[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7EVCCE1M145K3M4XK26PSPS":3,"skill-same:01M7EVCCE1M145K3M4XK26PSPS":52},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":24,"commitSha":25,"skillDirectory":26,"forksCount":14,"repositoryHtmlUrl":27,"repositoryHomepageUrl":6,"ownerHtmlUrl":28,"ownerBlogUrl":29,"ownerDescription":30,"updatedAt":31,"install":32,"resources":45},"01M7EVCCE1M145K3M4XK26PSPS","fp-finder-latex",null,"Run bounded independent review and revision for a mathematical prototype, filled presentation or LaTeX manuscript package. Fix the stage criterion, preserve exact versions, findings and usage, and report unresolved outcomes honestly. Simulated review, not guaranteed acceptance.","tdupu\u002Fagent-skills-public","tdupu","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F3440403?v=4","01M7EV32X9J0EFE5T4P8TKQTCQ","Taylor Dupuy","01M7EV32XZ1Z4MDBYAFRCJD3BP",0,"GPL-3.0","https:\u002F\u002Fgithub.com\u002Ftdupu\u002Fagent-skills-public\u002Fblob\u002Fmain\u002Fskills\u002Ffp-finder-latex\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: fp-finder-latex\ndescription: Run bounded independent review and revision for a mathematical prototype, filled presentation or LaTeX manuscript package. Fix the stage criterion, preserve exact versions, findings and usage, and report unresolved outcomes honestly. Simulated review, not guaranteed acceptance.\n---\n\n# fp-finder-latex\n\nParent: [LaTeX workflows](..\u002F..\u002FREADME.md).\n\nYou are the coordinator. Use separate native agents as referees and revisors;\nsequential dispatch is sufficient. Judge arguments, not author reputation.\nThe result is a simulated review with stated coverage, never human approval\nor formal verification. Keep a scientifically useful **WORKING** revision\neven when further findings remain; only independent review can make an exact\nsubmission **ACCEPTED** under its declared criterion.\n\n## Stage criterion\n\nWhen called by the [research-file synthesis workflow](..\u002F..\u002F..\u002F..\u002Fskills\u002Fmath-workflow\u002Freferences\u002Fsynthesis.md),\nrecord one of these distinct tracks before dispatch:\n\n- **Outline:** a prototype\u002Fspec, including Markdown when the project forbids\n  scratch TeX. Judge audience, exact targets, dependency order, source coverage,\n  negative findings and explicitly mapped gaps. Open proof obligations are\n  allowed only as named research gaps, outside the outline's proof-verification\n  scope. Missing or misleading gap accounting remains a blocking review finding;\n  outline acceptance never certifies the unproved claims.\n- **Presentation:** the filled candidate and its selected source\u002Fevidence map.\n  Judge the promised self-containedness, proofs, imports and narrative. A\n  Markdown candidate with LaTeX excerpts may receive source-level approval;\n  build\u002FPDF checks remain NOT RUN until the integrated artifact exists.\n- **Integrated manuscript:** the declared TeX root and input closure after\n  insertion. Check the actual build\u002FPDF and affected mathematical interfaces;\n  source-level approval alone cannot accept this track.\n\nThe coordinator passes the existing selection, audience, plan and authorization.\nDo not restart intake or force a prototype to meet a finished-paper criterion.\nUse the same independent referee\u002Frevisor loop below, adapted to the recorded\nstage: inventory planned items for outlines and actual claims for presentations.\nFreeze all reviewed inputs, including the selection\u002Fspec; altered inputs require\nreview of affected dependencies. Label any acceptance with its stage and hash.\nOrdinary manuscript review continues to require final build\u002FPDF checks.\n\n## Preflight and contract\n\nRead the nearest project instructions and canonical-file contracts. Load\n[using-latexpowers](..\u002Fusing-latexpowers\u002FSKILL.md),\n[using-mathpowers](..\u002F..\u002F..\u002Flean\u002Fskills\u002Fusing-mathpowers\u002FSKILL.md),\n[referee-report](..\u002Freferee-report\u002FSKILL.md) and the relevant shared\n[writer mechanics](..\u002F..\u002FWRITERS.md). Use their scoped proof\u002Fsource\u002Fwriting\nleaves in workers, without restarting the whole intake in every worker.\nKeep mathematical claim states in the existing\n[LEDGER contract](..\u002F..\u002F..\u002Frepo-docs\u002FLEDGER.md); this skill owns review\nhistory, not a competing proof-status system.\n\nBefore expensive work, verify inputs and primary sources, distinct-agent\ndispatch and any proof backend required by this track. Probe TeX\u002FPDF tools\nfor TeX targets; Markdown outline\u002Fsource review does not require them.\nRecord actual capability and model provenance.