| name | awesome-rebuttal |
| description | Global-installable, project-level academic rebuttal strategy skill for AI/ML/CV/NLP/Robotics papers. Use when authors need a workspace-local .awesome-rebuttal state folder, paper/code/review/venue-rule intake, JSON memory, snapshots, LaTeX/template handling for one-page rebuttals, reviewer stance analysis, strategy planning, experiment triage, safe author response drafting, or AC summaries under confirmed venue rules. |
Awesome Rebuttal
Use this skill as a global-installable, project-level rebuttal workspace assistant. The installed skill provides reusable procedures and assets; each actual paper/rebuttal workspace gets its own .awesome-rebuttal/ state folder for memory, snapshots, template state, and logs. Start by understanding the workspace, then collect evidence, persist memory, analyze strategy, and only then draft response text.
Operating contract
- Inspect the current workspace before content analysis.
- Create or use a project-local
.awesome-rebuttal/ state folder; never store runtime memory in the installed skill folder.
- If the workspace is empty, organize or recommend
Code/, Paper/, Reference/, and Temp/.
- If the workspace already contains files, infer the author's organization and adapt non-destructively.
- Ask how progress should be preserved:
manual_git, auto_git, or markdown_snapshot_only.
- Run the intake gate before analysis. Missing required inputs block drafting.
- Treat venue rules as user-provided or AI-searched + user-confirmed; never rely on stale built-in venue rules.
- Keep every factual claim grounded in
paper, code, review, venue_rules, user, or explicit inference.
- Never invent experiments, numbers, citations, reviewer positions, or venue permissions.
Language policy
- Interaction language: follow the user's language by default for questions, analysis reports, progress updates, and explanations.
- Submission language: final rebuttal artifacts must be written in English unless the user explicitly requests another submission language and the venue permits it.
- Memory: record this choice in
project_memory.language_policy and mirror any venue-specific exception in venue_rules.language_policy.
- Drafting rule:
11_response_writer.md, 12_template_designer.md, and 13_ac_summary_writer.md may discuss plans in the user's language, but author-response text, reviewer replies, AC summaries, OpenReview comments, and PDF rebuttal prose default to English.
- Terminology: preserve exact technical terms, metric names, method names, dataset names, and reviewer wording from the paper/reviews; translate only surrounding explanatory prose when needed.
- If the user provides Chinese strategy notes, convert them into professional English rebuttal prose rather than literal translation.
Shared questionnaire protocol
Whenever the skill hits missing, ambiguous, or confirmation-dependent information, first summarize what the user already provided and what the workspace evidence shows. Then ask a focused questionnaire instead of guessing.
Read references/core/user_questionnaire_protocol.md for the reusable questionnaire pattern. Use it especially in workspace bootstrap and intake, and reuse it later for venue-rule confirmation, experiment feasibility, versioning mode, or any strategy decision that materially changes the output.
Prefer structured choices when possible:
- single-choice for mutually exclusive paths
- multi-select for available inputs or constraints
- short text for pasted rules, reviews, paths, or URLs
- confirmation for inferred workspace maps or AI-found venue rules
Ask only for the smallest missing decision set needed for the next safe step.
Canonical response modes
Use these exact response_mode values across all memory files and capability handoffs:
openreview_per_reviewer — one reply/comment per reviewer thread.
unified_limited — one limited unified response where concerns are merged.
pdf_one_page — one-page PDF/LaTeX rebuttal.
global_comment — one global platform comment/text box.
hybrid — global summary plus per-reviewer replies.
openreview_markdown_latex — OpenReview-style Markdown comment with lightweight LaTeX math.
unknown — not confirmed yet.
Do not introduce aliases such as per_reviewer, global, global_text, one_page_pdf, or markdown_latex_hybrid in new memory. If user wording uses those terms, normalize to the canonical value and record the original wording in notes if useful.
Venue rules schema contract
Use references/memory-schemas/venue_rules.schema.json as the global rule-memory contract. Venue rules are runtime evidence, not built-in knowledge.
