Fork a Full Academic Pipeline: 18 Agents, 10 Gates, One Starter Prompt

Academic work is multi-artifact and review-hungry: slides that must not overflow, papers that must not misstate a number, replication packages that must actually run. claude-code-my-workflow is a 2k-star, MIT-licensed template that exports one economics professor's production setup — extracted from a PhD course — so you fork the whole pipeline instead of assembling it skill by skill.

Why This Skill Matters

The template's inventory is concrete: 18 agents, 61 skills, 37 rules, and 11 hooks (the README's own count in its What's Included section). The agents split review by dimension — proofreader for grammar, slide-auditor for layout, pedagogy-reviewer for teaching quality, r-reviewer for R code, plus a domain-reviewer/domain-referee/methods-referee/editor manuscript pipeline. Disciplines come preloaded for Economics and Political Science; forkers extend via journal profiles and discipline cards.

The trust mechanism is the differentiator. One command — ./scripts/backtest.sh — runs ten gates, which the README enumerates: surface-sync, skill integrity, model currency against the single source of truth, link and anchor resolution, Agent Skills spec conformance, staleness, repo hygiene, derived counts, ledger coverage, and a seeded hook battery that re-fires every active guard hook against the failure it targets. An installed pre-commit hook runs the suite plus a quality check (score ≥ 80) on every commit, so bypassing the skill no longer bypasses the review. Every artifact gets a 0–100 quality score with thresholds: 80 to commit, 90 for a PR, 95 for excellence.

Installation

The README's quick start assumes Claude Code, git, and Python 3; XeLaTeX and Quarto are needed for the included demos, R and the GitHub CLI are recommended:

git clone https://github.com/YOUR_USERNAME/claude-code-my-workflow.git my-project
cd my-project
./scripts/validate-setup.sh

validate-setup.sh reports exactly what is missing, with install links. Then start Claude Code and paste the starter prompt from the full guide, filling in your project name and a two-to-three-sentence description. Claude reads the configuration files, adapts them to your project, and enters contractor mode — for complex requests it first drafts a requirements specification with MUST/SHOULD/MAY priorities and clarity status (CLEAR/ASSUMED/BLOCKED) for you to approve. Verify the environment with two demo invocations: /compile-latex HelloWorld compiles the Beamer sample, /deploy HelloWorld renders the Quarto sample. If both succeed, delete the demos and start real work.

Real Workflow: Peer-Review a Manuscript

The review pipeline is where the agents earn their keep. /review-paper --peer <journal> runs the manuscript through the referee pipeline — and with --variance N it runs N referees with sampled dispositions and reports a decision distribution rather than a point estimate, a design the README traces to an AgentReview finding that 37% of decisions vary purely from disposition sampling.

Adversarial QA backs it up: the critic reads both the Beamer and Quarto artifacts and produces harsh findings, the fixer implements exactly what the critic found, and they loop until two consecutive rounds surface nothing new (a 5-round cap is the fallback, not the primary stop). Before submission, three verification skills run their own checks — /verify-claims re-verifies with a fresh-context verifier that has never seen the draft (fabricated citations and numerical contradictions fail closed), /audit-reproducibility cross-checks every numeric claim against the script output that produced it and records per-claim status in passport.yaml, and /humanize flags AI-voice tells, read-only by design.

Tips

• Fork first, clone second — the quick start assumes your fork, since the template is meant to diverge.
• Only need Python/R/Markdown? The README says you can skip XeLaTeX and Quarto entirely and head straight to /data-analysis, /review-paper, /lit-review, or /review-r.
• The gates are honest about their limits: the README describes the git-guardrails and root-of-trust hooks as best-effort textual tripwires that fail open on their own errors — read them as a change of channel, not a lock.
• /compress-session and /promote-memory manage context across sessions; committed MEMORY.md collects learnings that help every forker, machine-specific notes stay local.
• The full guide lives at psantanna.com/claude-code-my-workflow, with the repo's guide/ directory carrying the source.

When Not to Use This

The preloaded discipline profiles are Economics and Political Science — the README names those two, and says forkers extend the rest. A lab scientist or humanities scholar can use the machinery but inherits the journal conventions only after adapting profiles. And if your output is a codebase rather than documents — no slides, no manuscripts — a general CI pipeline with code review gives you the same protection for less setup; the gates here audit the surfaces academic work actually ships.

See the leaderboard for more skills.