Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Router and workflow hub for the Research Project Template — code, tests, manuscripts, pipeline, validation, docs, and reproducibility. USE WHEN the user wants to audit the repo, debug ./run.sh or execute_pipeline.py, add a feature, write tests, scaffold a manuscript, verify claims, fix cross-references, run validation gates, refactor, or prepare a Zenodo release, discover relevant template skills, improve agent routing, review external agentic operating models or retrieval references such as Ste
.claude/skills/docxology-template-workflows/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-18 | ✗→✓ | ▲ Improved | 69% | 0% |
| case-05 | ✗→✓ | ▲ Improved | 30% | 0% |
| case-06 | ✗→✓ | ▲ Improved | 34% | 0% |
| case-07 | ✗→✓ | ▲ Improved | 61% | 0% |
| case-08 | ✗→✓ | ▲ Improved | 60% | 0% |
When intent is clear, open the matching child SKILL.md directly. When ambiguous, use the routing table, then load exactly one child skill.
[[FIG:]] / labels.yaml mentioned.| Symptom or goal | Child skill | | --- | --- | | Fresh clone, "start", "set up the repo", first-time install + validation | startup/SKILL.md | | Research question, literature review, fact check, systematic review | deep-research/SKILL.md | | Draft, outline, revise, format, or disclose AI use in a manuscript | academic-paper/SKILL.md | | Peer review, methodology review, re-review, or calibration | academic-paper-reviewer/SKILL.md | | Research → write → verify → review → revise → finalize | academic-pipeline/SKILL.md | | Methods, methodology, stage contracts, artifact/evidence wiring | methods-orchestration/SKILL.md | | Pipeline stage failed, stuck, or flaky | pipeline-debugging/SKILL.md | | Regenerate-from-clean / determinism / double-run diff | reproducibility-audit/SKILL.md | | Publication-readiness sign-off across tests, evidence, methods, artifacts, and renders | publication-audit/SKILL.md | | Triple-check every manuscript claim; pre-submission | manuscript-claim-verification/SKILL.md | | [[FIG:]] / labels.yaml registry audit | manuscript-cross-references/SKILL.md | | New manuscript + project from research brief | manuscript-creation/SKILL.md | | Literature search corpus synthesis (LLM blocks) | literature-synthesis/SKILL.md | | Full repo audit (tests, docs, manuscript, pipeline) | comprehensive-assessment/SKILL.md | | Skill inventory, agent onboarding, routing hardening, external skill, Steward OS-style operating-model review, AutoResearch CLI measurement patterns, or LEANN semantic-retrieval guidance | agentic-use/SKILL.md | | Validation CLI, gates, markdown/PDF checks | validation-quality/SKILL.md | | New module or algorithm (thin orchestrator) | code-development/SKILL.md | | Tests under no-mocks policy | test-creation/SKILL.md | | End-to-end feature across layers | feature-addition/SKILL.md | | Clean-break refactor with migration | refactoring/SKILL.md | | New infrastructure/* package | infrastructure-module/SKILL.md | | AGENTS.md / README.md for a directory | documentation-creation/SKILL.md |
New project scaffold (one-shot): docs/guides/new-project-one-shot-prompt.md — not a separate skill; link from manuscript-creation when starting from zero.
docs/_generated/active_projects.md — never hard-code rotating paths.infrastructure/ or projects/<name>/src/ only; scripts orchestrate.docs/_generated/COUNTS.md.To assemble a custom subset of pipeline stages (discover → fetch → synthesize → write → verify → review → validate → publish) instead of running the whole --pipeline, follow the Workflow Composition Map.
These skills are indexed in .cursor/skill_manifest.json (uv run python -m infrastructure.skills write). Human index: README.md.
This workflow set includes original, template-native adaptations of patterns observed in Imbad0202/academic-research-skills: mode registries, data-access labels, provenance checkpoints, and benchmark disclosure. ARS content is not vendored or copied because ARS is CC-BY-NC-4.0 while this repository is Apache-2.0.
Several ARS ideas are now backed by deterministic infrastructure (code, not prompts), in keeping with the template's "Code Before Prompts" principle:
infrastructure/reference/verificationresolves each cited DOI/arXiv id/title against Crossref/OpenAlex/arXiv and flags fabricated/mismatched/anachronistic citations. Offline-first with a persistent SQLite cache; live resolution is opt-in. This is the anti-leakage tier-0 gate distilled from the ARS hallucination taxonomy and verification cache. CLI: python -m infrastructure.reference.verification verify <bib>.
infrastructure/validation/content/ai_writing.pyscores em-dash density, stock LLM phrasing, and sentence-length burstiness. CLI: python -m infrastructure.validation.cli prose-quality <path>.
