Claude Code My Workflow Agent Skill

My Claude Code Setup

Actively maintained. A summary of how I use Claude Code for academic work — slides, papers, data analysis, and more — packaged so you can fork it for your own research. See CHANGELOG.md for the latest changes.

Live site: psantanna.com/claude-code-my-workflow

A ready-to-fork foundation for AI-assisted academic work. You describe what you want — lecture slides, a research paper, a data analysis, a replication package — and Claude plans the approach, runs specialized agents, fixes issues, verifies quality, and presents results. Like a contractor who handles the entire job. Extracted from a production PhD course and extended by a growing community.

---

Quick Start (5–10 minutes, plus ~30 min for first-time installs)

Before you start: Claude Code, git and Python 3 are the minimum. Python 3 runs the hooks, the gate suite (./scripts/backtest.sh — 11 checkers) and the quality scorer, and is pre-installed on macOS/Linux. To run the included HelloWorld demos end-to-end you also need XeLaTeX (Beamer sample) and Quarto (Quarto sample). R and the GitHub CLI are recommended. Full list in Prerequisites below. Fastest path: clone first, then run ./scripts/validate-setup.sh — it reports exactly what's missing with install links.

Only need Python/R/markdown? You don't need XeLaTeX or Quarto. The agents, rules, skills, and orchestration patterns work for any text/code artifact. Skip the HelloWorld demos and head straight to /data-analysis, /review-paper, /lit-review, or /review-r.

Session 2 onwards: MEMORY.md (committed) collects generic [LEARN] entries that help all forkers; machine-specific notes accumulate in Claude Code's native auto memory (~/.claude/projects/<project>/memory/, machine-local, never committed). See .claude/rules/meta-governance.md for the distinction.

1. Fork & Clone

Unpublished work? Choose before you copy. GitHub keeps a fork of a public repository public, and its visibility cannot be changed. For a paper, data or anything not yet public, click Use this template on the repository page instead, set the new repository to Private, and clone that one: git clone https://github.com/YOUR_USERNAME/YOUR_REPO.git my-project. Fork only for work that can be public.

# Fork this repo on GitHub (click "Fork" on the repo page), then:
git clone https://github.com/YOUR_USERNAME/claude-code-my-workflow.git my-project
cd my-project
./scripts/validate-setup.sh        # reports missing tools with install links

Replace YOUR_USERNAME with your GitHub username.

2. Start Claude Code and Paste This Prompt

claude

Using VS Code? Open the Claude Code panel instead. Everything works the same — see the full guide for details.

Using Codex (now also a tab in the ChatGPT desktop app) or Antigravity? Start with AGENTS.md. Those agents read the skills in .agents/skills/: generated wrappers that point to the real instructions in .claude/skills/, for every skill one agent can run end to end. Claude Code's deny rules and guard hooks do not apply to them, so AGENTS.md lists what they must do by hand; the pre-commit gate and CI still apply. Skills built on several independent reviewers (/review-paper --peer, /seven-pass-review, /verify-claims) and the ones that manage Claude Code itself stay Claude-only — the list and the reasons are in scripts/sync-agent-skills.py.

Avoid prompt fatigue. On Claude Code ≥ 2.1.283, new interactive terminal and VS Code sessions start in auto mode by default (a classifier approves routine actions and prompts on risky ones). Where auto mode is unavailable, this template's .claude/settings.json ships broad allow rules (Edit(**), Write(**), Bash(*)), so routine work inside the repo runs unprompted. It does not set permissions.blockWritesOutsideWorkingDirectories (removed in v2.9.1), so Claude can change .claude/settings*.json itself. To turn that block on, add "blockWritesOutsideWorkingDirectories": true under permissions (Claude Code ≥ 2.1.314): Claude's file tools then refuse writes outside the working directories and every write to a Claude Code settings file, even inside the repository, and the worktree tools (EnterWorktree, ExitWorktree) leave the session. A true in any settings file applies, and lifting it takes a restart. If you still see prompts, press Shift+Tab to switch to Auto-accept edits (claude --permission-mode acceptEdits). For container/VM Bypass mode or restricted-data deny rules, see TROUBLESHOOTING.md.

Then paste the starter prompt from the guide, filling in your project details:

I am starting to work on [PROJECT NAME] in this repo. [Describe your project in 2–3 sentences.] I've set up the Claude Code academic workflow... Please read the configuration files and adapt them for my project. Enter plan mode and start.

The full guide has the complete starter prompt with all the details.

What this does: Claude reads all the configuration files, fills in your project name, institution, and preferences, then enters contractor mode — planning, implementing, and (within the skill you invoke) running the review + verify loop. You approve the plan, invoke a skill, and the skill handles the rest within its scope.

Heavily adapting CLAUDE.md for a non-academic project? Anthropic's built-in /init command will re-derive a CLAUDE.md from your codebase as a starting point. The pre-shipped CLAUDE.md in this template already covers the academic setup — you only need /init if your fork diverges substantially (e.g., a Python/ML project that doesn't use LaTeX or Quarto).

3. Verify Your Setup

Before building real lectures, confirm your environment works:

./scripts/validate-setup.sh        # Checks XeLaTeX, Quarto, Python, git, etc.

Then inside Claude:

/compile-latex HelloWorld          # Compiles Slides/HelloWorld.tex to PDF
/deploy HelloWorld                 # Renders Quarto/HelloWorld.qmd to HTML

If both succeed, delete Slides/HelloWorld.tex and Quarto/HelloWorld.qmd and start on your real work.

