ML Research Paper Pipeline: Design to Submission
An end-to-end pipeline for ML/AI research papers targeting NeurIPS, ICML, ICLR, ACL, AAAI, and COLM — experiment design tied to claims, incremental-safe execution with an experiment journal, negative-result handling, ensemble self-review with a meta-reviewer, and anonymization-safe submission prep.
A paper stays 'almost done' for months not because the results are bad, but because nothing bridges raw experiment output to an actual draft — so every writing session starts by re-deriving the story from JSON files, and half the numbers that make it into the draft were never traced back to a real result.
Who it's for: PhD students and postdocs writing conference or journal papers under deadline pressure, ML researchers running multi-experiment projects who need the results-to-writeup bridge to stop breaking, research engineers preparing a first paper submission, labs standardizing how experiments get logged, reviewed, and turned into drafts
Example
"Help me get from these experiment results to an ICML submission" → A structured experiment log bridging raw result files to the paper's story, a draft grounded in that log rather than hallucinated numbers, a 3-5 reviewer ensemble with a meta-review pass defaulting to negative bias, a separate claim-verification pass tracing every number back to its source, and an anonymization + pre-compilation checklist before the final PDF
CLAUDE.md Template
New here? 3-minute setup guide → | Already set up? Copy the template below.
npx ccplaybooks add ml-research-paper-pipeline# ML Research Paper Pipeline — Design to Submission
An end-to-end pipeline for producing publication-ready ML/AI research papers for venues like NeurIPS, ICML, ICLR, ACL, AAAI, and COLM. Covers the full lifecycle: project setup, literature review, experiment design, execution and monitoring, result analysis, drafting, self-review, and submission prep.
**This is not a linear pipeline — it's an iterative loop.** Results trigger new experiments. Reviews trigger new analysis. Handle the feedback loops; don't force a strictly sequential march through the phases.
## Core Philosophy
1. **Be proactive.** Deliver complete drafts, not endless clarifying questions — produce something concrete a scientist can react to, then iterate.
2. **Never hallucinate citations.** AI-generated citations have a meaningfully high error rate. Always fetch bibliography data programmatically; mark anything unverifiable as `[CITATION NEEDED]`.
3. **A paper is a story, not a pile of experiments.** State the single contribution in one sentence. If you can't, the paper isn't ready to write.
4. **Experiments serve claims.** Every experiment must state which claim it supports. Don't run experiments that don't connect to the paper's actual narrative.Get new playbooks like this one
One email a week with new Claude Code workflows. Free, like everything here.
No spam. Unsubscribe anytime.
What This Does
Covers the full lifecycle of an empirical ML paper as one iterative loop, not a strict sequence: project setup and contribution framing, a breadth-then-depth literature search with a mandatory citation-verification protocol (never generate a bibliography entry from memory), experiment design where every experiment must map to a specific claim, crash-safe execution with an experiment journal that captures the reasoning tree git alone can't, and result analysis that treats a well-explained negative result as genuinely publishable rather than automatically a dead end.
The two most load-bearing pieces are the experiment log — a structured bridge from raw result files to prose that prevents the drafting pass from hallucinating numbers it had to re-derive — and the self-review phase, which runs 3-5 independent reviewer personas biased toward negative findings, aggregates them with a meta-reviewer, adds a separate visual pass on the compiled PDF and a fresh-context claim-verification pass, before ever reaching submission prep (anonymization, formatting, and automated pre-compilation checks that catch broken citations and missing figures before the LaTeX compiler does).
Quick Start
Step 1: Create a Project Folder
mkdir research-paper && cd research-paper
Step 2: Download the Template
Click Download above, then:
mv ~/Downloads/CLAUDE.md ./
Step 3: Start the Pipeline
claude
Point Claude at your existing repo or results and ask it to work through Phase 0 (project setup and contribution framing) first — it'll propose the one-sentence contribution for you to confirm before anything else proceeds.
Tips & Best Practices
- Revisit the one-sentence contribution from Phase 0 every time a result changes the story — a paper that quietly drifts from its stated contribution reads as unfocused even when every individual section is solid.
- Treat "did they compare against X" as the default reviewer question and make sure the baseline table already answers it before submission — weak baselines are one of the most common rejection reasons.
- Don't force a positive spin a negative result doesn't support. Several strong venues (workshops, TMLR, NeurIPS Datasets & Benchmarks) explicitly want honestly reported null results with rigorous methodology.
Limitations
- Built around empirical ML papers for major ML/AI venues — a theory paper, a survey, or a different field's conventions need real adaptation, not a literal pass through every phase.
- Citation verification and human-evaluation design assume access to real APIs and platforms (a citation database, a crowdsourcing service) — without them, budget more manual verification time.
- A process framework, not a guarantee of acceptance — rigorous baselines and honest limitations sections improve the odds, but reviewer judgment is still reviewer judgment.