Every venue_rules.json should separate:
status and source: missing/user-provided/AI-found-pending-confirmation plus URL/path/retrieval notes.
response: canonical mode, platform, limits, per-reviewer/global/interactive/AC-summary permissions.
formatting: official template, PDF/LaTeX/Markdown support, figures/tables/appendix, and page-layout constraints.
content_permissions: new experiments/results, links, supplements, code links, references, revision commitments.
anonymity: anonymous requirement, self-citation, acknowledgements, and identity-risk notes.
confirmation: user confirmation timestamp, pending questions, and conflicts.
language_policy: interaction language follows user; final submission-facing prose defaults to English.
If rules are AI-searched, keep status: ai_found_pending_confirmation until the user confirms them. Unknown fields stay unknown; do not infer permissions silently.
Required workflow
Follow this order unless the user asks for a narrower capability:
- Workspace bootstrap — read
references/capabilities/00_workspace_bootstrap.md; create/use .awesome-rebuttal/, detect LaTeX environment, and if decisions are missing use references/core/user_questionnaire_protocol.md.
- Intake gate — read
references/capabilities/01_intake_gate.md; parse user-provided context first, then ask a questionnaire for blockers.
- Template management — read
17_rebuttal_template_manager.md after intake when response format requires a pdf_one_page template, openreview_per_reviewer Markdown scaffolds, global_comment Markdown, or openreview_markdown_latex comments.
- Information collection — read
references/capabilities/02_information_collection.md.
- Paper/code memory — read
03_paper_memory_builder.md and/or 04_code_memory_builder.md when paper or code context is present.
- Review indexing — read
05_review_normalizer.md; preserve raw reviews, build raw anchors, issue item index, and common issue index without strategic analysis.
- Review concern analysis — read
06_situation_analyzer.md; interpret indexed concerns semantically without priority ranking.
- Review-driven experiment planning — read
10_experiment_triage.md; generate numbered EXP-* experiment candidates from reviewer concerns and persist experiment_memory.json before final strategy planning.
- Priority and situation analysis — read
07_concern_atomizer.md; rank concern importance, classify rebuttal posture, build reviewer priority map, link P0/P1 evidence gaps to numbered experiments, and identify AC-facing decision facts.
- Strategy planning — read
08_strategy_planner.md; combine priority analysis, experiment memory, paper/code evidence, venue constraints, and user decisions. Use a questionnaire for strategy/experiment trade-offs before locking the plan.
- Snapshots/versioning — read
09_snapshot_maker.md and 16_rebuttal_versioning.md at each durable checkpoint.
- Writing — read
11_response_writer.md, 13_ac_summary_writer.md, and 12_template_designer.md only after strategy is evidence-backed and user-approved; 12 selects pdf_one_page, openreview_per_reviewer, global_comment, hybrid, or openreview_markdown_latex layout.
- Rehearsal — read
18_rebuttal_rehearsal.md once a draft exists; simulate reviewer/AC personas reading paper-vs-rebuttal in isolation, then route hardening back to 11/13. Advisory only; it never approves submission.
- Safety gate — read
14_safety_rule_checker.md before any final or paste-ready text/PDF/comment.
- Discussion rounds — read
19_discussion_round_handler.md after the first rebuttal is submitted; ingest reviewer/AC follow-up replies, track score/engagement per round, draft delta-only follow-ups (reusing the 18 answer bank), and route each through 14 before posting.
- Overleaf sync — read
15_overleaf_leaflink_sync.md only when the paper is on Overleaf/cn.overleaf or the user asks about cloud/local synchronization.
Read only the capability files needed for the current request.