---

How It Works

Goal-first, gate-enforced (the v2.0 shift)

You don't craft a perfect prompt — you state a goal and let the work loop toward it under gates. Specialist agents do the labor; enforcing gates decide when it's good enough; you adjudicate the disagreements they surface. Three things make that trustworthy:

• Real gates, not reminders. One command — ./scripts/backtest.sh — runs eleven gates: surface-sync, skill integrity, model currency against the SSoT, link and anchor resolution, Agent Skills spec conformance, staleness (including source-vs-published divergence), repo hygiene, derived counts (enumerable claims re-counted from disk), ledger coverage (the qualification ledger and the checks that actually run must agree in both directions, and every hook declared in settings must exist and be invocable), a seeded hook battery (every active guard hook is re-fired against the failure it targets, alongside clean controls, on every run), and a cross-platform portability suite (path, encoding, line-ending and git-output cases). A version-controlled pre-commit hook (./scripts/install-hooks.sh) runs it plus the quality check (≥80) on every commit. In the session, destructive git commands are refused by permission deny rules in .claude/settings.json, which Claude Code checks itself, with no hook process. Two PreToolUse hooks remain, and Claude Code starts each only for the commands it checks: dirty-tree-guard for git merge, pull and rebase, and issue-guard for gh. Only issue-guard is registered with "onFailure": "block" (Claude Code ≥ 2.1.295), so a machine without python3 no longer has Bash, Edit and Write blocked; its gh commands are refused until python3 works. The review runtime also re-checks any reviewer-introduced "fatal" finding before it counts.

  <details><summary><em>What the in-session guards catch, and what they miss (click to expand)</em></summary>

  The deny rules refuse git reset --hard, every git clean (dry runs included), git push --force and -f (--force-with-lease still runs), git add -A, --all and . wherever they appear among the arguments, and a checkout or restore of the whole tree (git checkout ., git checkout -- ., git checkout HEAD -- ., git restore ., git restore -- .; also git restore --staged ., which only unstages). They also refuse reads and edits of .env files and reads of *.pem, *.key and credentials* files. Claude Code applies a deny rule to every part of a compound command, including a subshell or a $(…) substitution, strips wrappers such as timeout and nohup, and keeps it in every permission mode. It does not match a rewritten spelling (git -C . push --force, git -c x=y push, /usr/bin/git, bash -c '…'), and Anthropic's permissions docs call Bash deny rules a guard on the invocation Claude usually produces, not a security boundary. dirty-tree-guard refuses a merge, pull or rebase while tracked files have uncommitted changes; untracked files do not count, --abort and --continue pass, and ALLOW_DIRTY_MERGE=1 overrides it. It reads the session's directory, so it misses a repository reached by cd elsewhere && or git -C, and it fails open. No hook refuses shell writes into .claude/settings*.json, .claude/hooks/ or .githooks/ any more; blockWritesOutsideWorkingDirectories still refuses file-tool writes to the settings files. Nothing here recovers anything either: the transcript and git reflog are an audit trail and a commit-history aid, and neither holds the bytes of an uncommitted edit or an untracked file.

  </details>

• Every gate is qualified, and the ledger is itself a gate. Each one has been shown a planted defect and confirmed to go red, with recall and false-alarm rate recorded in quality_reports/qualification/LEDGER.md — and that ledger is now load-bearing rather than aspirational: a registered check with no row there fails the build, and a row naming a checker that no longer exists fails too. Checks that have not been qualified are listed there by name as visible debt — because an unqualified check is not weak evidence, it is none. Run /vaccinate to qualify one.

• A real orchestration runtime. Reviews fan out to forked specialist agents, reduce over a shared finding schema, judge with a hallucination gate, and loop until dry — see orchestrator-protocol.md.

• Ground truth as a process. A mismatch isn't always a failure: a defensible, named alternative is recorded as EXPLAINED and carried into your response-to-referees, while genuine errors stay fail-closed.

This is not an autonomous daemon — the loop is always you- or skill-initiated, and you stay the auditor. Scheduled automation handles recurring chores and notifies only on findings — cloud Routines for committed-repo checks (weekly lit-delta), Desktop scheduled tasks for anything touching local data (inbox triage, and the nightly reproducibility check).

Contractor Mode

You describe a task. For complex or ambiguous requests, Claude first creates a requirements specification with MUST/SHOULD/MAY priorities and clarity status (CLEAR/ASSUMED/BLOCKED). You approve the spec, then Claude plans the approach and runs the right skill — or, for user-invoked skills such as /create-lecture, tells you which one to run (e.g. /create-lecture, /qa-quarto, /review-paper --adversarial). That skill implements the orchestrator runtime internally — implement, verify, review, fix, re-verify, score — and returns a summary when the work meets quality standards. Say "just do it" and it runs the full loop; commits still require an explicit /commit (which the pre-commit hook then gates).

Specialized Agents

Instead of one general-purpose reviewer, 18 focused agents each check one dimension. A representative sample:

• proofreader — grammar/typos
• slide-auditor — visual layout
• pedagogy-reviewer — teaching quality
• r-reviewer — R code quality
• domain-reviewer — field-specific correctness, slides (template — customize for your field)
• domain-referee / methods-referee / editor — manuscript peer-review pipeline (/review-paper --peer)

Each is better at its narrow task than a generalist would be. The /slide-excellence skill runs the slide-review agents in parallel; /review-paper --peer runs the paper-review pipeline. The same pattern extends to any academic artifact — manuscripts, data pipelines, proposals.

Adversarial QA

Two agents work in opposition: the critic reads both Beamer and Quarto and produces harsh findings. The fixer implements exactly what the critic found. They loop until dry — converging after two consecutive rounds surface no new issue (a 5-round cap is the fallback, not the primary stop). This catches errors that single-pass review misses.