These run fully offline in CI; only the optional --live resolution pass touches the network.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-18 | fail→pass | 10,504 | 5,855 | -44% | 1 | 1 | 0% | 1,653 | 2,791 | +69% | 0 | 0 | — |
case-01 | fail→fail | 34,160 | 7,217 | -79% | 1 | 1 | 0% | 5,686 | 2,288 | -60% | 0 | 0 | — |
case-02 | fail→fail | 4,081 | 4,614 | +13% | 1 | 1 | 0% | 240 | 2,033 | +747% | 0 | 0 | — |
case-03 | fail→fail | 3,566 | 4,866 | +36% | 1 | 1 | 0% | 230 | 2,072 | +801% | 0 | 0 | — |
case-04 | fail→fail | 9,209 | 3,835 | -58% | 1 | 1 | 0% | 1,715 | 2,429 | +42% | 0 | 0 | — |
case-05 | fail→pass | 10,565 | 2,998 | -72% | 1 | 1 | 0% | 1,804 | 2,338 | +30% | 0 | 0 | — |
case-06 | fail→pass | 10,878 | 4,116 | -62% | 1 | 1 | 0% | 1,673 | 2,248 | +34% | 0 | 0 | — |
case-07 | fail→pass | 9,184 | 5,136 | -44% | 1 | 1 | 0% | 1,626 | 2,611 | +61% | 0 | 0 | — |
case-08 | fail→pass | 8,496 | 3,305 | -61% | 1 | 1 | 0% | 1,425 | 2,287 | +60% | 0 | 0 | — |
case-09 | fail→pass | 9,764 | 2,794 | -71% | 1 | 1 | 0% | 1,607 | 2,216 | +38% | 0 | 0 | — |
case-10 | fail→pass | 13,321 | 6,602 | -50% | 1 | 1 | 0% | 2,071 | 2,483 | +20% | 0 | 0 | — |
case-11 | fail→pass | 9,888 | 3,553 | -64% | 1 | 1 | 0% | 1,409 | 2,239 | +59% | 0 | 0 | — |
case-12 | fail→pass | 7,292 | 3,341 | -54% | 1 | 1 | 0% | 1,368 | 2,428 | +77% | 0 | 0 | — |
case-13 | fail→pass | 6,118 | 2,611 | -57% | 1 | 1 | 0% | 1,131 | 2,153 | +90% | 0 | 0 | — |
case-14 | fail→pass | 12,475 | 5,022 | -60% | 1 | 1 | 0% | 2,205 | 2,720 | +23% | 0 | 0 | — |
case-15 | pass→pass | 9,870 | 3,754 | -62% | 1 | 1 | 0% | 1,702 | 2,447 | +44% | 0 | 0 | — |
case-16 | fail→pass | 11,377 | 3,185 | -72% | 1 | 1 | 0% | 1,881 | 2,344 | +25% | 0 | 0 | — |
case-17 | pass→pass | 15,895 | 7,349 | -54% | 1 | 1 | 0% | 2,277 | 3,127 | +37% | 0 | 0 | — |
case-19 | fail→pass | 22,580 | 3,138 | -86% | 1 | 1 | 0% | 2,047 | 2,293 | +12% | 0 | 0 | — |
case-20 | fail→pass | 8,998 | 4,590 | -49% | 1 | 1 | 0% | 1,630 | 2,460 | +51% | 0 | 0 | — |
case-21 | fail→pass | 8,556 | 3,147 | -63% | 1 | 1 | 0% | 1,448 | 2,215 | +53% | 0 | 0 | — |
case-22 | fail→pass | 8,906 | 6,677 | -25% | 1 | 1 | 0% | 1,506 | 2,709 | +80% | 0 | 0 | — |
case-23 | pass→pass | 8,227 | 5,644 | -31% | 1 | 1 | 0% | 1,631 | 2,817 | +73% | 0 | 0 | — |
case-24 | fail→pass | 36,224 | 18,368 | -49% | 1 | 1 | 0% | 3,470 | 4,714 | +36% | 0 | 0 | — |
case-25 | fail→fail | 13,838 | 10,737 | -22% | 1 | 1 | 0% | 2,933 | 3,897 | +33% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 25 cases were attempted, and 22 counted toward the lift figure. The other 3 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +68 percentage points is the difference between those two pass rates over the 22 comparable cases.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.