Core outputs
A complete strategy-first run should produce:
- Intake completeness report
- workspace-local
.awesome-rebuttal/memory/*.json for project, paper, code, reviews, experiments, strategy, responses, templates, AC summaries, safety, versioning, and optional Overleaf sync
- Reviewer stance map
- Atomic concern ledger
- Common concern clusters
- Response strategy matrix
- Numbered review-driven experiment memory and triage (
EXP-*)
- Format-aware response blueprint and evidence-aligned response draft
- Coverage map from reviewer concerns to response units
- Rehearsal findings from simulated reviewer/AC personas, with a prioritized hardening list and an anticipated follow-up answer bank
- Multi-round discussion state: reviewer engagement/score tracker, per-thread decisions, and delta-only follow-up drafts gated by safety
- Optional AC summary decision, fact ledger, and draft when rules allow
- Rule/safety checklist and
safety_memory.json final gate
- Rebuttal template report and active template copy when template-based response is needed
- Canonical
.awesome-rebuttal/snapshots/snapshot_memory.json plus generated user Markdown snapshot, or git checkpoint according to the selected versioning mode
Project-local state folder
For each actual rebuttal workspace, create or use:
.awesome-rebuttal/
├── memory/ # project/paper/code/review/experiment/strategy/response/template/rehearsal/discussion/safety/versioning memory
├── drafts/ # response blueprints, drafts, and coverage maps
├── snapshots/ # JSON and markdown reload snapshots
├── templates/ # active rebuttal templates copied/adapted for this project
├── logs/ # rule-search, compile, and validation logs
└── cache/ # disposable skill cache; safe to regenerate
The installed/global skill folder contains reusable instructions and assets only. Runtime memory, snapshots, fetched templates, and logs belong in the current workspace's .awesome-rebuttal/ folder.
Workspace convention
Recommended runtime workspace:
<rebuttal-workspace>/
├── Code/ # code, scripts, configs, reproduced outputs
├── Paper/ # paper PDF/LaTeX; Overleaf sync target if used
├── Reference/ # reviews, venue rules, reference papers, notes
└── Temp/ # temporary extraction, scratch drafts, search/cache outputs
Do not force this layout on an existing organized workspace. Map existing paths and record them in project memory.
Progress preservation modes
Ask the user to choose one:
manual_git: suggest checkpoint boundaries and commit messages; do not commit unless explicitly asked.
auto_git: create local milestone commits only; no push, history rewrite, or destructive git operations without explicit instruction.
markdown_snapshot_only: maintain .awesome-rebuttal/snapshots/snapshot_memory.json as the canonical reload entry and render .awesome-rebuttal/snapshots/REBUTTAL_SNAPSHOT.md or .awesome-rebuttal/snapshots/PROJECT_SNAPSHOT.md as the user-readable progress summary.
LaTeX environment policy
During workspace bootstrap, detect whether a local LaTeX toolchain is available (latexmk, pdflatex, xelatex, lualatex, bibtex/biber, kpsewhich, or tectonic). If no compiler is available and a PDF rebuttal is needed, ask the user to choose a setup path for their platform: MacTeX/BasicTeX on macOS, TeX Live on Linux, MiKTeX/TeX Live on Windows, Tectonic, Overleaf-only compilation, or skip local compilation for now.
Overleaf / LeafLink policy
LeafLink is not an advertisement. It is a conditional helper for authors whose paper is on cloud Overleaf.
- Ask whether the paper is on Overleaf/cn.overleaf.
- If yes, offer LeafLink as an optional way to sync the project into
Paper/: https://github.com/xiongqi123123/LeafLink
- If no, stay silent and use local paper files.
- Never include LeafLink text in final rebuttal, reviewer replies, AC summaries, or conference-submission-facing text.
Safety gates
Block finalization if any of these fail:
- Provenance gate — every factual statement has a source.
- Commitment gate — every promise is already done, explicitly approved, or framed as future work.
- Coverage gate — every reviewer concern is answered, intentionally deferred, or marked
needs_user_input.
- Venue-rule gate — format/links/supplementary claims are allowed by confirmed rules.
- Tone gate — no reviewer attacks, defensiveness, flattery manipulation, or unprofessional phrasing.