Quality Review

Every artifact gets a score (0–100). Scores below threshold halt the workflow and surface the findings — the user decides whether to fix or explicitly override:

• 80 — commit threshold
• 90 — PR threshold
• 95 — excellence (aspirational)

Framing honesty: Thresholds are advisory at the harness level — the /commit skill runs quality checks and halts on failure. And as of v2.0, running ./scripts/install-hooks.sh once installs a real pre-commit hook (.githooks/pre-commit) that runs the backtest gate suite (the hook battery only when a hook, its settings or the battery itself is staged; CI always runs everything) plus the quality (≥80) gate on every commit, so bypassing the skill no longer bypasses the review. Opt out per-commit with SKIP_QUALITY_GATE=1 or git commit --no-verify.

Context Survival

Plans, specifications, and session logs survive auto-compression and session boundaries. The PreCompact hook saves a context snapshot before Claude's auto-compression triggers, ensuring critical decisions are never lost. MEMORY.md accumulates learning across sessions, so patterns discovered in one session inform future work.

For forced compression (long pipelines, mid-plan handoffs), /compress-session (v1.9.0) distils the conversation into a structured note — decisions, next actions, and discarded-as-noise — instead of letting auto-compaction truncate. /promote-memory (v1.9.0) periodically harvests generic learnings from native auto memory to committed MEMORY.md via a five-critic council.

Verification Discipline (v1.7.0+)

Multiple complementary verification layers run before submission:

• /verify-claims (v1.7.0) — Chain-of-Verification with a fresh-context verifier that cannot self-confirm because it has never seen the draft. v1.9.0 adds HIGH/MED/LOW-WARN severity tiers; HIGH-WARN findings (fabricated citation, numerical contradiction) fail the verification closed — the draft is never reported as verified; /commit does not read these verdicts, so resolving them before committing is on you.
• /audit-reproducibility (v1.4.0; Stata coverage v1.9.0) — every numeric claim in the manuscript is cross-checked against the script output that produced it. v1.9.0 adds passport.yaml — a per-paper YAML state file with PASS/FAIL/STALE/UNVERIFIED status per claim.
• /humanize (v1.9.0) — detect AI-voice tells (boilerplate transitions, hedging stacking, sycophancy) before submission. Read-only by design; auto-rewriting degrades quality.
• /review-paper --peer <journal> --variance N (v1.9.0) — runs N referees with sampled dispositions and reports a decision distribution, not a point estimate. Motivated by AgentReview (EMNLP 2024) finding 37% of decisions vary purely from disposition sampling.

---

The Guide

For a comprehensive walkthrough, read the full guide (or see the source).

It covers:

1. Why This Workflow Exists — the problem and the vision
2. Getting Started — fork, paste one prompt, and Claude sets up the rest
3. The System in Action — specialized agents, adversarial QA, quality scoring
4. The Building Blocks — CLAUDE.md, rules, skills, agents, hooks, memory
5. Workflow Patterns — slides, research, reproducibility, presentation rhetoric, sequential adversarial audits, and more
6. The Ecosystem — extensions by clo-author, claudeblattman, MixtapeTools, autoresearch, ClaudeCodeTools, and a growing community
7. Customizing for Your Domain — creating your own reviewers and knowledge bases

2026 Features

The guide covers Claude Code's latest capabilities:

• Model lineup — the Fable tier (opt-in via /model fable or the best alias) is the top tier for long-horizon work; the Opus tier is the Claude Code default and the template's high-judgment tier. Current point versions, prices, effort defaults, and the provider-dependent alias table live in the single source of truth, model-versions.md — surfaces here stay tier-abstract so they cannot go stale, and the staleness gate fails the build when the SSoT's own expiry passes.
• Effort levels — /effort sets cost vs. thoroughness (low / medium / high / xhigh / max). Default effort differs by tier (per the model SSoT): the current Opus and Haiku tiers default to medium — and Opus's medium matches the prior Opus generation's high — while the Fable and Sonnet tiers default to high. Set effort explicitly (CLAUDE_CODE_SUBAGENT_EFFORT sets the subagent default), lower it before prompting for brevity, reserve xhigh for extended exploration and the ultracode toggle (dynamic workflows; since Claude Code 2.1.284 it no longer forces xhigh) for the largest autonomous runs, with workflowSizeGuideline to cap how many agents a workflow plans; maxEffortLevel caps effort on every surface.
• /goal <verifiable condition> (v1.9.0; Anthropic May 2026) — keep working across turns until a fast model confirms the condition holds. Pairs with /commit quality gates for verified-end-state runs.
• claude agents dashboard (v1.9.0; Anthropic May 2026) — single screen for parallel review work (/review-paper --peer, /slide-excellence).
• Cost-Conscious Composition — prompt-cache TTL (5-min default on API keys; on Claude subscriptions 1 hour for the main conversation but 5 minutes for subagents, unless you set subagentPromptCacheTtl), 70/20/10 model routing (Haiku/Sonnet/Opus), /usage monitoring (it absorbed /cost), Agent SDK credit-pool split (2026-06-15).
• Skill frontmatter — effort, context: fork, agent, hooks, disable-model-invocation (v1.8.0+), disallowed-tools (the actual tool restriction — allowed-tools only pre-approves), paths (glob-scoped auto-activation), and dynamic content ($ARGUMENTS, !command syntax)
• Permission modes — Manual (config value default), Accept edits, Plan, Auto (classifier-gated; the built-in starting mode for new interactive terminal and VS Code sessions on Claude Code ≥ 2.1.283 — on earlier versions, on Pro, Max, and Team since 2026-08-14 — and available on Bedrock / Google Cloud / Foundry without an opt-in flag), Bypass
• Hook handler types — five (command, HTTP, MCP tool, prompt, agent) across 33 hook events (Claude Code 2.1.318); hooks see effort.level and $CLAUDE_EFFORT (Apr 2026 Week 19)
• Advanced agent configuration — model, maxTurns, isolation, tool restrictions; model-routing.md rule codifies per-agent tier (v1.9.0)
• Worktree base ref (v1.9.0; Anthropic Apr 2026) — worktree.baseRef setting controls fresh (default; remote default-branch) vs head (local HEAD) for new worktrees
• Built-in skills — /fewer-permission-prompts, /team-onboarding, /autofix-pr, /powerup, Ultraplan, /loop (self-pacing)
• Plugins — /plugin (its Discover tab) for third-party extensions. Since 2.1.287 a plugin can also ship a mod, code inside Claude Code that on a personal account can override deny rules and hooks (see TROUBLESHOOTING)
• Mid-2026 additions (Jun–Sep 2026) — /skill-doctor (what each skill costs in context and how often it fires), claude plugin eval (score a plugin against test cases vs a no-plugin baseline), maxEffortLevel (cap effort on every surface — a hard ceiling for a fixed grant budget), built-in output styles (Concise for everyday work; Learning, which leaves parts for you to write — pairs with /scaffold-exercises; Explanatory; and Proactive, which skips plan mode unless you ask, so it works against plan-first), subagents running in the background and fork mode on by default, /fork (continue a copy of the conversation in a background session), /doctor (setup checkup), and autoUpdatesChannel: "stable" for deadline weeks