\nIf a required dependency is absent, say what is missing, the concrete setup\naction and which work it blocks; continue only independent work. Do not\ninvent a backend or call inline self-review independent. An explicitly\nauthorized inline mode must retain that limitation and cannot satisfy the\nindependent-review acceptance condition.\n\nCreate one run root in the project's permitted scratch area. Record:\n\n- User goal, canonical manuscripts, active-document boundaries, local input\n  closures and cross-paper dependencies; preserve existing work and archives.\n- Revision authorization and limits. Existing autonomous authorization uses\n  the manual referee-response route with writer mechanics and an item-to-tag\n  map. Use [revise](..\u002Frevise\u002FSKILL.md) only for its actual accepted-item-list\n  contract. A review-only request permits reports, not manuscript mutation;\n  obtain missing authorization at that boundary, without asking again for\n  authorization already granted. Real received reports retain their chosen\n  adjudication route.\n- A fixed criterion and track ID: genre, contribution bar, reader background,\n  correctness\u002Fsource standards, required coverage and package objective.\n  Do not infer a research-only venue from “high quality.” State necessary\n  assumptions; clarify a consequential unresolved choice. A later narrower\n  genre\u002Fclaim is a separate track, never retroactive closure of the original.\n- Referee\u002Frevisor identities and relevant prior authorship. A referee never\n  writes the manuscript revision it judges. Disclose same-task\u002Fmodel limits.\n- A positive round budget, default **8**, checked before every new dispatch.\n  Record any user-specified time\u002Ftoken limit. A limit is not a verdict.\n\nFor multiple papers, first assess their division and dependency graph.\nCombining or splitting is allowed within authorization, with a claim and\nfinding relocation map, stable aliases and review of affected interfaces.\nDo not improve apparent acceptance by hiding a difficult claim elsewhere.\n\n## One round\n\n1. **Freeze.** Snapshot sources, bibliography\u002Ffigures\u002Fpreambles and available\n   PDFs, with hashes and build\u002Ftool evidence. Never overwrite a reviewed\n   snapshot. Record an initial failed build or absent PDF explicitly;\n   source review may proceed, with PDF checking owed for manuscript acceptance.\n\n2. **Referee.** Dispatch a read-only reviewer using referee-report at proof\n   level. Read the source before weighing the author response. Inventory\n   active claims\u002Fproofs and catalogue entries; mark checked, imported and\n   unchecked coverage precisely. Test hypotheses, edge cases, counterexamples,\n   source editions and actual source-to-manuscript maps. Separate correctness,\n   contribution, presentation and the stronger package objective. A citation,\n   compilation or statement count is not proof. Imported theorems can be\n   legitimate journal inputs without making a package self-contained.\n\n   Each report gives ACCEPT, MINOR_REVISION, MAJOR_REVISION or REJECT under\n   the fixed criterion. Every finding has a stable ID, original objection\n   and locator, severity\u002Ftype, evidence, concrete closure requirement and\n   reviewing authority. New objections get new IDs. Preserve counterexamples\n   and source evidence. Reviewers must not manufacture objections for tone,\n   or approve from fatigue, pressure or iteration count.\n\n   Before issuing a verdict, the reviewer runs a report-completeness audit.\n   Every material objection discovered in the source, a prior report, a\n   source comparison, or an adversarial check appears in the current finding\n   table with a stable ID, exact locator, closure test, and one destination:\n   `RESOLVED-IN-REPORT`, `OPEN-TODO`, `IMPORTED-OBLIGATION`,\n   `OUT-OF-SCOPE`, or `INVALIDATED`. Findings may not\n   remain only in worker scratch or an ancillary calculation. The report\n   links each TODO to the manuscript response map. `RESOLVED-IN-REPORT`\n   requires the exact resolving passage and closure test; `OPEN-TODO`\n   requires an owner, evidence requirement, next source or computation, and\n   closure test; `IMPORTED-OBLIGATION` requires the precise source,\n   hypotheses, import boundary, and explicit permission under the fixed\n   criterion; `OUT-OF-SCOPE` requires the criterion and exclusion reason;\n   `INVALIDATED` requires the defeating evidence. ACCEPT is forbidden while\n   any material finding is `OPEN-TODO`, or while an imported row lacks those\n   source, hypothesis, boundary and criterion-permission fields.