erc-review

A reviewing kit for ERC Starting and Consolidator Grant proposals: a guide to what the panel is actually scoring, a review protocol that works top-down and stops at the highest level that is failing, and two tools that execute it — one for a person, one for an agent.

It exists because reviewing a grant draft usually goes wrong in the same way. You open the file, start at page 1, and fix sentences. Four hours later the prose is better and the proposal still is not fundable, because the problem was that the scientific question was never nameable and no amount of sentence work reaches that.

What is here

ERC_writing_guide.md	The rules. §1–12 are what the call and the panel want; §13 is the review protocol.
.claude/skills/erc-review/SKILL.md	A Claude Code skill that runs the judgment stages of §13 against a draft.
tools/erc_check.py	Settles everything a regex can settle, so neither you nor an agent spends attention there. No dependencies.
ercreview.sty	Margin-note macros for leaving findings in the draft, keyed to guide sections.
example/	A deliberately flawed skeleton, so the checker has something to find.

The protocol

Ten stages, most expensive mistake first:

0  census          what exists at all?
1  fundability     an ERC project, or an excellent next paper?
2  problem         a nameable question, with an instability and a cost?
3  layers          is each passage at its right altitude?
4  step-1          can B1 alone be scored? (the Step-1 panel never sees B2)
5  criteria        does every evaluation question have an address?
6  work plan       are work units and tasks correctly typed and bounded?
7  mechanics       do the sentences carry the argument's joints?
8  prose           fillers, hedges, voice, jargon
9  compliance      pages, placeholders, citations, funding table

Fixing an early stage rewrites everything downstream of it, so report only from the highest stage that has failures and say what you deferred. Stages 8 and 9 are mechanical; erc_check.py does them. Stages 1–7 are judgment, and each rubric item in guide §13.2 carries three fields — a test, the cost if it fails (which evaluation question, and for which reader), and what to suggest. The cost field is the point: a finding that cannot name the question it costs you is a preference, and should be dropped.

Using the checker

No dependencies beyond Python 3.9+. Run it from the directory holding your b1.tex and b2.tex; page counting additionally needs tectonic or any engine you point it at.

git clone https://github.com/borgr/erc-review.git
ERC=$(pwd)/erc-review

cd path/to/your/proposal
python3 $ERC/tools/erc_check.py --census --pages   # stage 0: what exists
python3 $ERC/tools/erc_check.py --pages            # everything, with page counts
python3 $ERC/tools/erc_check.py --only BLOCK       # just the submission blockers
python3 $ERC/tools/erc_check.py --json             # for an agent to parse

Findings print as file:line [code] message (guide §) and are triaged BLOCK (not submittable), SCORE (costs points on a named evaluation question), POLISH (costs goodwill). It also reports every section and task with word and sentence counts, flagging tasks too short to carry purpose–approach–contribution and moonshots over ~200 words.

Try it on the example first:

python3 tools/erc_check.py example/

Using the skill

Install it where Claude Code looks for skills, then ask for a review from your draft directory:

ln -s "$ERC/.claude/skills/erc-review" ~/.claude/skills/erc-review

The skill reports at most five findings at a time, each as where → what → why it costs → what to do, and offers three routes per finding: draft the replacement, mark it in the file with \ercscore{§}{...}, or ask you the question — the last whenever the fix needs a fact that is not in the draft. It edits stages 8–9 directly and reports a count; it never silently rewrites an argument.

Leaving findings in the draft

\usepackage{ercreview}          % after todonotes, if you use it
...
\ercblock{13.3}{funding table still blank}
\ercscore{9.2}{this is the rationale, not the objective -- the objective is missing}
\ercpolish{10}{"and then" hides the causal step}
...
\listoferccomments               % summary page of every open finding

\usepackage[off]{ercreview} drops all of them for the submission build.

Caveats

Verify the numbers against your own call. Everything here was checked on 11 August 2026 against the Information for Applicants to the Starting and Consolidator Grant 2027 Calls (v11.0) and the ERC Work Programme 2027, and guide §12 records that plus the four things that changed recently enough that older advice is actively wrong — Part II went from 14 pages to 7, feasibility left the Step-1 assessment, novel methodology stopped being a scoring clause, and the eligibility windows widened to 0–10 and 5–15 years post-PhD. The ERC reissues both documents each July. The authority is the pair for your call, not this repository.

This is one reading of one seminar plus the official forms. Guide §13's ordering claim — that fixing framing before prose is strictly better — is an editorial position, not an ERC policy. Corrections and disagreements are welcome as issues.

No worked examples from real proposals. The seminar deck this draws on contains examples from active ERC projects under an explicit no-redistribution restriction. Those are excluded; every illustration here is written fresh. See the guide's Provenance section for the sources and their terms.

Licence

Code (tools/, ercreview.sty, example/): MIT. Prose (ERC_writing_guide.md, SKILL.md, this README): CC BY 4.0. See LICENSE.