---

Use Cases

Academic Task	How This Workflow Helps
Lecture slides (Beamer/Quarto)	Full creation, translation, multi-agent review, deployment
Research papers	Literature review, manuscript review, simulated peer review (/review-paper --peer [journal]), reviewer-disposition variance reporting (--variance N)
Data analysis	End-to-end R pipelines (/data-analysis) or Stata pipelines via stata-mcp (/stata-replication, v1.9.0), replication verification, publication-ready output
Monte Carlo simulations	Reproducible simulation studies (/simulation-study, v1.10.0) — parameterized DGP, estimator grid, bias/RMSE/coverage/size/power with Monte Carlo SEs, dedicated sim-reviewer review pass
Package development	R package release gate (/r-package-check, v1.10.0) — devtools::document() + tests + R CMD check --as-cran + CRAN-policy triage + r-package-reviewer (Stata / Python checks on the roadmap)
Replication packages	AEA-compliant packaging, reproducibility audit trails, passport.yaml claims provenance (v1.9.0)
Presentations	Rhetoric of decks principles, visual audit, cognitive load review
Research proposals	Structured drafting with adversarial critique
Preregistration	OSF / AsPredicted / AEA RCT Registry-ready document (/preregister --style) — full workflow in Pattern 16
Manuscript submission discipline	/humanize (detect AI voice), /verify-claims HIGH-WARN fail-closed reporting (a fabricated citation is never reported as verified), reviewer-disposition variance

Disciplines preloaded: Economics (top-5 journal profiles, R conventions) and Political Science (APSR / AJPS / JOP profiles, formal-theory + survey-experiment paper types, conjoint/cjoint conventions). Forkers extend for psych / sociology / public-health via journal profiles + paper types + discipline cards.

One repo, many project types

This workflow is designed as a single hub for an entire research program — not one paper at a time. The same CLAUDE.md, rules, agents, and quality gates serve courses and lectures, papers and referee reports, data analysis and replication packages, Monte Carlo simulation studies (/simulation-study + sim-reviewer), and the R package release gate (/r-package-check + r-package-reviewer) — all new in v1.10.0. On the roadmap: Stata / Python package checks (SSC / PyPI) and personal-productivity workflows. See .claude/references/v2.0-backlog.md for what's next.

---

What's Included

<details> <summary><strong>18 agents, 62 skills, 37 rules, 9 hooks</strong> (click to expand)</summary>

Agents (.claude/agents/)

<!-- surface-sync-table: agents -->

Agent	What It Does
proofreader	Grammar, typos, overflow, consistency review
slide-auditor	Visual layout audit (overflow, font consistency, spacing)
pedagogy-reviewer	13-pattern pedagogical review (narrative arc, notation density, pacing)
r-reviewer	R code quality, reproducibility, and domain correctness
tikz-reviewer	Merciless TikZ diagram visual critique
beamer-translator	Beamer-to-Quarto translation specialist
quarto-critic	Adversarial QA comparing Quarto against Beamer benchmark
quarto-fixer	Implements fixes from the critic agent
verifier	End-to-end task completion verification
domain-reviewer	Template for your field-specific substance reviewer
claim-verifier (v1.7.0)	Chain-of-Verification fact-checker in a fresh context (never a conversation fork)
editor (v1.5.0)	Journal editor for /review-paper --peer (desk review + referee selection + synthesis)
domain-referee (v1.5.0)	Disposition-primed substance referee for --peer mode
methods-referee (v1.5.0+)	Paper-type-aware methodology referee (6 paper types)
humanize-auditor (v1.9.0)	Read-only AI-voice auditor invoked by /humanize
promote-memory-council (v1.9.0)	Five-critic council for [LEARN] promotion to MEMORY.md
sim-reviewer (v1.10.0)	Monte Carlo simulation reviewer — DGP/estimand match, Monte Carlo SE, coverage-vs-truth, claims↔tables parity
r-package-reviewer (v1.10.0)	R package-source reviewer — DESCRIPTION/NAMESPACE hygiene, roxygen completeness, testthat coverage, CRAN-policy red flags

Skills (.claude/skills/)

Start here on Day 1 — a short list to try first:

• Explore & frame: /interview-me (formalize a fuzzy idea) · /lit-review (literature synthesis + gap check)
• Run & audit data: /data-analysis (end-to-end R pipeline; or /stata-replication for Stata) · /audit-reproducibility (verify manuscript numbers ↔ code outputs)
• Review & ship: /review-paper --peer <journal> (simulated editor + 2 referees, plus /verify-claims fact-check) · /commit (run gates and commit cleanly)
• Teach: /create-lecture (draft a lecture deck) · /compile-latex (build the Beamer PDF)