\n\n   Apply an example admissibility gate to every calculation used as evidence.\n   It must be source-native with exact hypotheses and source interfaces, or\n   interpretation-compatible with an explicit typed map into the interpretation\n   catalogue that preserves the relevant operations. The workflow does not use\n   the labels \"toy model\", \"toy example\", or \"naive calculation\". An\n   illustrative construction that fails the gate is excluded from the verdict\n   and recorded only as a rejected diagnostic, never as a counterexample to\n   the manuscript's source-level theorem.\n\n3. **Adjudicate and revise.** Check findings against evidence; a referee is\n   not automatically correct. Keep disputed\u002Fconflicting items open with\n   reasons until resolved by mathematical evidence or an independent focused\n   adjudicator; seek the user only for a genuinely missing choice. Within\n   existing authorization, a different agent makes tagged, bounded revisions\n   and maps every finding to an edit\u002Fevidence locator or an explicit unresolved\n   response. The revisor proposes closure, never awards it. Preserve human\n   comments and dormant material under project rules. Prove new claims or\n   mark their unresolved status; don't insert a plausible argument as proved.\n   The response map must cover the referee's complete finding table and every\n   linked investigation TODO. An imported obligation may remain open only\n   when the fixed criterion expressly permits that import and the report gives\n   its precise source, hypotheses, and boundary.\n\n4. **Validate a working candidate.** Check the coherent diff and complete\n   item accounting. For TeX targets, build changed roots with isolated\n   auxiliaries and inspect logs, references\u002Fcitations and compiled prose\u002Fformulas. TeX\n   comments can swallow text without a build error. Keep required AI\u002Fsoftware\n   statements accurate and at the project's specified introduction location.\n   Record regressions; repair\u002Frevert them or leave them explicitly unresolved.\n   A supported partial revision may be retained as WORKING, with unchecked\n   claims visible. It is not an accepted article. Freeze its next submission.\n\n5. **Return to the referee.** Give the original referee the exact new snapshot,\n   complete response map and earlier reports. Verify each item as CLOSED,\n   PARTIALLY_ADDRESSED, OPEN or RETRACTED, with reasons and revised locators.\n   A retraction retains the original finding and evidence. Check all changed\n   arguments and affected dependencies, including new mistakes. Retain limits\n   for unchanged material; do not claim a fresh complete proof audit from a\n   diff check. A changed hash creates an unreviewed version and invalidates\n   affected dependent claims, without erasing historical acceptance of the\n   old version. No acceptance silently transfers between versions or tracks.\n\n   The returned report repeats the completeness table, including findings\n   discovered during revision. A clean build or polished exposition cannot\n   close a mathematical item whose closure test was not met. If an example\n   used in the prior round fails the admissibility gate, remove it from the\n   evidentiary chain and preserve the correction in the response history.\n\n6. **Account and decide.** Update the per-round records by delegating:\n   [[update-ai-usage]] for provenance (call\u002Fagent IDs, requested versus\n   observed models and harness, tool and source use, failed and discarded\n   attempts, source-check limits) and [[update-tokens]] for the priced\n   counters. Both consolidate upward into the repository's master\n   `ai-usage.md` and `tokens.md`; root summaries reference those records.\n   Their contract — never invent totals or costs, never treat unavailable as\n   zero, label every estimate — is `AI-POLICY.md` (AI15, AI16) and is not\n   restated here. The per-round `usage.md` index stays this skill's own\n   artifact: it links the round's `ai-usage.md` and `tokens.md` rather than\n   duplicating them, and neither delegate writes or consolidates it.\n   Round-local duties that remain this skill's: do not double-count mirrors\n   and do not add overlapping intervals.\n\n   Accept only when all in-scope papers and affected package interfaces meet\n   the fixed criterion and required coverage on the exact current tuple, with\n   no mandatory unresolved items or unchecked load-bearing proofs under that\n   track. Outline and source-level acceptance carry their stated limits;\n   integrated manuscript acceptance requires final build\u002FPDF checks. Record\n   permitted imported prerequisites. A larger\n   self-containedness goal needs its own dependency closure; article approval\n   does not imply it. Further substantive repairs start another round within\n   budget, not a repeat of the same report to solicit a nicer verdict.