External references

Curated bibliography of external Claude Code work that informs empn-core's design — skill libraries and pipelines, frameworks and methods, community catalogues. Maintained for two reasons: to credit where patterns came from, and to give future contributors a place to look when an idea seems novel before re-deriving it from scratch.

This is a reference, not a sync target. We borrow patterns, not code.

Skill libraries & pipelines

Personal or org Claude Code skill stacks and research pipelines close to empn-core's genre. Two lines each: what it is, then how we use or read it.

• clo-author — Hugo Sant'Anna: a Claude Code scaffold that automates the empirical-paper production function across five agent stages (discover → strategy → execute → review → submit, quality-gated at ≥95). Use it: gh repo fork hugosantanna/clo-author --clone, run claude in the folder, then drive with /discover lit [topic] or /review --methods paper/main.tex.
• coarse — David Van Dijcke: a multi-stage AI peer-review pipeline (domain classify → criteria → literature search → parallel reviewers → adversarial proof-check → editorial dedupe). Already our upstream: the procedural-discipline layer and the review-steelman skill come from it. Worth re-reading once a quarter to see whether their prompts have moved on. Use it: uvx coarse-ink review paper.pdf --api-key sk-or-... for a structured markdown review.
• Stanford STORM — researches a topic, surfaces multiple perspectives via simulated expert Q&A, then writes a full report with citations (Co-STORM adds a human-in-the-loop variant in the same repo). Borrow the method, not the code: its multi-perspective pre-writing maps almost exactly onto scope-question + map-positions + the adversarial personas, and cited long-form output is what draft-brief / synthesize-literature produce. Reference implementation, not a product — a research artifact (STORM / Co-STORM papers, 2024) with low maintenance since late 2025 (last commit 2025-09; ~30k stars). For maintained OSS in this niche use GPT Researcher below; for one-shot runs, hosted deep-research tools now do the same job.
• GPT Researcher — autonomous deep research → 5–6pp report in PDF/Docx/MD. Explicitly inspired by STORM. Direct peer to the deep-research skill.
• Pedro Sant'Anna — claude-code-my-workflow — econ-focused workflow, 10 agents + 19 commands, MEMORY.md pattern for cross-session learning. Read for workflow shape, not skills to lift. (Different person from Hugo Sant'Anna above.)
• Lu Han — econ-writing-skill — one deep skill synthesising 50+ econ-writing guides; 100-point rubric, named-persona reviewers, 4-tier ranked source registry. Source of our three-persona adversarial pass and source-routing-by-writing-step.
• Chris Wiles — claude-code-showcase — a skill-eval hook scoring prompts against skills at UserPromptSubmit. Recast into our anti-ai-tone Stop hook; the pre-prompt variant is still an open candidate.

Frameworks & methods

• BMAD Method — an agentic build framework running a project from ideation through planning to implementation via specialised agents and guided workflows. Dev-oriented and broader than empn-core; read it for workflow-structuring ideas, not policy content. Use it: install into Claude / Cursor / Copilot, then start with its bmad-help skill.

Community catalogues

Awesome-lists and wikis to watch for new patterns — not sync targets.

• claudeblattman.com/resources — Chris Blattman's curated pointer list (Sant'Anna, Murchison, Killeen, Tane, Fernandez, Wiles, Mollick). A maintained index, not a skills repo.
• ai-econ-wiki — catalogues academic-economist skills; references an APSA 2026 taxonomy of econ Claude Code skills.
• awesome-econ-ai-stuff — "AI Skills for Economists" awesome-list. Thin today; the right community to watch.
• alirezarezvani/claude-skills — 337-skill mega-list. Pattern catalogue; the orchestration/, dossier, aeo, compliance-os directories repay a read.
• ComposioHQ/awesome-claude-skills — see content-research-writer for its iterate-on-citations writing loop.
• BehiSecc/awesome-claude-skills — best-curated for verifier-style and knowledge-management skills.
• conorbronsdon/avoid-ai-writing — 21 AI-writing pattern categories with substitution table. Compare against shared/anti-ai-tone.md for additive categories.
• Astro-Han/karpathy-llm-wiki — two-layer KB (immutable raw/ + synthesised wiki/) with append-only op log and a broken-cross-ref lint. Pattern for shared/ consistency checking.

Patterns adopted

The following patterns have been pulled into empn-core, each credited to its source. Listed roughly in order of adoption.

