CCFA-Skills: a 17-skill agent family for CCF-A paper storylines

The problem CCFA-Skills was built for

The README opens with a scene rather than a feature list: experiments finish at two in the morning, results look better than expected, and the manuscript has quietly drifted. The original research question is buried under related work. The method description no longer matches the mechanism in the code. New experiments answer the last round of reviews but split the argument into disconnected threads. Each local edit improved something; the paper as a whole became harder to read.

The stated diagnosis is that one long prompt cannot do retrieval, experiment design, writing and reviewing at the same time. Those tasks demand different behaviour: retrieval has to stay faithful to sources, experiments have to follow a protocol, writing has to serve the argument, and review has to stay independent. CCFA-Skills answers by splitting the work into 17 named skills and keeping the link between research question, evidence and conclusion alive across handoffs.

The audience is narrow on purpose. This is for researchers preparing submissions to CCF-A venues, which is why the repository ships LaTeX template material (ccf-latex-templates) and venue guides rather than a general prose assistant. If you write in Word and never touch a BibTeX file, most of the family is dead weight.

How the 17 skills hand work to each other

The repository is organised as a family, not a monolith. Each top-level directory is one skill: ccf-common, ccf-pipeline-orchestrator, ccf-project-scaffolder, ccf-idea-reviewer, ccf-idea-optimizer, ccf-literature-searcher, ccf-literature-monitor, ccf-experiment-designer, ccf-integrity-auditor, ccf-paper-reviewer, ccf-paper-writer, ccf-humanization, ccf-rebuttal-writer, ccf-visual-composer, ccf-submission-checker, ccf-paper-to-exemplar and ccf-skill-forger. Two plugin manifests sit at the top level, .claude-plugin/ and .codex-plugin/, which is how the family is exposed to different agents.

There is a fixed entry order. According to the README, every CCFA task enables ccf-humanization first, then ccf-common, and only then the specific skill. The first governs tone: it removes defensive hedging, mechanical enumeration, excessive dashes and internal engineering phrasing while keeping limitations and disclosures that affect the conclusion. The second carries the collaboration, scope, evidence and file rules. On handoff, the already-active rules are reused and only what changed or was lost is refreshed.

The design consequence is worth stating plainly. Loading fewer skills is presented as a feature: the README argues that loading more skills makes results more vulnerable to irrelevant rules. That is the opposite of the usual advice to give an agent maximum context, and it means the orchestrator (ccf-pipeline-orchestrator) matters as much as any individual skill. If you bypass it and load a dozen skills at once, you are working against the architecture.

Installing CCFA-Skills and running a first review

The README links separate getting-started pages for Codex, Claude Code, Cursor, Gemini CLI and other agents, plus an auto-update page. The one-line install shown in the README targets Codex with a global install and copies every skill:

npx skills add mikubaka88/CCFA-Skills --global --agent codex --skill '*' --yes --copy

After installation, the README says to state the research problem directly rather than invoking a skill by name. The example prompts are plain sentences, not commands. A first useful run is idea triage, which the README gives in this form:

严格评审这三个选题并排序，指出各自最可能被拒的原因。

That is Chinese for: review these three topics strictly, rank them, and name the most likely reason each would be rejected. The README's other starter prompts cover literature search for the past three years, experiment design with an explicit instruction not to fabricate results, rewriting a method section in CVPR style while preserving equations and citations, and drawing a method architecture figure that starts as an aesthetic draft before asking whether to rebuild it as PPTX.

One practical note: the README shows the install command in a PowerShell code block, so Windows users should expect to run it in PowerShell rather than bash. The README does not document an uninstall or rollback path, so if you install globally you should know where your agent keeps its skill directory before you start.

Where the family stops being the right tool

The most important limitation is stated by the project itself. The README says missing facts stay missing until they are reliably found or verified, and one of the example prompts explicitly instructs the experiment designer not to invent results. That is a boundary, not a capability: CCFA-Skills will not fill a gap in your evidence, and if your claim table is incomplete, the integrity auditor has nothing to audit.

Second, the family does not replace judgement. The README draws a line between the reviewer and the writer: the reviewer judges, the writer revises, and you look at the problem before deciding how to change anything. If you want a tool that rewrites the paper for you, this is the wrong shape of product.