\n\n## Non-success and completion\n\nAfter two successive rounds with the same substantive open obligations and\nno new resolving evidence, reassess strategy. Continue only with a recorded\nmaterially different approach within the remaining budget. A blocker needing\nnew research or unavailable evidence can stop its track while independent\nwork finishes. On exhaustion, unresolved disagreement or inability to progress,\nreport **NOT CONVERGED** with exact blockers and the best working snapshot;\nnever manufacture an acceptance or weaken the criterion silently.\n\nReturn manuscript paths, exact reviewed versions, per-track verdicts, closed\nand remaining findings, package dependencies and accounting links. Preserve\nall reports and rejected candidates. Clean up only task-created worktrees,\nbranches and disposable files after preserving evidence; no implicit commit,\npush, submission or publication. Simulated acceptance means only that the\nrecorded reviewers found no remaining mandatory objection under that scope.\n\nThis is a thin manuscript adapter, not the generic META-FP formula running\nunchanged. That formula already permits LaTeX growth; its APPROVING quality\nfloor differs from retaining a partial WORKING revision. To stabilize this\nskill's own text, use [fp-finder-skill](..\u002F..\u002F..\u002F..\u002Fskills\u002Ffp-finder-skill\u002FSKILL.md);\nits per-file shrink rule does not govern manuscripts.\n\n## Example and validation\n\nExample prompt: “Use fp-finder-latex on these two canonical papers as\nspecialist exposition. Revise autonomously for at most four rounds; preserve\nthe separate goal of a self-contained package.” Paths come from the caller.\n\n[Behavioral fixtures](assets\u002Fscenarios.json) are portable inputs for bounded\nagent dry runs. For the final skill hash, give a separate agent each input,\nrequest its decision and next actions under this skill, and save the prompt,\nresponse and independent comparison with the expected observable outcome.\nThe inputs are fictional tests, not real manuscript states. Distinguish\nexecuted cases from supplied-only cases. Hash\u002Faccounting checks certify only\nthose mechanics; semantic trials do not certify general mathematical judgment.\nAfter any promoted change, rerun affected cases on the final tuple.\n\n## Negative results travel with the positive ones (LX11, LX12)\n\nThis skill renders previous work into a new artifact, so the negative-result\nfloor of `..\u002F..\u002FPOLICY.md` applies. Before finishing, enumerate the inputs'\nrefutations, counterexamples, ill-posed proposals, and failed expectations, and\ncarry each into the output as a numbered statement (or a labelled question when\nunproved), naming the expectation it corrects: \"it is natural to expect $X$; in\nfact $Y$\". Report the count carried and, for anything deliberately left out,\nthe scope reason. Dropping a refuted claim together with its refutation is a\nrevise-level violation: the refutation is the surviving result.\n","# fp-finder-latex\n\nParent: [LaTeX workflows](..\u002F..\u002FREADME.md).\n\nYou are the coordinator. Use separate native agents as referees and revisors;\nsequential dispatch is sufficient. Judge arguments, not author reputation.\nThe result is a simulated review with stated coverage, never human approval\nor formal verification. Keep a scientifically useful **WORKING** revision\neven when further findings remain; only independent review can make an exact\nsubmission **ACCEPTED** under its declared criterion.\n\n## Stage criterion\n\nWhen called by the [research-file synthesis workflow](..\u002F..\u002F..\u002F..\u002Fskills\u002Fmath-workflow\u002Freferences\u002Fsynthesis.md),\nrecord one of these distinct tracks before dispatch:\n\n- **Outline:** a prototype\u002Fspec, including Markdown when the project forbids\n  scratch TeX. Judge audience, exact targets, dependency order, source coverage,\n  negative findings and explicitly mapped gaps. Open proof obligations are\n  allowed only as named research gaps, outside the outline's proof-verification\n  scope. Missing or misleading gap accounting remains a blocking review finding;\n  outline acceptance never certifies the unproved claims.\n- **Presentation:** the filled candidate and its selected source\u002Fevidence map.\n  Judge the promised self-containedness, proofs, imports and narrative. A\n  Markdown candidate with LaTeX excerpts may receive source-level approval;\n  build\u002FPDF checks remain NOT RUN until the integrated artifact exists.