<!-- surface-sync-table: skills -->

Skill	What It Does
/compile-latex	3-pass XeLaTeX compilation with bibtex
/deploy	Render Quarto + sync to GitHub Pages
/extract-tikz	TikZ diagrams to PDF to SVG pipeline
/proofread	Launch proofreader on a file
/visual-audit	Visual layout audit of a deck (the slide-auditor checks, run inline)
/pedagogy-review	Launch pedagogy-reviewer on a file
/review-r	Launch R code reviewer
/qa-quarto	Adversarial critic-fixer loop (loops until dry; 5-round cap is a fallback)
/slide-excellence	Combined multi-agent review
/translate-to-quarto	Full Beamer-to-Quarto translation — Phase 0 pre-flight plus 12 translation phases (1–11 and a 6.5 pedagogy review)
/vaccinate	Measure whether a check, gate, or AI reviewer actually detects the failure it targets — seeds defects + a clean control, reports recall and false-positive rate into a qualification ledger
/adjudicate-review	Turn incoming findings — AI review, referee report, linter, second model — into verified fixes. Every finding is a CANDIDATE until checked against the source
/blast-radius	Before and after changing anything shared (return value, schema, default, units), enumerate every consumer and actually run them
/credible-claims	Research brief before delegating, claim record after. Keeps faster execution from being mistaken for credible evidence
/differential-audit	Compare two implementations — a port, a replication, a refactor, a version upgrade — so that agreement means something
/oracle-review	Run an external frontier-model referee (Claude Code → another vendor's frontier model via the Oracle CLI) and adjudicate what comes back
/verify-artifact	Prove the file you are about to send IS the thing you mean — rebuild, verify integrity, diff against source
/voice-profile	Extract a written voice profile from your own prior papers, then audit drafts against it — the positive counterpart to /humanize, which only detects AI tells
/validate-bib	Cross-reference citations against bibliography
/verify	Run the backtest and the quality score on what is about to be committed; Claude Code tells Claude to run it before committing (2.1.286+), a reminder next to the enforcing pre-commit hook
/devils-advocate	Challenge design decisions before committing
/create-lecture	Full lecture creation workflow
/commit	Run the gates, then commit on a branch; --pr also pushes and opens a pull request. Never merges — that waits for your say-so
/lit-review	Literature search, synthesis, and gap identification
/research-ideation	Generate research questions and empirical strategies
/interview-me	Interactive interview to formalize a research idea
/review-paper	Manuscript review: structure, econometrics, referee objections
/data-analysis	End-to-end R analysis with publication-ready output
/learn	Extract non-obvious discoveries into persistent skills
/context-status	Show session health and context usage
/deep-audit	Adversarial audit of a theory, proof, paper, codebase or set of claims — decompose, independent skeptics, a separate judge per finding, fix, re-verify; repo-wide consistency is one application
/permission-check	Diagnose permission layers when prompts fire unexpectedly
/audit-reproducibility	Enforce tolerance thresholds on paper ↔ code numeric claims
/new-diagram	Scaffold a TikZ diagram from the snippet gallery with prevention + review
/respond-to-referees	R&R response-letter generator (maps referee comments to revisions)
/seven-pass-review	Seven-pass adversarial manuscript review (parallel forked subagents)
/checkpoint	Structured session-handoff snapshot (state + plan pointers + next actions). Companion to narrative session logs.
/preregister	Generate a preregistration document (OSF / AsPredicted / AEA RCT Registry style) from a research spec
/verify-claims (v1.7.0)	Chain-of-Verification fact-check (fresh-context verifier). HIGH/MED/LOW-WARN severity tiers (v1.9.0); HIGH-WARN fails closed in the report (the draft is never presented as verified).
/humanize (v1.9.0)	Detect AI-voice tells in academic prose (10 detection categories; read-only, no rewrite)
/compress-session (v1.9.0)	Distil current session into structured notes (decisions, next actions, discarded-as-noise) before auto-compaction
/promote-memory (v1.9.0)	Five-critic council that votes on which [LEARN] entries graduate from native auto memory to MEMORY.md
/stata-replication (v1.9.0)	End-to-end Stata pipeline via the stata-mcp MCP server (mirrors /data-analysis for R-first projects)
/simulation-study (v1.10.0)	Scaffold + run a reproducible Monte Carlo study — parameterized DGP, estimator grid, seeded replications, bias/RMSE/coverage/size/power with Monte Carlo SEs
/r-package-check (v1.10.0)	R package release gate — devtools::document() + tests + R CMD check --as-cran, triage ERROR/WARNING/NOTE vs CRAN policy, r-package-reviewer pass
/replication-package (v2.0)	Assemble a submission-ready DCAS / openICPSR replication package — standard README, dataset manifest, computational-requirements capture, Table/Figure → script:line map, confidential-data deposit note (blocks on /audit-reproducibility FAIL)
/challenge	Stress-test a finding against the choices you didn't make — specification curve over the discrete forks, then named sensitivity statistics (E-value, Cinelli–Hazlett RV, Oster δ) against the identifying assumption
/capture-environment (v2.0)	Snapshot the computational environment for a replication package — renv.lock + sessionInfo.txt (R), requirements.txt / environment.yml / uv.lock (Python), Stata version + ado list, seeds/RNG, optional pinning Dockerfile
/power-analysis (v2.0)	Power / required-N / minimum-detectable-effect for study design — two-arm RCT (clustering/ICC, unequal allocation), multi-arm corrections, simulation-based power for non-standard designs; feeds /preregister
/disclosure-check (v2.0)	Statistical-disclosure-limitation pre-screen for restricted/confidential-data outputs (small cells, complementary-suppression gaps, dominance, PII); CRITICAL/WARNING/OK + gate
/grant-proposal (v2.0)	Scaffold an NSF/NIH/ERC/foundation grant proposal by composing primitives (spec → aims/methods, delegated DMP + facilities, coherence pass + requirements checklist)
/data-management-plan (v2.0)	Funder-compliant Data Management Plan (NSF / NIH DMS 2023 / ERC / Horizon Europe) — folds in disclosure-avoidance + IRB constraints and a replication-package/environment plan; outputs a draft + funder checklist
/coauthor-brief (v2.0)	Collaborator handoff brief — what changed since last brief, per-artifact state, open questions, reproduce-locally + restricted-data access steps
/triage-inbox (v2.0)	Schedulable academic inbox + calendar triage via Gmail/Calendar MCP — classifies referee requests, R&R/editor, co-author threads, seminar/conference invites, grant/admin deadlines; proposes one human-gated action each (draft reply, calendar hold, scaffold a referee project, /coauthor-brief, snooze); emits a digest + referee-obligations tracker; degrades gracefully when MCP is absent; never auto-sends
/diagnose (v2.0)	Root-cause a wrong/failing empirical result — disciplined reproduce → minimise → hypothesise → instrument → fix loop; tuned for research-code bugs (type coercion, NA/merge blow-ups, clustering/SE choice, seed/package-version drift); --no-fix localizes without editing
/submission-disclosures (v2.1)	The submission-time disclosure block: AI-use disclosure matched to the target journal's verified-current policy, CRediT contributor roles, conflict-of-interest, and data-availability statements (NOT statistical disclosure — that's /disclosure-check)
/syllabus (v2.0)	Build/restructure a course syllabus from a topic or reading list — course description + prerequisites, week-by-week schedule (topic→readings→deliverables), measurable learning objectives, assessment scheme + rubric, standard policies (late work / AI use / integrity / accessibility), and a per-week work-list mapping weeks to /create-lecture decks; economics-aware (PhD metrics/micro/macro sequences, undergrad)
/teach-from-paper (v2.0)	Reads a paper end-to-end and pitches it to a stated audience level — lecture outline (motivation → setup → key result → method → takeaways), the 3-5 results worth presenting with intuition, a slide skeleton for /create-lecture, discussion questions, and a problem-set brief for /scaffold-exercises
/respond-to-eval (v2.0)	Teaching analogue of /respond-to-referees — clusters course-eval comments into themes, weights by frequency (signal vs noise), classifies Keep / Change / Investigate / Out-of-scope, and drafts concrete changes mapped to the syllabus + slide decks; saves the plan to quality_reports/teaching/
/scaffold-exercises (v2.0)	Scaffold a graded problem set across analytical/empirical/coding types, with worked solutions and "why this matters" explainers emitted to a separate solution key
/new-skill (v2.0)	Scaffold a new skill that follows this repo's conventions — interviews for purpose, triggers, and tools, writes .claude/skills/<name>/SKILL.md from the template with frontmatter/body that pass check-skill-integrity.py first try, then reminds to add the surface-table rows
/issues (v2.6)	GitHub issues as the project's to-do list and memory — files each confirmed finding as its own issue after checking open and closed issues for a duplicate, lists what is open for the work at hand, and closes an issue with a comment recording what was wrong, what was tried, and how it was fixed; shows every draft and posts nothing without a yes; warns before posting to a public repository