Third, the figure pipeline is staged rather than finished. The method architecture example in the README is generated as a PNG visual draft first, and only after the composition is confirmed can it be rebuilt as SVG, vector PDF or a PPTX made of native objects. If your workflow needs an editable vector figure immediately, plan for the extra round trip. The README also notes that the numeric showcase figures are illustrative and do not represent paper conclusions, so they should not be read as evidence of anything.

How it differs from a single writing prompt or a reference manager

The closest comparison is not another paper-writing tool but the habit most researchers already have: one long prompt that asks a general assistant to search, draft and critique in the same conversation. The README names that pattern as the problem, on the grounds that retrieval, protocol-following, argumentation and independent review interfere with one another when they share a context window. CCFA-Skills splits them into separate skills with a fixed activation order, so the reviewing step is not the same voice as the revising step.

A reference manager solves a different slice again. It stores and formats citations; ccf-literature-searcher and ccf-literature-monitor are about finding related work, datasets, benchmarks and public baselines, and ccf-integrity-auditor is about checking that claims, numbers, terminology, figures and citations line up across the manuscript. Neither of those is a bibliography database, and the README does not claim it is.

There is also a self-improvement path that most prompt collections lack: ccf-skill-forger exists to improve the skills themselves, resolve conflicts and run pre-release checks, and ccf-paper-to-exemplar extracts reusable writing methods from a paper you supply rather than copying its sentences.

Maintenance, licensing and the cost of keeping up

The repository is not archived, and the last push was on 2026-09-16. The release history is short but dense: v0.9.0 on 2026-08-13, v0.10.0 on 2026-09-06, and v0.10.0-r2 on 2026-09-16, the last one labelled as a gpt6 adaptation. The README describes the r2 update as an ICLR 2027 alignment that calibrates anonymity, page limits, the year template and the AI usage statement check, with the venue guide living at ccf-paper-writer/references/venue-guides/iclr.md. The version number itself stayed at 0.10.0.

That cadence tells you what the upgrade cost looks like: venue guides and review checklists are the parts that go stale, and they are the parts being revised. If you use the family across a submission cycle, expect to re-read the venue guide before you rely on the submission checker. There is a CHANGELOG.md at the top level and an auto-update page under docs/getting-started/, so the update path is documented even though the README does not describe rollback.

The licence is MIT. In practice that means you can copy, modify and redistribute the skills, including inside a company or a lab, provided the copyright notice and permission notice travel with the copies. It does not give you any warranty, and it does not cover the third-party material the README references: the LLaVA-4D figure screenshot is credited to its authors and marked as used for non-commercial academic composition study, with the README asking that the maintainer be contacted for removal. If you plan to reuse anything from assets/visual-showcase/references/, that is a separate question from the MIT grant.

What the README does not answer

Several things a prospective user would want to know are simply absent. There is no documented rollback or uninstall procedure for the global install. The README does not state which agent versions were validated, only that per-agent installation pages exist. It does not describe how skills are stored on disk after installation, so the --copy flag's effect is not explained beyond the command itself.

The evaluation directory exists at the top level, and there is a demo/attention-is-all-you-need/ example, but the README does not present measured results from either, and this article does not claim any. Similarly, the repository has a 实验结果.md file at the top level, but nothing in the README ties it to a published benchmark.

The honest summary is that CCFA-Skills is a structured workflow with a clear internal discipline and a thin public record of outcomes. You can read the skill definitions in the repository before installing anything, which is the cheapest way to decide whether the split matches how you actually work.

Editorial conclusion

Adopt CCFA-Skills if you already write LaTeX papers for CCF-A venues and work inside a skill-aware agent such as Codex, Claude Code, Cursor or Gemini CLI; the family's value is the handoff discipline, not any single prompt. Skip it if you want a generic writing assistant, a reference manager, or a tool that verifies your experimental numbers for you: the README states that missing facts stay missing until found, so nothing is invented on your behalf. Before relying on it for a deadline, verify three things yourself: that your agent reads the plugin directory it installs, that the ICLR 2027 guide under ccf-paper-writer/references/venue-guides/iclr.md matches the current call, and that ccf-integrity-auditor can actually see your claim table.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. mikubaka88/CCFA-Skills on GitHub
4. README
5. Releases