\n- **Integrated manuscript:** the declared TeX root and input closure after\n  insertion. Check the actual build\u002FPDF and affected mathematical interfaces;\n  source-level approval alone cannot accept this track.\n\nThe coordinator passes the existing selection, audience, plan and authorization.\nDo not restart intake or force a prototype to meet a finished-paper criterion.\nUse the same independent referee\u002Frevisor loop below, adapted to the recorded\nstage: inventory planned items for outlines and actual claims for presentations.\nFreeze all reviewed inputs, including the selection\u002Fspec; altered inputs require\nreview of affected dependencies. Label any acceptance with its stage and hash.\nOrdinary manuscript review continues to require final build\u002FPDF checks.\n\n## Preflight and contract\n\nRead the nearest project instructions and canonical-file contracts. Load\n[using-latexpowers](..\u002Fusing-latexpowers\u002FSKILL.md),\n[using-mathpowers](..\u002F..\u002F..\u002Flean\u002Fskills\u002Fusing-mathpowers\u002FSKILL.md),\n[referee-report](..\u002Freferee-report\u002FSKILL.md) and the relevant shared\n[writer mechanics](..\u002F..\u002FWRITERS.md). Use their scoped proof\u002Fsource\u002Fwriting\nleaves in workers, without restarting the whole intake in every worker.\nKeep mathematical claim states in the existing\n[LEDGER contract](..\u002F..\u002F..\u002Frepo-docs\u002FLEDGER.md); this skill owns review\nhistory, not a competing proof-status system.\n\nBefore expensive work, verify inputs and primary sources, distinct-agent\ndispatch and any proof backend required by this track. Probe TeX\u002FPDF tools\nfor TeX targets; Markdown outline\u002Fsource review does not require them.\nRecord actual capability and model provenance.\nIf a required dependency is absent, say what is missing, the concrete setup\naction and which work it blocks; continue only independent work. Do not\ninvent a backend or call inline self-review independent. An explicitly\nauthorized inline mode must retain that limitation and cannot satisfy the\nindependent-review acceptance condition.\n\nCreate one run root in the project's permitted scratch area. Record:\n\n- User goal, canonical manuscripts, active-document boundaries, local input\n  closures and cross-paper dependencies; preserve existing work and archives.\n- Revision authorization and limits. Existing autonomous authorization uses\n  the manual referee-response route with writer mechanics and an item-to-tag\n  map. Use [revise](..\u002Frevise\u002FSKILL.md) only for its actual accepted-item-list\n  contract. A review-only request permits reports, not manuscript mutation;\n  obtain missing authorization at that boundary, without asking again for\n  authorization already granted. Real received reports retain their chosen\n  adjudication route.\n- A fixed criterion and track ID: genre, contribution bar, reader background,\n  correctness\u002Fsource standards, required coverage and package objective.\n  Do not infer a research-only venue from “high quality.” State necessary\n  assumptions; clarify a consequential unresolved choice. A later narrower\n  genre\u002Fclaim is a separate track, never retroactive closure of the original.\n- Referee\u002Frevisor identities and relevant prior authorship. A referee never\n  writes the manuscript revision it judges. Disclose same-task\u002Fmodel limits.\n- A positive round budget, default **8**, checked before every new dispatch.\n  Record any user-specified time\u002Ftoken limit. A limit is not a verdict.\n\nFor multiple papers, first assess their division and dependency graph.\nCombining or splitting is allowed within authorization, with a claim and\nfinding relocation map, stable aliases and review of affected interfaces.\nDo not improve apparent acceptance by hiding a difficult claim elsewhere.\n\n## One round\n\n1. **Freeze.** Snapshot sources, bibliography\u002Ffigures\u002Fpreambles and available\n   PDFs, with hashes and build\u002Ftool evidence. Never overwrite a reviewed\n   snapshot. Record an initial failed build or absent PDF explicitly;\n   source review may proceed, with PDF checking owed for manuscript acceptance.\n\n2. **Referee.** Dispatch a read-only reviewer using referee-report at proof\n   level. Read the source before weighing the author response. Inventory\n   active claims\u002Fproofs and catalogue entries; mark checked, imported and\n   unchecked coverage precisely. Test hypotheses, edge cases, counterexamples,\n   source editions and actual source-to-manuscript maps. Separate correctness,\n   contribution, presentation and the stronger package objective. A citation,\n   compilation or statement count is not proof. Imported theorems can be\n   legitimate journal inputs without making a package self-contained.