ERC draft pre-review unofficial Source on GitHub

An unofficial check of your ERC draft before you submit

Paste one prompt into a chatbot with your Starting or Consolidator Grant draft. It asks the model to read the draft against the ERC’s 2027 evaluation questions and flag routine problems: vague hypotheses, no risk plan, weak panel fit, a CV disconnected from the project, each with a quote from your draft and a repair. Four more checks follow: a second pass, a resubmission audit, interview preparation, and a route for Claude Code or Codex.

Tested on synthetic drafts only. It does not write proposal content, predict how the ERC will score you, or replace human review.

By Tomáš Havránek and Zuzana Iršová, Charles University · meta-analysis.cz

Independent and unofficial. Not produced, endorsed or approved by the European Research Council or ERCEA.

The pre-review

Basic · recommended for most applicants

  1. Copy the prompt

    basic/prompt.md

    Show the prompt

  2. Open a paid chat with training turned off

    Use a paid account and the provider’s strongest reasoning setting rather than the fast default. Before you paste anything, turn off training on your data (ChatGPT, Claude, Gemini) and check your institution’s rules. Paying does not make the text confidential, because the provider still receives it, and opting out of training does not delete it. Remove names and institutions you do not need, including other people’s, and do not paste embargoed or patent-sensitive material unless you have sanitised it.

  3. Paste the prompt, then your draft, then your stage

    Paste the prompt into a new chat, then your draft, and end with one line that names your stage.

    Stage: Pitch
    An idea sketch, or an abstract plus a draft CV. Methodology and a complete CV are not expected yet.
    Stage: Part I + CV
    The Extended Synopsis (Part I of the Scientific Proposal) in progress, with a draft CV.
    Stage: Full proposal
    The complete B1 (abstract, Part I, CV and track record) and whatever B2 you have. The panel reads only B1 at Step 1.

    If you forget the stage, it asks. If the stage does not match what you pasted, it says so and reviews what is there.

  4. Read the report

    The prompt asks for a report in this order (an example on a synthetic draft).

    ERC StG/CoG critique
    1. In plain words: four to six sentences, written so that you can forward them.
    2. Eligibility: a pass or a "cannot check from what you supplied" on the eligibility window and the time commitment, from the dates in your CV.
    3. Stage and Step-1 risk: the stage you declared and what you actually supplied, plus, at Pitch and Part I + CV, a risk level from low to critical.
    4. Provisional marks: a breakdown the ERC form does not have, on its sub-questions 1.1 to 1.4 for the project (1 to 5) and 2.1 to 2.3 for you as principal investigator (Outstanding to Non-competitive). None at Pitch stage.
    5. Score caps triggered: this package’s own heuristics for missing elements, not ERC rules, each named in plain words.
    6. Findings: up to 6 at Pitch, 8 at Part I + CV and 10 for a full proposal, ranked by severity, each with a verbatim quote from your draft and a concrete repair.
    7. What the next stage needs to add
    8. Workshop hand-off: one finished paragraph to send with your draft.

    The risk level and the marks are the model’s reading of structural completeness. They do not predict how the ERC will score you. Before acting on a finding, search your draft for the exact words it quotes, and check your own arithmetic: one synthetic draft carries a planted budget error, and two of three recorded runs missed it.

After the pre-review

Standard and Advanced cost more of your time and are cross-checks on a draft that has already been through the pre-review.

Standard: three roles in one chat

A more thorough pass with the same model · 20 to 30 minutes

Three roles read the draft in turn, in one chat. The Panel Reviewer is the pre-review above. The Devil’s Advocate looks for counterarguments the Panel Reviewer did not raise. The Writing Coach looks only at how the draft reads and is told to add no content, so check that each edit keeps your meaning.

  1. Run the pre-review above and leave the chat open.

  2. Paste this into the same chat, without your draft.

    standard/02_devils_advocate.md

    Show the prompt

  3. When it has answered, paste this into the same chat.

    standard/03_writing_coach.md

    Show the prompt
  4. Merge the three yourself. Where two roles agree, look there first, but that is one model agreeing with itself and not independent confirmation. Bring the workshop your synthesis and up to three questions. More on the synthesis.

Resubmission audit

Did the new draft fix what the last panel flagged? · check eligibility first

First check whether you may resubmit at all

Under Work Programme 2027, Table 6, a B or C at Step 1 in the 2026 Starting or Consolidator Grant call means you cannot apply to any of the five 2027 main grant calls: Starting, Consolidator, Advanced, ERC Plus and Synergy. So does a C at Step 1 in a 2025 Starting, Consolidator or Advanced Grant call. A B or C at Step 1 in the 2026 Advanced Grant call bars all of these except Synergy. An A at Step 1 without an invitation to Step 2, or an A or B at Step 2, carries no restriction. The rules change between Work Programmes, so check the current one first.

The next panel will not see your previous Evaluation Report, so the new draft cannot answer it. The report is evidence of weaknesses a fresh panel will tend to find again, and the audit asks whether each one is gone.

  1. Run the pre-review above first.
  2. Copy the panel comment and the individual reviews into a text file. Remove reviewer-identifying detail and other people’s personal information, but keep the substance.

  3. Paste the prompt into a new chat.

    resubmission/prompt.md

    Show the prompt

  4. Then paste the report and the new draft, each under its label:

    --- PREVIOUS EVALUATION REPORT ---
(panel comment and individual reviews)

--- NEW DRAFT ---
(Part I + CV, or the full proposal)

For each criticism you get a verdict (addressed, partially addressed or not addressed) with quotes from both documents. It does not predict the new panel’s score. Not yet tested on a real resubmission cycle.

Interview preparation

For Step 2 invitees · a question bank, then a stress test of your answers

It prepares you for a human mock panel and does not replace one. The question-bank prompt speaks in a panel member’s voice so that you hear the questions as a panel would put them. A panel member or remote referee must never use it to prepare a real interview.

  1. Paste the prompt into a new chat, then your complete B1 and B2. You get twenty to thirty likely questions.

    mock_interview/01_question_bank.md

    Show the prompt
  2. Draft a spoken answer of about a minute to each question.

  3. Stress-test one answer at a time, each in a fresh chat: paste the prompt, the question, your answer and, if it is short, the part of the proposal the answer relies on. Revise, then take the answers to a human mock panel.

    mock_interview/02_answer_stress_test.md

    Show the prompt

About ninety minutes, plus drafting the answers. Not yet used in a real Step 2 cycle.

Advanced: Claude Code and Codex

Experimental · manual setup · sends the draft to two providers

It runs the mad-research protocol against the same rubric in Claude Code and the Codex CLI. It takes thirty to forty-five minutes, has not been verified end to end, and is the only option that sends your draft to two providers.

In a pre-registered study by this package’s authors (Havránek and Iršová, 2026, not yet peer-reviewed), the authors of 44 published economics meta-analyses preferred a single pass by one frontier model to the debate protocol Advanced adapts. Grant proposals were not tested. On six proposals to a UK funder, an unrelated group found that a council of AI personas did no better than a single pass (Thorne and others, 2026, a preprint), so there is no evidence that Advanced does better. Setup.

Good to know