research-gstack

AI-assisted workflow skills for CS academic labs — systems, cloud, and networking research.

Adapted from gstack by applying the same sprint-lifecycle philosophy to the research lifecycle: ideate → plan → run → analyze → write → submit → reflect.

The Research Lifecycle

IDEATE → PLAN → RUN → ANALYZE → WRITE → SUBMIT → REFLECT
/idea-   /pi-  /cluster /results /paper  /submit   /lab-
clinic   review  -run    -audit   -sprint   -gate    retro
  +        +            +          +         +
/lit-   /methods /analysis   /pi-    /rebuttal
scan    -review    -run       review   -room
           +
    /design-experiments

Skills

Ideation

• /idea-clinic — Shape a fuzzy idea into a durable research charter with 6 diagnostic questions, premise challenge, and landscape awareness
• /lit-scan — Map related work, identify baseline requirements, anticipate reviewer comparisons

Planning

• /pi-review — PI-level scope and contribution review: mode selection (broaden/cherry-pick/hold/narrow), strongest claim, what to cut
• /methods-review — Lock experiment design before compute: hypothesis map, baselines, ablations, figure plan, compute budget
• /design-experiments — Explore experimental design space: generate 4-5 designs, compare tradeoffs, pick the best; or audit an existing design for flaws
• /research-autoplan — Full automated pipeline: idea-clinic → pi-review → methods-review with auto-decisions, only surfaces genuine judgment calls

Execution

• /cluster-run — Run tracking and registry: register experiments, poll status, compare configs, link figures to run IDs
• /analysis-run — Execute planned analysis, gather artifacts, log deviations from expectations

Results & Writing

• /results-audit — Claim-evidence audit: two-pass review (critical + informational), fix-first heuristic
• /paper-sprint — Turn results into narrative: one-sentence claim, contribution framing, section plan, result story, LaTeX scaffold
• /sanity-check — Pre-submission numbers audit: verify text matches tables, all citations present (report-only)

Publication

• /submission-gate — Final pre-submission checklist: 8 checks, binary READY/NOT READY verdict
• /rebuttal-room — Reviewer response planning: cluster concerns, map to evidence, triage experiments, draft responses
• /artifact-release — Package code/data/models for release: license check, REPRODUCE.md, artifact evaluation prep

Lab Operations

• /lab-retro — Weekly lab retrospective: experiment velocity, project status, researcher breakdown, process debt
• /onboard-student — New lab member onboarding: reading list, toolchain setup, expectations, first tasks
• /mentoring-review — Student progress review: progress assessment, technical assessment, praise + growth area
• /authorship-lock — Freeze author order pre-submission: CRediT contributions, acknowledgments draft
• /deadline-mode — Submission deadline tightening: must-have vs. deferrable, time budget, crunch rules

Safety & Governance

• /data-guard — Data governance overlay: stops on PII, license issues, IRB requirements, redistribution concerns
• /careful — Destructive action warnings: warns before rm -rf data/, overwriting configs, deleting checkpoints
• /freeze — Restrict edits to one directory: prevents accidental changes outside target scope
• /guard — Combined safety: freeze + careful together

Debugging & Verification

• /investigate — Systematic root cause analysis: 4-phase debugging, Iron Law (no fixes without root cause)
• /reproduce-check — Iterative reproducibility verification: find and fix reproduction failures
• /second-opinion — Cross-model review: get an independent AI perspective, with PI Sovereignty rules

Utilities

• /document-release — Post-publication doc updates: README, lab wiki, project status
• /unfreeze — Remove freeze boundary

Core Principles

Run the Full Experiment — When ablations are cheap, do the complete set. Three more GPU-hours to answer the question cleanly beats a weak ablation.

Read Before Running — Layer 1 (established baselines) → Layer 2 (recent arXiv, scrutinize) → Layer 3 (first-principles insight the crowd missed). The paper that gets remembered found what everyone else overlooked.

PI Sovereignty — Claude recommends. PIs decide. Two models agreeing is a signal, not a mandate. The researcher has context the model lacks.

State

All persistent state lives in ~/.labstack/:

• runs/ — experiment run registry (JSON per run)
• analytics/ — skill usage log
• projects/ — per-project state and plans
• retros/ — weekly retrospective history
• data-governance/ — data clearance records

Setup

git clone <this-repo> ~/.claude/skills/research-gstack
# Skills are now available as /idea-clinic, /pi-review, etc.