\n\n   Each report gives ACCEPT, MINOR_REVISION, MAJOR_REVISION or REJECT under\n   the fixed criterion. Every finding has a stable ID, original objection\n   and locator, severity\u002Ftype, evidence, concrete closure requirement and\n   reviewing authority. New objections get new IDs. Preserve counterexamples\n   and source evidence. Reviewers must not manufacture objections for tone,\n   or approve from fatigue, pressure or iteration count.\n\n   Before issuing a verdict, the reviewer runs a report-completeness audit.\n   Every material objection discovered in the source, a prior report, a\n   source comparison, or an adversarial check appears in the current finding\n   table with a stable ID, exact locator, closure test, and one destination:\n   `RESOLVED-IN-REPORT`, `OPEN-TODO`, `IMPORTED-OBLIGATION`,\n   `OUT-OF-SCOPE`, or `INVALIDATED`. Findings may not\n   remain only in worker scratch or an ancillary calculation. The report\n   links each TODO to the manuscript response map. `RESOLVED-IN-REPORT`\n   requires the exact resolving passage and closure test; `OPEN-TODO`\n   requires an owner, evidence requirement, next source or computation, and\n   closure test; `IMPORTED-OBLIGATION` requires the precise source,\n   hypotheses, import boundary, and explicit permission under the fixed\n   criterion; `OUT-OF-SCOPE` requires the criterion and exclusion reason;\n   `INVALIDATED` requires the defeating evidence. ACCEPT is forbidden while\n   any material finding is `OPEN-TODO`, or while an imported row lacks those\n   source, hypothesis, boundary and criterion-permission fields.\n\n   Apply an example admissibility gate to every calculation used as evidence.\n   It must be source-native with exact hypotheses and source interfaces, or\n   interpretation-compatible with an explicit typed map into the interpretation\n   catalogue that preserves the relevant operations. The workflow does not use\n   the labels \"toy model\", \"toy example\", or \"naive calculation\". An\n   illustrative construction that fails the gate is excluded from the verdict\n   and recorded only as a rejected diagnostic, never as a counterexample to\n   the manuscript's source-level theorem.\n\n3. **Adjudicate and revise.** Check findings against evidence; a referee is\n   not automatically correct. Keep disputed\u002Fconflicting items open with\n   reasons until resolved by mathematical evidence or an independent focused\n   adjudicator; seek the user only for a genuinely missing choice. Within\n   existing authorization, a different agent makes tagged, bounded revisions\n   and maps every finding to an edit\u002Fevidence locator or an explicit unresolved\n   response. The revisor proposes closure, never awards it. Preserve human\n   comments and dormant material under project rules. Prove new claims or\n   mark their unresolved status; don't insert a plausible argument as proved.\n   The response map must cover the referee's complete finding table and every\n   linked investigation TODO. An imported obligation may remain open only\n   when the fixed criterion expressly permits that import and the report gives\n   its precise source, hypotheses, and boundary.\n\n4. **Validate a working candidate.** Check the coherent diff and complete\n   item accounting. For TeX targets, build changed roots with isolated\n   auxiliaries and inspect logs, references\u002Fcitations and compiled prose\u002Fformulas. TeX\n   comments can swallow text without a build error. Keep required AI\u002Fsoftware\n   statements accurate and at the project's specified introduction location.\n   Record regressions; repair\u002Frevert them or leave them explicitly unresolved.\n   A supported partial revision may be retained as WORKING, with unchecked\n   claims visible. It is not an accepted article. Freeze its next submission.\n\n5. **Return to the referee.** Give the original referee the exact new snapshot,\n   complete response map and earlier reports. Verify each item as CLOSED,\n   PARTIALLY_ADDRESSED, OPEN or RETRACTED, with reasons and revised locators.\n   A retraction retains the original finding and evidence. Check all changed\n   arguments and affected dependencies, including new mistakes. Retain limits\n   for unchanged material; do not claim a fresh complete proof audit from a\n   diff check. A changed hash creates an unreviewed version and invalidates\n   affected dependent claims, without erasing historical acceptance of the\n   old version. No acceptance silently transfers between versions or tracks.