gauk-ai-feedback

Tomáš Havránek, Charles University · more tools, data and code at meta-analysis.cz

An unofficial pre-submission review for applications to the Charles University Grant Agency (GA UK, Grantová agentura Univerzity Karlovy). It reads a draft the way an opponent will and tells the student what to fix before the faculty deadline. Formal problems come first: a missing section, a page limit, a budget over its caps, a timetable that follows the academic year. Then the substance, above all whether the application says what the project will add and whether the methods and the time requested can deliver it.

Unofficial. This is not a GA UK or Charles University tool, and neither has reviewed or endorsed it. Its rules are transcribed from the published documents of the 24th round and may contain mistakes; where those documents or your faculty say something different, they are right. It cannot tell you whether you will be funded, and it never writes any part of your application. Read DISCLOSURE.md before using it.

Česky: README.cs.md.

The easiest way is the web page: tjhavranek.github.io/gauk-ai-feedback, in Czech and English, with nothing to install. It copies the prompt for you and can check your attachments in your own browser without uploading them.

Quick start

1. Ask your project leader whether you may paste the draft into a chatbot (see Before you paste anything).
2. Open the prompt in English or Czech and click the copy icon in the top right corner of the grey box.
3. Paste it into a new chat, then paste your draft after it. Start with Microsoft Copilot via office365.cuni.cz, signed in with your university CAS account: Charles University says that version has commercial data protection. The prompt is the same for ChatGPT, Claude and Gemini. If a chatbot will not take all of it, download the prompt as a file from the web page and attach it instead. Pasting other people's personal data, such as the leader's CV, still needs their explicit consent.

Do this at least a few days before your faculty's deadline, so there is time to revise. That is all most students need; the rest of this page explains the review, the local formal check and the agentic route.

Three ways to use it

You need nothing but a chatbot for the first. The second is for students who are comfortable with a command line; the third also works in the Claude and ChatGPT desktop apps, without one.

1. With any chatbot: paste one prompt, then your draft. This is the main route and the one to start with.
2. The local formal check: a Python script that runs on your computer, sends nothing anywhere, and reports what can be counted: pages, the first-year budget totals, sections, the years in the timetable, the length of the leader's publication list.
3. With an agentic tool, optional and experimental: Claude Code, Codex or a similar agent runs the check, reads your files, writes the review and then verifies every quote in it against your text.

Before you paste anything

A chatbot sends whatever you paste to the company that runs it. The declaration of honour says the project text may not be given to anyone outside the team without the project leader's consent, so ask your leader, usually your supervisor, first. Leave out what should not travel: unpublished data, and the leader's CV unless they agree. Check whether your faculty has rules on AI tools; the university's guidance is at https://ai.cuni.cz/AI-81.html. An account with training switched off lowers the risk, but it does not make sharing permitted. The local check sends nothing anywhere.

1. With a chatbot

1. Open dist/prompt_en.md (English) or dist/prompt_cs.md (Czech) and copy everything from === PROMPT BEGIN === to === PROMPT END ===.
2. Start a new chat and paste the prompt. After it, paste your draft: the web-form fields and the text of the project proposal, and the CVs if you want them read too. Say which section you apply in (A, B or C), for how many years, and which version of the form you use. If the draft is long, send it in several messages and say when it is complete.
3. If you can, also paste the self-report block, dist/self_report_en.md or dist/self_report_cs.md, filled in with the numbers from the application's own character counters and your files. A chatbot cannot count characters or pages reliably, so the prompt forbids it to guess; without the block, those lines come back as NOT MEASURED.
4. Read the review. It always has the same parts: a short summary, the formal check, a list of what it could not check, a paragraph on the contribution, at most ten findings with a verbatim quote and a concrete fix each, and a paragraph you can forward to your supervisor.

Revise and run it again once or twice. Further runs on the same draft tend to repeat themselves, and by then your supervisor is the better reader.

2. The local formal check

Install Python 3.10 or newer. On macOS the command is python3; on Windows, py works if python is not found. Then, in this folder:

python -m pip install -r requirements.txt

Put your application into a folder of its own. The file names only need to contain a recognisable word: navrh or proposal for the project proposal, cv_resitel or cv_pi and cv_vedouci or cv_supervisor for the CVs, literatura or references for the list of references. Each name should fit one attachment only; the script does not guess between two candidates. A Word draft of the proposal is read as well. Copy templates/form.yml into the folder and fill in the web-form fields you have. Then run:

python checker/gauk_check.py one my_application/

The result is check.md in the folder. Paste it after the prompt in step 1, and the chatbot will take its numbers as given.