Research Workflow

Feature	What It Does
Exploration folder	Structured explorations/ sandbox with graduate/archive lifecycle
Fast-track workflow	60/100 quality threshold for rapid prototyping
Simplified orchestrator	implement → verify → score → done (no multi-round reviews)
Enhanced session logging	Structured tables for changes, decisions, verification
Merge-only reporting	Quality reports at merge time only
Math line-length exception	Long lines acceptable for documented formulas
Workflow quick reference	One-page cheat sheet at .claude/WORKFLOW_QUICK_REF.md

Rules (.claude/rules/)

Rules use path-scoped loading: always-on rules load every session; path-scoped rules load only when Claude works on matching files. Adherence degrades as instruction files grow (official guidance: keep CLAUDE.md under ~200 lines), so less is more.

Always-on (no paths: frontmatter — load every session):

Rule	What It Enforces
plan-first-workflow	Plan mode for non-trivial tasks + context preservation
session-logging	Three logging triggers: post-plan, incremental, end-of-session
progress-reports	GitHub as memory — issues as defect memory, quality_reports/ as work memory, MEMORY.md as lesson memory
repo-hygiene	Scratch must not become main — enforced by check-repo-hygiene.py on every commit
prompt-shaping (v2.0)	Ambient habit — shape informal/ambiguous requests before acting (replaces the retired /prompt + /prompt-only skills)

Path-scoped (load only when working on matching files):

