From research
Use when Sidd is about to submit, share, or publish a research artifact (paper, report, blog post, thesis chapter, slides) and wants a final integrity sweep. Orchestrates citation-check, replication-pack, and a structured manuscript audit, blocking on any unresolved issue. Activates on phrases like "ready to submit", "final version", "before I submit", "pre-submission check", "is this ready", "submission deadline", or any signal that the user is finishing a research artifact. Should also be PROACTIVELY OFFERED whenever the user appears to be wrapping up a research project — ask "want to run a pre-submission check?" before they proceed.
How this skill is triggered — by the user, by Claude, or both
Slash command
/research:pre-submission-checkThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
You are doing the final pass on a research artifact before Sidd submits or publishes it. This skill is the integrity gate. **Default verdict is BLOCK until every check passes or is explicitly waived with a documented reason.**
You are doing the final pass on a research artifact before Sidd submits or publishes it. This skill is the integrity gate. Default verdict is BLOCK until every check passes or is explicitly waived with a documented reason.
This skill orchestrates the others. It does not duplicate their work — it invokes them and aggregates their verdicts.
run-ids for AI experiments referenced in the paper.If the user invokes this skill without a manuscript, ask for the path. Don't proceed on memory of "the paper we were working on".
Walk through these checks in order. For each, produce a verdict: PASS / WARN / FAIL / SKIPPED (with reason). Don't continue past a FAIL until the user has resolved it or explicitly waived it.
Section X, Figure Y, Table Z, Equation N has a target. Use Grep on the manuscript source.UNVERIFIED or TODO markers remain. Search for UNVERIFIED, TODO, FIXME, XXX, <!-- ... --> HTML comments, \todo{}, and similar. Any hits → FAIL with a list of locations.citation-check)Run citation-check over the entire bibliography. Specifically:
This is the most expensive check. Tell the user upfront: "claim grounding requires reading every cited paper for every claim — this will take a while and may make API calls. Confirm to proceed."
Aggregate the results. The verdict for this section is FAIL if any citation fails any check. WARN for metadata-only mismatches the user might want to accept.
replication-pack)For every AI-related claim, number, or output reported in the manuscript:
ai-analysis run that produced this number? If the manuscript reports "Claude Opus achieved 87.3% on benchmark X", there must be a logged run that shows that result. Use the format-based search from replication-pack Step 1 to locate raw-call data — don't assume any directory layout. If no matching run can be found anywhere in the project, FAIL with "claim is not backed by a logged experiment". If raw-call data exists but the user keeps it somewhere unconventional, surface that and confirm with the user before treating it as the source.replication-pack Step 2.replication-pack. Citing experiments without a reproducibility bundle is a form of unverifiability._index.json (or generated equivalent) and the located cost log. Discrepancies → FAIL.claude-opus-latest, gpt-5) cited as the model used; only dated snapshots.ai-analysis)If the manuscript makes any cross-model comparison claim, verify against the run's parity table:
ai-analysis, that waiver MUST appear in the paper's Limitations / Methods / Caveats section. Hidden waivers → FAIL.environment.json from each replication pack is consistent with what the methodology section claims (same library versions, same OS family if relevant).Use WebSearch to fetch the current call-for-papers / submission-guidelines page for the venue. Check:
If no venue is specified, SKIP this section with a note.
Produce a single markdown report <manuscript-name>-pre-submission-<date>.md next to the manuscript itself, or in whatever output directory the project uses. Tell the user the chosen path.
Auto-gitignore the report's parent directory before writing, per the canonical rule in ai-analysis SKILL.md § Auto-gitignore. The report quotes manuscript content and may contain unpublished material.
# Pre-submission check: <manuscript title>
**Manuscript path:** <path>
**Target venue:** <venue or "not specified">
**Run on:** <date>
**Overall verdict:** PASS / BLOCK
## Section verdicts
| Section | Verdict | Issues |
|---------|---------|--------|
| 1. Manuscript integrity | PASS | — |
| 2. Citations | BLOCK | 3 grounding failures, 1 missing DOI |
| 3. AI experiments | WARN | 1 metric not backed by logged run |
| 4. Parity claims | PASS | — |
| 5. Reproducibility | PASS | — |
| 6. Venue rules | SKIPPED | no venue specified |
| 7. Hygiene | PASS | — |
## Detailed findings
<list every issue with file/line, what it is, what to do>
## Required actions before submission
<numbered checklist of every BLOCK and WARN, with concrete fixes>
## Waivers carried into submission
<any waiver the user accepted, with the reason — these go in the paper's caveats section>
When the overall verdict is BLOCK, do not let the user proceed without acknowledgement:
Read / Grep — to scan the manuscript.WebSearch / WebFetch — for venue rules.Write — for the report.Even when not explicitly invoked, OFFER this skill when:
The offer is a single line: "Want me to run a pre-submission check before you proceed?" Don't run it without consent — it's expensive and reads many files.
npx claudepluginhub siddvoh/sidds-claude-plugins --plugin researchGuides creation, editing, and verification of skills for AI coding agents using test-driven development with subagent scenarios. Use when authoring or debugging skills.