Example

Technical Details took the images apart. This page shows one of them at work: the tex image and its editorial skill, which staff use on papers before submission and students on bachelor, master and doctoral theses before their supervisor reads a chapter. The example is a fictional chapter, reviewed here end to end, and the reviewed PDF is embedded below so that you can read the notes yourself.

Who uses the editorial skill, and for what

A draft that goes to a supervisor or a co-author is usually read for two things at once: whether the physics is right and whether the text is. The second reading costs the reader time that the first one needs. The editorial skill takes the second reading off their desk. It reads the LaTeX source of a chapter or a paper and attaches a note wherever the text has a spelling or grammar mistake, wording that could be better, a visibly wrong statement, a broken reference, an inconsistency, or a point that needs clarifying. Every note names the issue and the fix.

The institute has a firm rule for papers and theses: their text is written by people, in full, and no agent generates or rewrites it. The Research section explains why. The editorial skill is built to fit that rule, not to bend it. It never changes the text: the body words stay exactly as written, and every note is a suggestion. The agent’s remarks themselves are not copied into the document either; the author reads each one, decides, and types the correction by hand. Everything, the agent’s comments included, passes through a human brain before it reaches the manuscript. And the skill works inside the author’s own repository, on a branch named after the agent and the date, so the review is a set of commits the author can read, merge or delete, and the main branch is untouched. The skill’s own text says so in a few sentences:

The chapter under review

The chapter under review is invented for this page: five pages of a bachelor thesis on a standard model of a quantum magnet, the transverse-field Ising chain, written by a fictional student at a fictional university, with an introduction, the model, its exact solution by the Jordan–Wigner transformation, a phase diagram, a numerical check and a summary. Its mistakes were planted so that every one of the skill’s six note categories (listed in a table further down) occurs at least once. The draft is a single file, chapter.tex . The following frame shows it as the student submitted it, so that you can compare it with the reviewed version further down:

The review, step by step

From the author’s chair the review is a short loop: open the project in the sandbox, ask for the pass, wait, read the notes, and type the corrections yourself. The steps below are that loop as it was run for this page, with what the agent did in between.

1. Open the project in the sandbox

   Go to the directory of the thesis or paper and start the tex image there. The first start in a directory asks whether it may set up the two briefing files; answer yes. You land in a shell inside the container, in the same directory, and start the agent of your choice in its unattended form, the one with the d suffix that the workflow introduced: inside the sandbox there is no reason to approve each step.

   cd ~/theses/example
   git push          # a local commit is not a backup
   secdev tex
   clauded
   

2. Ask for the pass

   One sentence is enough. Where to work, which branch to create, what to record and how to build the PDF all come from the skill, which the agent reads before it starts.

   >Prompt to the agent:

   Please do an editorial pass over chapter.tex. British English. Do not change the text.

3. Let the agent work

   The skill prescribes six steps, and the agent follows them without further questions unless something is missing. In this run:

   • The agent found the directory to be a Git working tree, so it chose branch mode, recorded the starting commit and the state of the working tree, and created and switched to annotations-claude-2026-09-26. Nothing is edited before that switch.
   • The agent identified chapter.tex as a standalone master document: a report with amsmath, tikz, hyperref and one project macro, built with pdfLaTeX through latexmk.
   • The agent settled on the document type and language: a thesis chapter in British English, which tells it what to weigh, such as notation drift and references that no longer point where the text says.
   • The agent read the file from top to bottom and attached 25 notes with the two macros the skill provides. A span wraps the exact words at issue and gives them a pale background; a point note sits after a construct that cannot be wrapped, such as a cross-reference or a displayed equation. Two notes carry a tutor panel, an extra box that states the general rule, because the same kind of mistake occurred more than once. The body text was not changed.
   • The agent added two lines to the end of the preamble: the package that typesets the notes, loaded from the skill’s own directory instead of being copied into the project, and the review metadata with the exact model identifier and a timestamp frozen at the start of the review.
   • The agent built the document with latexmk, with the skill’s asset directory on the input path. The build converged in three passes with no warning that a margin note could not be placed, which is the skill’s test that the PDF is usable. It rendered every page, checked that the diff against master contains only annotation markup and the two preamble lines, committed the review in two chunks, left the branch checked out, and reported.

4. Read the PDF

   The agent’s closing message names the branch, the commits and the PDF, and the PDF is the thing to read. A circled, coloured number in the text marks each note, and the same number heads the note’s box in the nearest margin, in the colour of its category. The reviewed chapter, page by page:

   The annotated source behind the PDF is chapter-annotated.tex . Its diff against the draft consists of the two preamble lines and the notes, each an \MNOTE or \MNOTEspan macro placed after the words it concerns.

   The six categories, with their colour and symbol in the PDF and the number of notes each received in this run:

   Category	Meaning	Colour and symbol	Notes
   Language	Objective spelling or grammar mistake	Red, spelling check	5
   Style	Valid wording that could be improved	Blue, pen	3
   Factual	Clear substantive scientific or mathematical error	Amber, warning triangle	4
   Reference	Wrong or broken citation or cross-reference	Purple, broken link	4
   Consistency	Conflicting notation, terminology or language conventions	Teal, scales	6
   Clarification	Ambiguity, missing definition, or a question needing context	Grey, question mark	3

   The distinction the skill insists on is between a convention and a fact. Writing the coupling as $J$ in one place and as a calligraphic letter in another is a consistency note; writing that the gap is four times the distance from the critical point when the equation two lines above says twice is a factual note. Where the evidence is not enough to call something wrong, the note asks instead.

5. Fold the results into the draft, by hand

   Decide note by note. Type each correction you accept into the draft on master, in your own words; the branch is never merged into the text, because the institute’s rule is that no agent output, not even a suggested wording, goes into a thesis or a paper except through the author. Commit the corrected draft. Keep the branch as a record of the review or delete it. Nothing was pushed by the agent: the skill keeps every commit local, and in an ordinary secdev container there is no key to push with. You push from the host, with your own credentials.

6. Run the pass again

   One pass is seldom the last. Corrections bring new wording, a chapter grows, and a note you disagreed with may make more sense once the paragraph around it has changed. Ask for another pass on the corrected draft: the skill starts a new branch with the new date, so reviews never overwrite each other, and the earlier branch stays as a record. A pass that finds nothing says so, and that is the signal to send the chapter to your supervisor, or the paper to the journal.