Rule	Triggers On	What It Enforces
orchestrator-protocol	.claude/skills/, .claude/agents/, orchestration refs	Goal-first review runtime: fan-out → reduce → judge (+ hallucination gate) → loop-until-dry — loads while fan-out skills and reviewer agents are written; the skills link it directly
meta-governance	top-level docs, guide/, docs/, templates/, .claude/, gate scripts	Template vs. working project distinctions — maintainer work only
verification-protocol	.tex, .qmd, docs/	Task completion checklist
single-source-of-truth	Figures/, .tex, .qmd	No content duplication; Beamer is authoritative
quality-gates	.tex, .qmd, *.R	80/90/95 scoring + tolerance thresholds
r-code-conventions	*.R	R coding standards + math line-length exception
tikz-visual-quality	.tex	TikZ diagram visual standards
beamer-quarto-sync	.tex, .qmd	Auto-sync Beamer edits to Quarto
pdf-processing	master_supporting_docs/	Safe large PDF handling
proofreading-protocol	.tex, .qmd, quality_reports/	Propose-first, then apply with approval
no-pause-beamer	.tex	No overlay commands in Beamer
replication-protocol	*.R	Replicate original results before extending
knowledge-base-template	.tex, .qmd, *.R	Notation/application registry template
orchestrator-research	*.R, explorations/	Simple orchestrator for research (no multi-round reviews)
exploration-folder-protocol	explorations/	Structured sandbox for experimental work
exploration-fast-track	explorations/	Lightweight exploration workflow (60/100 threshold)
tikz-prevention (v1.3.0)	Slides/**, Figures/**, Preambles/**	TikZ pre-flight grep checks (P3/P4 collision avoidance)
agent-authored-code (v2.5)	**/*.sh, **/*.py, **/*.R, **/*.do	The bugs are usually ours: dry-run before any bulk edit, resolve paths before cd, monitor by PID file not pgrep, cover every terminal state
writing-with-ai (v2.5)	**/*.tex, **/*.qmd, **/*.md, **/*.Rmd	Internal vs external-facing documents; why a model cannot make its own output stop reading as model output; the human-readable standard
issue-ledger (v2.5)	.github/**	Which work gets a GitHub issue (shipped bugs, deferred improvements, blocking owner decisions, one summary issue per review round; never restricted data), the evidence standard — denominator, positive/negative control, explicit non-scope — and the closing comment
tikz-measurement (v1.3.0)	Slides/**, Figures/**, Preambles/**, scripts/**	Bézier curve depth math + 6-pass collision protocol (from MixtapeTools)
content-invariants (v1.6.x)	.tex, .qmd, Preambles/, scripts/R/**	Numbered invariants INV-1–INV-12 (palette sync, Beamer↔Quarto notation parity, no overlays, at most two boxes per slide, set.seed once, relative paths, …) that critics cite by number
cross-artifact-review (v1.4.0)	master_supporting_docs/, .tex, .qmd	Paper ↔ code dependency graph; auto-invokes /review-r + /audit-reproducibility
post-flight-verification (v1.7.0)	Skills generating factual claims	Chain-of-Verification protocol with a fresh-context verifier
summary-parity (v1.7.0)	CHANGELOG.md, README.md, .qmd, skill/rule/agent .md	Anti-whack-a-mole: re-verify summaries against their bodies
model-routing (v1.9.0)	.claude/agents/**/*.md, .claude/skills/**/SKILL.md	70/20/10 architect/editor split (Haiku/Sonnet/Opus)
review-fencing (v2.5.1)	.tex/.qmd, master_supporting_docs/, quality_reports/ audit dirs, claim-verifier.md	Reviewer independence is a property of the environment — neutral copy outside the checkout, prior verdicts excluded, own reading first, positive controls fenced from committed answer keys
stata-code-conventions (v1.9.0)	**/*.do, scripts/stata/**	Stata header scaffold, numbered pipeline, esttab, clustering discipline, AEA compliance
simulation-conventions (v1.10.0)	**/*simulation*.R, **/*_sim.R, explorations/**	Monte Carlo discipline: DGP/estimand, L'Ecuyer seeding, Monte Carlo SE, coverage-vs-truth, raw-result storage
r-package-conventions (v1.10.0)	R/**, tests/**, DESCRIPTION, NAMESPACE, man/**	R package-source standards: no library() in R/, roxygen NAMESPACE, Imports/Suggests, testthat 3e, CRAN policy
confidential-data (v2.0)	data/**, **/*.dta, **/restricted/**, **/confidential/**	Restricted/IRB-data protocol: never commit raw data, disclosure clearance before release, restricted-data-safe multi-author git topology
inference-robustness (v2.0)	scripts/**/*.R, scripts/**/*.do, scripts/**/*.py	Multiple-testing (FWER/Romano-Wolf vs FDR/Anderson sharpened-q, pre-register the family) + specification-curve / leave-one-out / wild-cluster-bootstrap robustness

Templates (templates/)

Template	What It Does
session-log.md	Structured session logging format
quality-report.md	Merge-time quality report format
exploration-readme.md	Exploration project README template
archive-readme.md	Archive documentation template
requirements-spec.md	MUST/SHOULD/MAY requirements framework with clarity status
constitutional-governance.md	Template for defining non-negotiable principles vs. preferences
skill-template.md	Academic skill creation template with domain-specific examples
decision-record.md	Architectural decision record (ADR) template
journal-profile-template.md	Journal profile for /review-paper --peer editor calibration
preregistration-template.md (v1.8.0)	Preregistration document scaffold (OSF / AsPredicted / AEA RCT)
passport-template.yaml (v1.9.0)	Per-paper YAML passport for numeric-claim provenance (used by /audit-reproducibility)
response-to-referees.md	R&R response document scaffold
executor-contract.md (v2.5.1)	Dispatchable goal contract for a delegated task — goal, acceptance bar, exact paths, gates it must pass, output contract, and the mechanisms the executor may refuse
triangulation-ledger.md	What each robustness check will mean, fixed and committed before it runs; joins spec-ledger.md on date and Script:line
screening-rubric.md (v2.5.1)	Screening rubric — written before the screen runs, with per-candidate evidence, an adjudication table, and a dispatcher spot-check

