The landscape, honestly surveyed

How it compares

There are good AI grading products, and several of them will suit you better than this one. Here is what you get here that you cannot get from them, what you give up in exchange, and the searches behind every claim so you can repeat them.

Surveyed 2026‑08‑24, re-verified 2026‑08‑26. Prices below are quoted from each vendor's own pricing page on the date shown, and they change.

The case for this one

What you get here that the products do not offer

Most of these are structural rather than featural, which is why no amount of product polish elsewhere closes them.

01

No new vendor sees your students' work

This is the big one for anyone who has been through a procurement review. Adopting a commercial grader means adding a company to your institution's list of processors handling education records, and getting that approved. Here there is no such company: the plugin adds no destination of its own, and the only service in the loop is your own Claude session, under terms your institution may already cover.

02

No licence, per seat or per student

Free, CC BY 4.0, forever, for any number of students. You pay for your own model usage and nothing else. There is no plan to age out of and no per-student count to true up at renewal.

03

Your rubric, not a vendor's idea of good work

The standard being applied is a markdown file you wrote and can edit mid-term. In a hosted product the scoring model is the product, and you tune it at the edges. Here every precedent your course sets is written back into your own rulebook and is yours to read, change, or delete.

04

Nothing to be locked into

Every artifact is a plain file on your machine: rubric, findings, matrix, precedents, letters. There is no export step because there is nothing to export from, and if you stop using this tomorrow you keep everything, in a format you can still read in ten years.

05

It is not limited to essays

Most AI graders are essay graders. This reads whatever you point it at — code, notebooks, PDFs, reports with figures, mixed submissions — because the discipline is “cite the source” rather than a text-scoring model.

06

You can read the whole method

Including the parts that failed. The rules, the agent prompts, the scripts, and the run statistics are all in the open, so you can audit the auditor rather than trust a scoring model you are not permitted to inspect.

And the rigour layer on top of those: blind per-student auditing, a cross-student consistency pass, an explicit bound that detects a loop auditing its own repairs, and a write-back step that makes the instruction set stronger every term. No surveyed project combines more than two of the five (Figure 11).

What it costs, side by side

 PriceCapWho else holds the student work
Grade it like an audit Free (CC BY 4.0) plus your own model usage None Nobody new. Your own Claude session, under your existing terms
CoGrader Free tier; $15/mo billed annually, or $19/mo 100/mo free; 350/mo on StandardThe vendor
EssayGrader Free tier; $6.99 / $14.99 / $34.99 per month 50 / 100 / 350 / 800 essays per monthThe vendor
Gradescope Institutional quote; not published publicly Per institutional licenceThe vendor (Turnitin)
Class Companion Free to teachers; LMS integration needs a school plan, price not published Per planThe vendor
PrairieLearn Free and open source, self-hosted NoneNobody new, if you self-host

Prices read from each vendor's own pricing page on 2026-08-26. CoGrader and EssayGrader publish theirs; Gradescope and Class Companion do not publish institutional pricing, so those rows say so rather than guessing. Free tiers are real and may be all a small course needs — the point of this table is not that the products are expensive, it is that a per-seat price and a third-party processor are two costs this one does not have.

Choose this one if

You care more about being able to defend a grade than about grading quickly; your institution's approval process makes adding a vendor slow or impossible; your submissions are not plain essays; you want the standard being applied to be a document you own; or you want the same method for something that is not grading at all.

First, the honest part

Where others are genuinely ahead

Having made the case, here is the other side of it, at full strength. If what you need is convenience, LMS integration, vendor support, or proven scale across a whole institution, one of the products below will serve you better. This is a rigour layer, not a convenience rival, and pretending otherwise would waste your time.

  • Commercial products are ahead on everything operational. Gradescope (AI answer-grouping at institutional scale), CoGrader, Brisk, Class Companion and others have real interfaces, LMS integration, grade passback, support contracts, and compliance positioning. None of that exists here.
  • PrairieLearn is ahead on battle-tested infrastructure and a human-override workflow proven at course scale, in an open-source platform with a real community.
  • LLM-AES is ahead on academic validation — peer-reviewed at LAK25. This project has honest internal run statistics, which is not the same thing as external peer review, and should not be confused with it.
  • homework-grader had the calibration idea first. Its quantitative teacher-calibration was ahead of this project's original design, and /grade-audit calibrate exists because that was the right idea. Credited here rather than quietly absorbed. (It appears dormant: at survey time its repository was created and last pushed on the same day, 2026‑02‑22, at 7 stars — which is part of why the idea needed a maintained home.)