Every finding is marked BLOCKING, ADVISORY or UNKNOWN. BLOCKING is a clear breach of a written rule, such as six pages where five are allowed. ADVISORY is measured but open to a reasonable explanation, such as a word in the budget that suggests an ineligible cost, or a proposal that seems to be in a different language from the form. UNKNOWN means the script could not tell, for example from a scanned PDF. The report also lists the values it measured without finding a problem. There is no "all clear". A report with no findings still ends with the rules the script cannot see: whether you are within the standard period of study, how many projects you are already on, whether your leader has recommended the application, and your faculty's own deadline.

The script does not check character counts yet, and of the budget it checks only the first-year totals you enter in form.yml: the annual ceiling, the wage cap and the stipend share. Read the character counts off the counters in the application, and check the other caps against the budget rules in the prompt.

3. With an agentic tool (optional, experimental)

Download this repository as a ZIP (or clone it), copy your application files into a folder inside it, for example my_application, open the repository folder in Claude Code and ask it to check your application, for example "check my GA UK application in the folder my_application". No terminal is needed: Claude Code is the Code tab of the Claude desktop app, with a paid plan, and Codex is part of the ChatGPT desktop app. The agent runs the Python scripts itself; if Python is missing, it says so and asks before installing it. The project skill in .claude/skills/gauk-feedback/ makes it follow agentic/RUNBOOK.md: run the formal check, read your files, write the review under the same rules as the chatbot prompt, and verify every quote against your text with tools/verify_quotes.py before showing it to you. Codex and other agents that read AGENTS.md follow the same runbook. The agent never changes your files.

For a second opinion, agentic/ describes how to run two heavier tools on a proposal: mad-research, an audit by Claude and Codex, and paper-workshop, a workshop of Claude reviewers. Both were built for research papers, neither has been tried on a GA UK application, and there is no evidence that they give better advice than the single pass. Use them only if you already know them, and never let paper-workshop rewrite your application.

What the review will not do

It will not estimate your chance of funding, give a score or a grade, or compare you with other applicants. It has not seen the other applications, and a number of that kind mostly discourages people who should apply.

It will not write or rewrite your text, even if you ask. You sign a declaration that you prepared the application independently. The review says what is missing and where it belongs; you write it.

It is not for evaluators. Opponents, rapporteurs, board members and faculty officers should not use it on applications they assess: assessment is confidential, and a chatbot would send someone else's application to a third party. Faculty and GA UK staff are welcome to point students to it before the faculty deadline.

The AI declaration

The application asks whether you used AI in preparing it and, if you did, how and to what extent. Using this review counts, so note which tool and model you used. The applicant guide states that the answer is not a criterion for assessing the project. A template for describing AI use is at https://ai.cuni.cz/AI-81.html.

Limits

It covers new applications in the 24th round (the call opens on 1 October 2026; projects start in 2027) and nothing else: not continuation requests, not final reports. Faculty-specific rules and deadlines are not covered, and the faculty deadline is the one that binds you. The rules expire on 1 February 2027, after which the checker refuses to run until they are updated.

Its automated tests use invented applications, and only a handful of real drafts have been checked with it; see docs/VALIDATION.md. Language models miss problems that a specialist in your field would see, and they can be wrong with confidence. Push back on a finding you disagree with, and take the scientific questions to your supervisor.

How it is built

The rules live in one place, rules/, each with its published source. The prompts in dist/ are generated from them and from the texts in src/, and a test fails when a generated file no longer matches its sources.

rules/          the rules, in YAML, with their sources
src/            the hand-written prompt texts, with slots for the rules
dist/           the generated prompts and rubric; do not edit
checker/        the local formal check
tools/          build, text extraction, quote verification
templates/      form.yml for your web-form fields
agentic/        the agentic runbook and the optional integrations
tests/          invented applications and the test suite

To rebuild and test, install requirements-dev.txt and run python tools/build.py, python tests/make_fixtures.py and python tests/test_checker.py.

Already running a project

The switch at the top of the page has three settings: a new application, a continuation request with its annual report, and a final report. On either report setting the page copies a different prompt, which reviews the report itself: whether the comment on the spending ties each sum to what it bought and explains every transfer, whether each result the report mentions is attached, what a deferred assessment requires, and what a finished project is assessed on. It never says whether the year's work justifies continuing or what mark the project would get; the rapporteur and the boards decide that.

The report prompts are about 10,500 characters, a quarter of the application one, because they carry no application rules. The file check and the agentic route stay with new applications. Every rule rests on the published Řeším projekt page, which also has a worked example of the spending comment.

Problems and corrections

If a rule is wrong or has changed, please open an issue naming the rule and the published source. Do not paste any part of a real application into an issue. For anything else, there is a short anonymous feedback questionnaire linked from the page.

Author and licence

Written by Tomáš Havránek (meta-analysis.cz), Charles University, who is a member of the GA UK Grant Council; DISCLOSURE.md sets out what follows from that. The design follows erc-ai-feedback, a similar review for ERC Starting and Consolidator Grant proposals. MIT licence.