Kim Ccf Research

kim-ccf-research — Master Orchestrator

The kim-ccf-research series is a 25-skill workflow for ML/AI PhDs targeting CCF-A venues (TPAMI, TKDE, AAAI, KDD, ICML, WWW, ICLR, CVPR). This skill is the entry point: it detects the current stage of the user's paper project and routes to the correct leaf skill.

The orchestrator performs no substantive work itself. It only:

• inspects the project root for stage evidence (docs/idea_report.md, docs/implementation.md, docs/dev_log.md, docs/manuscripts/paper.md, ccfa-review-reports/, etc.),
• chooses the next leaf to invoke,
• and enforces the shared cross-skill controls (handoff mode, task mode, evidence standards, privacy).

All real research/writing/review work happens in the leaf skills.

Quick Start

Three trigger shapes:

Examples:

/kim-ccf-research
→ Detects docs/idea_report.md Part 1 exists but Part 2 missing.
  Recommends: kim-ccf-research-idea (idea concretization).

/kim-ccf-research survey graph anomaly detection with contrastive learning
→ Calls kim-ccf-research-explore with the direction argument.

/kim-ccf-research review my paper
→ Calls kim-ccf-research-review.

Trigger Conditions

Load this skill when the user:

• Wants to start a paper project from scratch.
• Asks "what should I do next?" inside an ongoing paper project.
• Names a CCF-A venue and wants the matching workflow.
• References kim-ccf-research-* literally.
• Says "research pilot", "CCF-A 流水线", "论文流水线", "顶会投稿助手".

Do NOT load this skill when:

• The user is doing exploratory web search unrelated to a paper project. → Use kim-ccf-research-literature directly.
• The user is doing generic code work without a research context. → Use kim-ccf-research-code or kim-ccf-research-prototype directly.
• The user wants to read a paper for understanding, not to write one. → Use nature-reader or kim-ccf-research-translate directly.

Pipeline Stages (12-stage chain)

Replaces ResearchPilot's 7-stage chain with a 12-stage CCF-A-targeted chain:

explore → idea → idea-review (optional gate) → literature → experiment →
implement → code → iterate → plan → write-method → write-experiments →
write-abstract → write-introduction → write-related → write-conclusion →
review → humanize → rebuttal → integrity-audit → submission → translate

Pipeline State Machine

Detects stage by inspecting the project root:

docs/idea_report.md missing               → stage 0: nothing started
docs/idea_report.md Part 1 only           → stage 1-2: explore / idea
docs/idea_report.md Parts 1-2             → stage 3: idea-review (gating)
docs/related_work_corpus.md               → stage 4 done
docs/exp_plan.md                          → stage 5 done
docs/implementation.md                    → stage 6 done
docs/dev_log.md non-empty                 → stage 7-8 done
docs/manuscripts/paper.md with `=== 论文架构 ===` → stage 9 done
docs/manuscripts/02_method.md             → stage 10a done
docs/manuscripts/03_experiments.md        → stage 10b done
docs/manuscripts/01_abstract.md           → stage 10c done
docs/manuscripts/00_introduction.md       → stage 10d done
docs/manuscripts/04_related.md            → stage 10e done
docs/manuscripts/05_conclusion.md         → stage 10f done
ccfa-review-reports/<slug>.md             → stage 11 done
docs/rebuttal/v*.md                       → rebuttal active

If project root is ambiguous (no docs/ and no .git/), ask the user.

Routing Rules

Replace deleted ResearchPilot stages with their CCF-A counterparts:

For all other intents, see ../kim-ccf-research-shared/references/routing.md which holds the full intent → leaf matrix.

Handoff Protocol

Before invoking any leaf, the orchestrator loads ../kim-ccf-research-shared/references/handoff-modes.md and respects:

• partial (default): confirm before invoking another leaf. Use for ambiguous cases (e.g., "should I run literature before idea-review?").
• full: silent passthrough. Use when the user has explicitly authorized a chain (e.g., /kim-ccf-research full-pipeline).

If a leaf's metadata.ccf_skill_controls.respect_session_denylists is true, honor any per-session denylist the user set earlier.

Quality Standards

Before any leaf produces a scored or published output:

1. Load ../kim-ccf-research-shared/references/review-output-standards.md for scoring and visible-output discipline.
2. Load ../kim-ccf-research-shared/references/privacy-and-evidence.md before any external WebSearch / WebFetch.
3. Load ../kim-ccf-research-shared/references/ccf-a-venue-map.md before any venue-specific work.
4. Load ../kim-ccf-research-shared/references/artifact-contracts.md to write to the canonical paths.

Integration with Non-Series Skills

The orchestrator defers to existing non-series skills for these intents (do NOT re-implement):

If a delegated skill is invoked mid-orchestrator, return to the orchestrator on completion and re-evaluate the pipeline stage.

Anti-Patterns

The orchestrator MUST NOT:

• ⚠️ IRON RULE: Perform substantive research, writing, or review work itself.
• ⚠️ IRON RULE: Skip idea-review before experiment for CCF-A submissions.
• ⚠️ IRON RULE: Invoke submission without first running integrity and humanize.
• ⚠️ IRON RULE: Drop or silently skip a stage on user request without a written justification in docs/dev_log.md.
• ⚠️ IRON RULE: Cite by raw URL without a source-registry.yaml ID when one matches.

Version Info

metadata.version: 1.0.0 — initial release. metadata.last_updated: 2026-08-20. metadata.depends_on enumerates the 23 leaf skills + shared. If Claude Code honors depends_on at orchestrator load time, those leaves are preloaded; otherwise the orchestrator dynamically invokes each by name.

Bilingual Trigger Vocabulary

EN: research pilot, paper pipeline, end-to-end paper, research to paper, CCF-A workflow, idea optimization, literature review, experiment design, research code, paper review, rebuttal, submission check.

中文: 研究助手入口, 论文流水线, CCF-A 流水线, 顶会投稿, idea 具象化, 方向探索, 文献综述, 实验设计, 代码实现, 整稿审阅, rebuttal 回复, 投稿检查, 学术流水线入口, 路由到下一个阶段.