Pattern	Source	Where it lives in empn-core	Commit
Three-state provenance (Verified / Prior / Unverified) for quantitative claims, with inline citation tags on every figure in body prose	—	shared/empn/house-style.md + shared/scripts/check_inline_tags.py enforced by review-brief	7fba76c
Set-construction discipline (population source · membership predicate · completeness residual) before listing actors / sectors / MS	—	shared/_common/set-construction.md — read by scope-research, map-positions, contacts-discover, review-adversarial, scope-question	7fba76c
Load-bearing claim audit (restate falsifiably → predict → confront with case data → verdict → consequence)	—	review-adversarial § Load-bearing claim audit	7fba76c
Falsifier sentence + regime stratification on multi-regime windows	—	draft-brief § Three rules that govern the argument	7fba76c
scope-question skill — enumerate 3–5 rival framings, pin the terminus, name the falsifier before drafting	—	skills/scope-question/ (new skill)	7fba76c
Round-trip-verify modelling pattern — Python as single source of truth, generates Excel + HTML + values, verify() proves equality and catches drift	—	skills/create-model/ (new skill) + shared/scripts/empn_xlsx_helpers.py + pymodel.py engine	41647c8
Document-level provenance label (Sourced / Mixed sourcing / Reasoned priors) — line 1 of every multi-paragraph report	—	shared/empn/house-style.md § Document-level provenance label; wired into draft-brief, search-eu-file, synthesize-literature	41647c8, 8181c8c
Source-routing by writing step — map narrative-arc acts to first-tier shared source doc	gap visible in hanlulong/econ-writing-skill (ranked registry without writing-step mapping)	shared/_common/source-routing.md + pointer from narrative-arc.md	c970303
Three-persona adversarial pass (Council Negotiator / Industry Lobbyist / Academic) — three named reviewers each scanning the same draft through their own lens, two-persona convergence upgrades severity	hanlulong/econ-writing-skill (Methodologist / Field Expert / Writing Critic)	review-adversarial § Three-persona pass (deep mode)	02af355
Resolver gate caps verdict — any failure row from check_inline_tags, crossref_lookup, eurlex_resolve, or url_check caps verdict at NEEDS REVISION	—	review-brief § How to run it step 4	7e7065b
Anti-AI-tone Stop hook — passive register-drift check after every assistant response (forbidden phrases, AI lexical tells, em-dash density, didactic openers, first-person voice, …). Emits a systemMessage warning when violations fire; never blocks. Closes the gap where anti-ai-tone.md only loaded when a writing skill triggered.	ChrisWiles/claude-code-showcase .claude/hooks/ pattern, recast from UserPromptSubmit (pre-prompt) to Stop (post-response)	plugins/empn-core/hooks/hooks.json + shared/scripts/anti_ai_tone_check.py + shared/scripts/anti_ai_tone_stop_hook.py; workspace template gets a "Writing for humans — load on every turn" section	2bf72a3

Patterns considered, skipped

• hanlulong's banned-word list and anti-AI structural red flags — already covered by shared/anti-ai-tone.md with better-sourced substitution tables (Economist Style Guide, IMF Editorial Style Guide 2024, Gardner / ECA).
• Murchison / Killeen "Goals.yaml as single source of truth" — shared/ already plays this role; the Goals.yaml shape solves a personal-OS problem empn-core doesn't have.
• Ethan Mollick "Models / Apps / Harnesses" framing — audience-level explainer, not a skill pattern.
• Boris Tane "Research → Plan → Annotate → Implement" — generic developer workflow; nothing energy / policy-shaped.
• AEO post-publish citation monitor (from alirezarezvani) — would land near website-article, but high build cost for uncertain value; revisit if EMPN starts publishing for LLM-citation-visibility specifically.

Open candidates — patterns worth investigating

Not yet adopted, listed roughly in priority order:

• Weighted 100-point scoring rubric on editorial-review skills (hanlulong) — would turn diagnostic review output into actionable. Bracket thresholds (90–100 ready / 80–89 hold for prose pass / etc.) let a reader scan one number instead of reading the whole review. Proposed landing: review-structure and review-prose.
• Session lifecycle skills /spin-up and /wrap-up (Spina, APSA 2026 taxonomy) — front-load workspace context at session start, capture decisions at session end. Complements per-project skills (project-init, project-retrospective) at the per-session edge. Would live alongside the project-* family.
• Glossary skill (Spina, APSA 2026) — maintained term-and-acronym index. Real friction for new EMPN members given the acronym density (CBAM, NECP, RED III, FSR, ESABCC, MFF, ETS, OLP). Proposed landing: new glossary skill that maintains 02_Knowledge/_glossary.md.
• Chunked PDF extraction (Cunningham /split-pdf) — handles the upstream "this 600-page bundle is too big to ingest" step before map-consultation or synthesize-literature. Complement to distill-document (which compresses a single document).
• Wiki-lint over shared/ (karpathy-llm-wiki pattern) — extend audit-repo to check that every shared/ doc referenced in a SKILL.md still exists and that monitor digest cross-refs resolve. Adjacent to audit-repo, not a new skill.
• Topic-dossier with institutional memory (alirezarezvani dossier pattern) — file-level dossier that compounds across briefs, closing the gap between one-off briefs and institutional memory. Bigger build than the others; defer until after the lighter picks land.
• UserPromptSubmit skill-eval hook (Chris Wiles) — JSON rules file scoring ambiguous prompts and suggesting skills, complementing help. UX improvement, low priority. (The Stop-event sibling pattern was adopted as the anti-ai-tone hook — see Patterns adopted table.)

Maintaining this doc

Add a row to the Patterns adopted table whenever a new adoption lands; cite the commit. Move an open candidate to Patterns considered, skipped if it gets ruled out, with a one-line reason. The point is to keep both the credit trail and the institutional memory of "we already looked at X" findable.