\n\n   The returned report repeats the completeness table, including findings\n   discovered during revision. A clean build or polished exposition cannot\n   close a mathematical item whose closure test was not met. If an example\n   used in the prior round fails the admissibility gate, remove it from the\n   evidentiary chain and preserve the correction in the response history.\n\n6. **Account and decide.** Update the per-round records by delegating:\n   [[update-ai-usage]] for provenance (call\u002Fagent IDs, requested versus\n   observed models and harness, tool and source use, failed and discarded\n   attempts, source-check limits) and [[update-tokens]] for the priced\n   counters. Both consolidate upward into the repository's master\n   `ai-usage.md` and `tokens.md`; root summaries reference those records.\n   Their contract — never invent totals or costs, never treat unavailable as\n   zero, label every estimate — is `AI-POLICY.md` (AI15, AI16) and is not\n   restated here. The per-round `usage.md` index stays this skill's own\n   artifact: it links the round's `ai-usage.md` and `tokens.md` rather than\n   duplicating them, and neither delegate writes or consolidates it.\n   Round-local duties that remain this skill's: do not double-count mirrors\n   and do not add overlapping intervals.\n\n   Accept only when all in-scope papers and affected package interfaces meet\n   the fixed criterion and required coverage on the exact current tuple, with\n   no mandatory unresolved items or unchecked load-bearing proofs under that\n   track. Outline and source-level acceptance carry their stated limits;\n   integrated manuscript acceptance requires final build\u002FPDF checks. Record\n   permitted imported prerequisites. A larger\n   self-containedness goal needs its own dependency closure; article approval\n   does not imply it. Further substantive repairs start another round within\n   budget, not a repeat of the same report to solicit a nicer verdict.\n\n## Non-success and completion\n\nAfter two successive rounds with the same substantive open obligations and\nno new resolving evidence, reassess strategy. Continue only with a recorded\nmaterially different approach within the remaining budget. A blocker needing\nnew research or unavailable evidence can stop its track while independent\nwork finishes. On exhaustion, unresolved disagreement or inability to progress,\nreport **NOT CONVERGED** with exact blockers and the best working snapshot;\nnever manufacture an acceptance or weaken the criterion silently.\n\nReturn manuscript paths, exact reviewed versions, per-track verdicts, closed\nand remaining findings, package dependencies and accounting links. Preserve\nall reports and rejected candidates. Clean up only task-created worktrees,\nbranches and disposable files after preserving evidence; no implicit commit,\npush, submission or publication. Simulated acceptance means only that the\nrecorded reviewers found no remaining mandatory objection under that scope.\n\nThis is a thin manuscript adapter, not the generic META-FP formula running\nunchanged. That formula already permits LaTeX growth; its APPROVING quality\nfloor differs from retaining a partial WORKING revision. To stabilize this\nskill's own text, use [fp-finder-skill](..\u002F..\u002F..\u002F..\u002Fskills\u002Ffp-finder-skill\u002FSKILL.md);\nits per-file shrink rule does not govern manuscripts.\n\n## Example and validation\n\nExample prompt: “Use fp-finder-latex on these two canonical papers as\nspecialist exposition. Revise autonomously for at most four rounds; preserve\nthe separate goal of a self-contained package.” Paths come from the caller.\n\n[Behavioral fixtures](assets\u002Fscenarios.json) are portable inputs for bounded\nagent dry runs. For the final skill hash, give a separate agent each input,\nrequest its decision and next actions under this skill, and save the prompt,\nresponse and independent comparison with the expected observable outcome.\nThe inputs are fictional tests, not real manuscript states. Distinguish\nexecuted cases from supplied-only cases. Hash\u002Faccounting checks certify only\nthose mechanics; semantic trials do not certify general mathematical judgment.\nAfter any promoted change, rerun affected cases on the final tuple.\n\n## Negative results travel with the positive ones (LX11, LX12)\n\nThis skill renders previous work into a new artifact, so the negative-result\nfloor of `..\u002F..