researchstack

name

researchstack

description

Workflow-first research skill pack for computer systems, networking, and AI papers. Use when Codex needs to act like a rigorous paper team rather than a generic assistant: scoping a paper, stress-testing an idea, refining a thesis, mapping related work, reproducing prior papers, designing experiments, auditing code and artifacts, writing sections, improving LaTeX layout, designing figures, simulating reviewer feedback, or preparing rebuttals for venues such as ICLR, ASPLOS, SC, NSDI, and SIGCOMM.

Researchstack

Use this root skill as the routing layer for the rest of the pack.

Runtime Bootstrap

Run this first when the host supports shell preambles:

_ROOT=$(git rev-parse --show-toplevel 2>/dev/null || pwd)
RESEARCHSTACK_ROOT="$HOME/.codex/skills/researchstack"
[ -d "$_ROOT/.agents/skills/researchstack" ] && RESEARCHSTACK_ROOT="$_ROOT/.agents/skills/researchstack"
[ -d "$_ROOT/bin" ] && RESEARCHSTACK_ROOT="$_ROOT"
RSTACK_BIN="$RESEARCHSTACK_ROOT/bin"
RSTACK_HOST=$("$RSTACK_BIN/researchstack-host-detect" 2>/dev/null || echo "unknown")
RSTACK_PROJECT=$("$RSTACK_BIN/researchstack-project-slug" --root "$_ROOT" --print-json 2>/dev/null || echo "{}")
RSTACK_MEMORY=$("$RSTACK_BIN/researchstack-memory-status" --root "$_ROOT" --print-json 2>/dev/null || echo "{}")
RSTACK_CONFIG=$("$RSTACK_BIN/researchstack-config" show 2>/dev/null || echo "{}")
echo "HOST: $RSTACK_HOST"
echo "PROJECT: $RSTACK_PROJECT"
echo "MEMORY: $RSTACK_MEMORY"
echo "CONFIG: $RSTACK_CONFIG"

Interpret the bootstrap like this:

• If memory exists and the user is unsure what to do, route to researchstack-next-step.
• If memory is empty and the user only has a broad area, route to researchstack-idea-finder.
• If memory is empty and the user already has a thesis candidate, route to researchstack-lab-intake.
• If routing files are missing in the repo, recommend bin/researchstack-install-routing.
• If installs look stale or duplicated, recommend bin/researchstack-doctor.

Read references/workflow.md first when the request spans multiple stages of a paper. Read references/memory.md when the project is ongoing and prior decisions should carry across sessions.

Default behavior when the user says "help me push this paper" or asks for a broad research assist:

1. Start with researchstack-next-step if the user is mid-project and asks what to do next.
2. Start with researchstack-idea-finder if the user only has a broad area, not a paper-sized thesis.
3. Otherwise start with researchstack-lab-intake unless the thesis is already crisp.
4. Run researchstack-idea-review before investing in polishing.
5. If the idea is salvageable but weak, switch to researchstack-idea-refine.
6. Build the related-work boundary with researchstack-literature-map.
7. Lock evaluation with researchstack-experiment-design.
8. Move immediately into researchstack-experiment-ops so runs, seeds, and result promotion are controlled.
9. Use researchstack-artifact-audit and researchstack-code-review before trusting any figure or claim.
10. Only then move to researchstack-paper-write, researchstack-figure-studio, and researchstack-paper-layout.
11. Use researchstack-submission-gate before any external submission.
12. Finish with researchstack-peer-review and, if needed, researchstack-rebuttal-coach.

Do not jump straight to writing or figure polish when the thesis, evidence, or process are still unstable.

Route to the most specific skill:

• Use researchstack-lab-intake for raw project framing, venue targeting, and turning a vague direction into a paper plan.
• Use researchstack-next-step when the user is mid-project, stuck, or unsure which research skill to use next.
• Use researchstack-idea-finder when the user has an area of interest, venue intuition, or resource constraint but not yet a one-paper thesis.
• Use researchstack-idea-review for skeptical novelty and risk review.
• Use researchstack-idea-refine for guided improvement after the idea is weak but still promising.
• Use researchstack-literature-map for related-work positioning and gap analysis.
• Use researchstack-learn for project memory, venue memory, experiment policies, and long-term researcher preferences.
• Use researchstack-paper-reproduction when the user provides a paper PDF and wants to reconstruct or reproduce it.
• Use researchstack-experiment-design for evaluation planning before implementation or before paper lock.
• Use researchstack-experiment-ops for runbooks, logging discipline, result triage, and experimental process management.
• Use researchstack-artifact-audit for reproducibility and evidence tracing.
• Use researchstack-code-review for research-code correctness and claim alignment.
• Use researchstack-paper-write for outline and section drafting.
• Use researchstack-paper-layout for LaTeX structure, notation, tables, captions, and visual consistency.
• Use researchstack-figure-studio for figure and table planning.
• Use researchstack-submission-gate for final readiness and venue-fit decisions.
• Use researchstack-peer-review for harsh conference-style reviews.
• Use researchstack-rebuttal-coach for point-by-point response strategy.

If the user says "what do we already know about this project?" or wants the pack to remember durable decisions, route to researchstack-learn first.

When the user asks for an end-to-end research push, follow this sequence:

1. Next-step triage if the project is already in flight and the right next move is unclear.
2. Idea finding if the project starts from a broad area rather than a thesis.
3. Intake and venue target.
4. Skeptical idea review.
5. Guided idea refinement if salvageable.
6. Literature positioning.
7. Experiment plan.
8. Experiment operations and result promotion.
9. Artifact and code audit.
10. Writing and layout.
11. Submission gate.
12. Simulated peer review.
13. Rebuttal coaching if needed.

Use this next-step branch whenever the user is unsure what to do next:

1. researchstack-next-step
2. the one or two recommended follow-up skills
3. researchstack-learn after the step if a durable decision was made

Use this idea-finding branch when the user starts with a topic area, not a paper-sized idea:

1. researchstack-idea-finder
2. researchstack-lab-intake
3. researchstack-idea-review
4. researchstack-idea-refine if needed
5. researchstack-literature-map

Use this reproduction branch when the task starts from an existing paper rather than a new internal project:

1. researchstack-paper-reproduction
2. researchstack-experiment-design
3. researchstack-experiment-ops
4. researchstack-artifact-audit
5. researchstack-peer-review

The reproduction branch is about reconstructing what the paper actually did, identifying ambiguity, and making a clean verdict about reproducibility. It is not only about whether one number matched.

Keep the team behavior strict:

• Challenge novelty claims early.
• Prefer falsifiable hypotheses over inspirational framing.
• Keep claims narrower than the current evidence, not broader.
• Call out missing baselines, ablations, and deployment realism.
• Distinguish systems evidence, networking evidence, and ML evidence instead of mixing standards loosely.