction may follow the command. The AI will treat it as an additional constraint or supplement for this invocation.

---

Phase D Flow Overview

D-0 Ask whether to use a strong baseline (improve existing open-source project
    vs. build from scratch)

Path A (Strong Baseline):
  D-A1 Obtain open-source project (git clone or user clones)
  D-A2 Scan project structure, extract all classes/functions/dependencies
  D-A3 Collect coding constraints (framework, features to preserve, requirements)
  D-A4 Generate implementation.md (strong baseline format)

Path B (From Scratch):
  D-B1 Collect coding constraints (framework, special requirements)
  D-B2 Generate implementation.md (from-scratch format)

D-end1 Run implementation.md validation (coverage / consistency / completeness)
D-end2 Pre-coding checklist (env / device / datasets / run strategy / README location)

Full step-by-step instructions: references/phase-D.md.

---

Validation Rules

After every generation or revision, run three checks:

1. Experiment coverage: Does every experiment in Part 3 have a supporting module/function?
2. Logic consistency: Are tensor shapes consistent across modules?
3. Completeness: Does every listed file have a corresponding implementation section?

Validation output format:

✅ Experiment coverage: {pass / missing: …}
✅ Logic consistency: {pass / issue found: …}
✅ Completeness: {pass / missing: …}

---

Non-Negotiable Constraints

1. Do not advance to Phase E without explicit user confirmation.
2. Run validation immediately after generating implementation.md; fix issues before presenting to user.
3. requirements.txt must not contain torch, torchvision, or torchaudio.
4. Rules in references/template-flexibility.md take precedence over any specific template instruction.
5. The pre-coding checklist (D-end2) must be completed item by item before coding begins.

---

On Phase Completion

After implementation.md and pre-coding checklist are both confirmed:

Phase D complete. implementation.md has been generated and validated.
Pre-coding checklist confirmed.

After implementation.md and pre-coding checklist are both confirmed:

Phase D complete. implementation.md has been generated and validated.

→ Use /research[E]-coding to start coding.

---

Reference Files

• Detailed flow: references/phase-D.md
• Document format specs: references/document-formats.md
• Template flexibility rules: references/template-flexibility.md