\u002FPOLICY.md` applies. Before finishing, enumerate the inputs'\nrefutations, counterexamples, ill-posed proposals, and failed expectations, and\ncarry each into the output as a numbered statement (or a labelled question when\nunproved), naming the expectation it corrects: \"it is natural to expect $X$; in\nfact $Y$\". Report the count carried and, for anything deliberately left out,\nthe scope reason. Dropping a refuted claim together with its refutation is a\nrevise-level violation: the refutation is the surviving result.\n",{"name":5,"description":7},"main","8d13efb5e66306aeae6e3b948bc463e7e2db1b65","skills\u002Ffp-finder-latex","https:\u002F\u002Fgithub.com\u002Ftdupu\u002Fagent-skills-public","https:\u002F\u002Fgithub.com\u002Ftdupu","https:\u002F\u002Ftdupu.github.io\u002Findex.html","Associate Professor, Department of Mathematics and Statistics, University of Vermont --- last updated Fall 2019","2026-11-13 09:07:19",{"source":8,"skillName":5,"commands":33,"prompt":43,"zipAvailable":44},[34,37,40],{"tool":35,"command":36},"npx","npx skills add tdupu\u002Fagent-skills-public --skill fp-finder-latex",{"tool":38,"command":39},"bunx","bunx skills add tdupu\u002Fagent-skills-public --skill fp-finder-latex",{"tool":41,"command":42},"pnpm","pnpm dlx skills add tdupu\u002Fagent-skills-public --skill fp-finder-latex","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Ftdupu\u002Fagent-skills-public\u002Fblob\u002Fmain\u002Fskills\u002Ffp-finder-latex\u002FSKILL.md",true,[46],{"path":47,"resourceType":48,"resourceTypeDesc":49,"sizeBytes":50,"sourceHtmlUrl":51},"assets\u002Fscenarios.json","ASSET","资产文件",4650,"https:\u002F\u002Fgithub.com\u002Ftdupu\u002Fagent-skills-public\u002Fblob\u002Fmain\u002Fskills\u002Ffp-finder-latex\u002Fassets\u002Fscenarios.json",[53,60,67,74,81,88],{"skillNo":54,"name":55,"nameZh":6,"description":56,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":57,"categories":58,"keywords":59,"publishedAt":19},"01M7EW4PT586KKE7VRCAD80Z9K","writing-skills","Use when creating new skills, editing existing skills, or verifying skills work before deployment","https:\u002F\u002Fgithub.com\u002Ftdupu\u002Fagent-skills-public\u002Fblob\u002Fmain\u002Fsuperpowers\u002Fskills\u002Fwriting-skills\u002FSKILL.md",[],[],{"skillNo":61,"name":62,"nameZh":6,"description":63,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":64,"categories":65,"keywords":66,"publishedAt":19},"01M7EW4HV5V74AF3PTV8N4T0PS","writing-plans","Use when you have a spec or requirements for a multi-step task, before touching code","https:\u002F\u002Fgithub.com\u002Ftdupu\u002Fagent-skills-public\u002Fblob\u002Fmain\u002Fsuperpowers\u002Fskills\u002Fwriting-plans\u002FSKILL.md",[],[],{"skillNo":68,"name":69,"nameZh":6,"description":70,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":71,"categories":72,"keywords":73,"publishedAt":19},"01M7EW4DNKGP5CR1RAJP2J9PHA","verification-before-completion","Use when about to claim work is complete, fixed, or passing, before committing or creating PRs - requires running verification commands and confirming output before making any success claims; evidence before assertions always","https:\u002F\u002Fgithub.com\u002Ftdupu\u002Fagent-skills-public\u002Fblob\u002Fmain\u002Fsuperpowers\u002Fskills\u002Fverification-before-completion\u002FSKILL.md",[],[],{"skillNo":75,"name":76,"nameZh":6,"description":77,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":78,"categories":79,"keywords":80,"publishedAt":19},"01M7EW4BA2QP88VBFVXDVJM7XQ","using-superpowers","Use when starting any conversation - establishes how to find and use skills, requiring skill invocation before ANY response including clarifying questions","https:\u002F\u002Fgithub.com\u002Ftdupu\u002Fagent-skills-public\u002Fblob\u002Fmain\u002Fsuperpowers\u002Fskills\u002Fusing-superpowers\u002FSKILL.md",[],[],{"skillNo":82,"name":83,"nameZh":6,"description":84,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":85,"categories":86,"keywords":87,"publishedAt":19},"01M7EW47F2JG5FBESDJNP3PKMV","using-git-worktrees","Use when starting feature work that needs isolation from current workspace or before executing implementation plans - ensures an isolated workspace exists via native tools or git worktree fallback","https:\u002F\u002Fgithub.com\u002Ftdupu\u002Fagent-skills-public\u002Fblob\u002Fmain\u002Fsuperpowers\u002Fskills\u002Fusing-git-worktrees\u002FSKILL.md",[],[],{"skillNo":89,"name":90,"nameZh":6,"description":91,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":92,"categories":93,"keywords":94,"publishedAt":19},"01M7EW44N8556Y556NX9HJT4KP","test-driven-development","Use when implementing any feature or bugfix, before writing implementation code","https:\u002F\u002Fgithub.com\u002Ftdupu\u002Fagent-skills-public\u002Fblob\u002Fmain\u002Fsuperpowers\u002Fskills\u002Ftest-driven-development\u002FSKILL.md",[],[]]