And what you give up by choosing this one

No interface beyond a terminal, no LMS integration or grade passback, no support contract, no certification, and no peer review. It requires Claude Code and a willingness to edit markdown. Model usage is not free even though the tool is, and a round over N submissions spawns N auditors plus a lead pass, three to five times over. If that list is disqualifying, it genuinely is disqualifying, and one of the products above is the better answer.

Direct comparison

Things you could install or adopt instead

ProjectTypeEvidence-cited judgments Independent audit passWrite-backHuman in the loop Published validation
Grade it like an auditMethod + Claude Code plugin Every judgment named and evidenced; praise audited too Blind subagent per submission, lead consistency pass, explicit convergence bound Every caught failure becomes a durable rule Per-fix approval; one human reader outside the loop Honest run statistics, failures included
homework-graderClaude Code skill Mandatory quote per dimension Bias controls and κ-calibration; single grader, no second reviewer Manual rubric refinementTeacher calibrates and reviews flags Built-in κ/ρ/MAD machinery; no external statistics
teaching-skillsClaude Code skill suite Evidence-bound outputsDeterministic gates; no blind fan-out Per-term improvement recordProfessor checkpoints
claude-edu-pluginsPlugin (Moodle/Kahoot) Partial
LLM-AESResearch framework Score-orientedDual-process (LLM plus supervised model) YesPeer-reviewed (LAK25)
PrairieLearn AI gradingCourse platform feature Per-submission rationale; full prompt transparency Human review and override instead The human grade always winsInstructor-reported accuracy

Capabilities for homework-grader and teaching-skills are taken from their READMEs rather than from code inspection; commercial claims are vendor claims. If this table misses a competitor, please open an issue — it follows the same rule as everything else here: it gets corrected, and the correction is the point.

The claim, stated so it can be checked

What is actually unique

Not “AI grading.” Plenty of tools do that, and several do it better than this one in every way a busy department would care about. What no surveyed project combines is these five at once:

 Blind per-student fan-outCross-student consistency pass Convergence boundStructural write-backFailure-inclusive stats Total
Grade it like an audit 5
homework-grader 0
teaching-skills 1
LLM-AES 1
PrairieLearn 0
claude-edu-plugins 0
Figure 11. Scored from public documentation, generously where ambiguous. The honest reading is not “this one is best” — it is that these five properties are unoccupied territory, and a project that wants them has to build them.

The convergence bound is the one most likely to be undervalued. Every grading loop can detect problems; far fewer can detect that they have started auditing their own repairs, which is how a careful process turns into an expensive one that never terminates. That failure happened here, on a real run, and the rule exists because of it.

Worth noting on the official side: Anthropic's k12-teacher-skills (400 stars at re-verification, up from 394) covers lesson planning and differentiation and does not grade. Official adjacent interest; unoccupied niche.

Show your work

The searches behind the absence claims

Claiming “no existing Claude Code skill does audit-style grading” requires showing the searches that failed to find one. Run 2026‑08‑24, re-verified 2026‑08‑26:

  1. Four awesome-lists — hesreallyhim/awesome-claude-code, travisvn/awesome-claude-skills, karanb192/awesome-claude-skills, BehiSecc/awesome-claude-skills — grepped for grade, grading, grader, rubric, essay, homework, teacher, assess. Zero student-work grading entries.
  2. GitHub API repository searches — “claude skill grading”, “claude grader”, “claude-code rubric”, “claude skill essay grading”, by stars. No audit-style grading skill; the only additional grading skill was a 2-star single-purpose IELTS scorer.
  3. Web search for blind or independent per-submission subagent grading — generic subagent tutorials only.
  4. The SkillsMP marketplace — query “grading”. No student-work grading skills.
  5. The official anthropics/skills repository — 17 skills, none education or grading. Verified via secondary walkthroughs; marked high-confidence rather than file-by-file.
  6. Re-verified on 2026‑08‑26 — a fresh recency-sorted search surfaced no new grading skill in the intervening period. The one true peer remains dormant. The landscape not moving is itself the finding.