</details>

---

Prerequisites

Tool	Required For	Install
Claude Code 2.1.314+	Everything. The shipped settings need 2.1.295 (onFailure) and 2.1.314 (the write boundary); an older version skips them without a warning	claude.ai/install
git	Clone + version control	git-scm.com
Python 3 (3.9+), runnable as python3	The hooks and the gate scripts. Without a working python3 the hooks do not run, and gh commands are refused, because issue-guard is registered to fail closed (#179 for Windows)	Preinstalled on macOS/Linux; python.org for Windows
XeLaTeX	LaTeX compilation (Beamer HelloWorld, real lectures)	TeX Live or MacTeX
Quarto	Web slides (Quarto HelloWorld, real lectures)	quarto.org/docs/get-started
R	Figures and analysis (/data-analysis, scripts/R/ template)	r-project.org
pdf2svg	TikZ → SVG for Quarto (/extract-tikz)	brew install pdf2svg (macOS), apt install pdf2svg (Debian)
gh CLI	PR / issue workflow	brew install gh (macOS), apt install gh (Debian)
poppler (optional)	Beamer page screenshots, page-to-frame map and log digest (scripts/beamer-qa.py, run by /visual-audit)	brew install poppler (macOS), apt install poppler-utils (Debian)
Playwright for Python (optional)	Browser-measured slide QA (scripts/slide-qa.py, run by /visual-audit, /qa-quarto, /slide-excellence); drives your installed Chrome	python3 -m venv ~/.venvs/slide-qa && ~/.venvs/slide-qa/bin/pip install playwright, then export SLIDE_QA_PYTHON=~/.venvs/slide-qa/bin/python

Minimum to fork this template: Claude Code + git + Python 3 (Python is already installed on macOS/Linux).

Minimum to run the included HelloWorld demos end-to-end: add XeLaTeX (for /compile-latex HelloWorld) and Quarto (for /deploy HelloWorld).

Your real lectures may need more — R for scripts/R/ analyses, pdf2svg if you use TikZ extraction, gh CLI if you use the PR-based commit workflow, Playwright if you want slide overflow measured in a browser. ./scripts/validate-setup.sh reports whether R, the gh CLI and Playwright are installed; it does not check pdf2svg (command -v pdf2svg does).

---

Adapting for Your Field

1. Fill in the knowledge base (.claude/rules/knowledge-base-template.md) with your notation, applications, and design principles
2. Customize the domain reviewer (.claude/agents/domain-reviewer.md) with review lenses specific to your field
3. Update the color palette — this is a two-surface contract: change the HEX values at the top of both Preambles/header.tex (Beamer/TikZ) and Quarto/theme-template.scss (Quarto slides) so they agree. Then run ./scripts/check-palette-sync.sh to verify. Forgetting one surface silently produces mismatched Beamer vs. Quarto renderings. See Preambles/README.md for the full contract and the TikZ style library.
4. Add field-specific R pitfalls to .claude/rules/r-code-conventions.md
5. Fill in the lecture mapping in .claude/rules/beamer-quarto-sync.md
6. Customize the workflow quick reference (.claude/WORKFLOW_QUICK_REF.md) with your non-negotiables and preferences
7. Set up the exploration folder (explorations/) for experimental work

---

Additional Resources

• Claude Code Documentation
• Writing a Good CLAUDE.md — official guidance on project memory

---

Origin

This infrastructure was extracted from Econ 730: Causal Panel Data at Emory University, developed by Pedro Sant'Anna using Claude Code over 6+ sessions. The course produced 6 complete PhD lecture decks with 800+ slides, interactive Quarto versions with plotly charts, and full R replication packages — all managed through this multi-agent workflow. The patterns are domain-agnostic: the same agents, rules, and orchestrator work for any academic project.

---

Community & Extensions

Research groups across economics, energy, political science, and engineering have forked and adapted this workflow — 15+ research groups at the March 2026 survey; the repository has since passed 3,145 forks and 1,680 stars (GitHub, 2026-11-04 — a fork is not a research group, so read the survey figure and the fork count as different things). The infrastructure (orchestrator, hooks, quality gates) transfers without modification.

Extended workflows:

• clo-author by Hugo Sant'Anna (UAB) — Paper-centric research workflows (release 26.09) with 25 specialized agents and 18 skills (adversarial worker-critic pairs plus referees, data-engineer, verifier), script-enforced submission gates, /coherence internal-consistency audit, /explain author comprehension gate, simulated blind peer review, and AEA replication compliance. The /review-paper --peer <journal> pipeline in this template is adapted from clo-author with Hugo's permission (pipeline shape, 6-way disposition taxonomy, journal-calibration schema, paper-type branching). Thanks, Hugo.
• claudeblattman by Chris Blattman (U Chicago) — Comprehensive guide for non-technical academics: executive assistant workflows, proposal writing, parallel critics with a separate synthesis, and self-improving configuration
• MixtapeTools by Scott Cunningham (Baylor) — The Rhetoric of Decks, Referee 2 (a five-audit replication protocol whose audits include an R ↔ Stata ↔ Python cross-language replication), and Blindspot (a "peripheral vision" audit of empirical output)
• Paul Goldsmith-Pinkham's research tools (Yale) — claude-container (sandboxed Docker execution), exhibit-review (spatial table/figure review), and margin (source-first writing review)
• posit-dev/skills by Posit (MIT) — Skills for R packages, Quarto and accessibility, including alt-text for figure alt text in Quarto documents
• awesome-ai-for-economists by Lu Han (Bank of Canada) — Curated directory of AI workflows, MCP servers (mcp-for-stata, FRED, World Bank, Census), and replication tools for economists
• autoresearch by Andrej Karpathy — Constraint-based autonomous research with program.md as constitutional document
• Claude-Code-Presentation (formerly ClaudeCodeTools) by Alessandro Spina (UTS) — "The Editor" persona: seven-audit sequential paper review protocol

See the guide's ecosystem section for detailed descriptions, design principles, and more resources.

---

Versioning & Contributing

• What's new: see CHANGELOG.md. We follow loose semver — breaking changes get major bumps so you can decide when to pull updates.
• How to contribute: see .github/CONTRIBUTING.md. Pull requests are open to collaborators only — propose a generalizable improvement in an issue; fork-specific work stays in your fork.
• Pin to a version: git checkout $(git describe --tags --abbrev=0) pins the newest release tag; CHANGELOG.md says what each one contains.

---

License

MIT License. See LICENSE.