CAMPUS
Get startedSign up

Agents · sub-agents · skills

Your agents

4 agents, one per role, 19 domain sub-agents and 54 skills. The agent receives the request, the sub-agent chains the steps, the skill does the work. You keep the decision.

Install the agents in Claude: connectors, plugins, skills →

Map of agents, sub-agents and skills, with field variants and routines

How the agents work together

SOURCESROOT FOLDERROUTINESAGENTS & SKILLSARTIFACTSyour decisions become rulesMailread-onlyCalendarwrite after approvalCloud storageshared documentsPubMed · trainingpublic sourcesDownloadswhat comes inDepartment scheduleread-onlyRules and plans_SYSTEMWork foldersflatLocal mail indexfast searchResearch databaseaggregates onlyHolding areafiles in doubtArchivenever deleteMorningtriage, daily planEveningprojects, decisionsFilingevery eveningWeeklydeadlines, watchMonthlyCV, index, backupmail-triageby actionfilingrule of doubtorchestratorroutes, never doesmanuscript-reviewnever overwritecongress-talkplan approved firstliterature-watchweekly digestnew skillyour next building blockDashboardanticipation, in a databaseApproval pagesapprove · fix · discardMonitoringstatus lightsSkills mapwho does whatYoudecide, send, sign
  • flow
  • read-only
  • plan or rule
  • hand-over
  • database write
  1. 1

    What comes in

    Everything starts from six sources. The AI reads them; it writes to none of them without your approval. The department schedule stays strictly read-only.

  2. 2

    One root folder

    Whatever the AI must see lives in a single folder: rules and plans, flat work folders, a mail index for fast search, a holding area for files in doubt, and an archive instead of a bin.

  3. 3

    Routines that read a plan

    Four or five scheduled routines, not twenty-five. Each reads its run plan in the rules folder: adding a task means adding a line to the plan.

  4. 4

    Agents and skills do the work

    Routines call the agents and their skills. An orchestrator does nothing itself: it qualifies the request and hands over to the right building block (purple dashes).

  5. 5

    Artifacts to see

    Skills write into the artifacts' databases (gold lines): the dashboard, the approval pages, the monitoring page, the skills map. The page stays the same; its data changes.

  6. 6

    You, at the end of every chain

    Nothing leaves without you: sending, signing, writing to the calendar. Your decisions on approval pages become rules that the next routine will read.

  7. 7

    Follow one link end to end

    An email arrives → it is indexed → the morning routine applies its plan → the triage skill sorts it by action → it writes it into the dashboard → you decide. Every arrow is written somewhere: a plan, a hand-over or a database write.

  8. 8

    Add a skill

    An instruction comes back three times? Write a skill, declare who calls it (here the orchestrator) and where it writes (the dashboard), then add it to the skills map. A link that is not written does not exist.

  9. 9

    The improvement loop

    Your corrections on approval pages are reviewed every week and become rules or skill updates, which you install after reading. The system improves without ever acting in your place.

Organiser Agent

The augmented secretariat, shared by all three roles: what comes in, where it is filed, what is coming.

↓ Download the agentConnectorFree sign-up · 5 downloads

Dashboard Agent

Daily roadmap, projects, deadlines to anticipate, documents retrieved.

dashboarddashboard-writedecision-queue

Skill details (3)
  • dashboardP1
    A single surface: daily roadmap, projects, deadlines.“what do I need to prepare”, “open a project”Guide →↓ skill
  • dashboard-writeP2
    Add a deadline or update a project from the conversation.“add a deadline”, “update this project”Guide →↓ skill
  • decision-queueP1
    One numbered queue of everything awaiting a decision.“what is waiting for my decision”, “accept 1 3 5”Guide →↓ skill
↓ All skills of this sub-agent campus-dashboard-agent

Archivist Agent

Root folder, typology, filing, inventory.

filingfile-typologyinventorysearch

Skill details (4)
  • filingP1
    File by copy, resolve doubts, learn.“file this”, “tidy my downloads”Guide →↓ skill
  • file-typologyP2
    Recognise a file's profile before filing it.“what type of file is this”, “where does this go”Guide →↓ skill
  • inventoryP3
    Cross-inventory every two months: what is dormant, diverging or missing.“run the inventory”Guide →↓ skill
  • searchP1
    Find a file, a conversation or a slide, stating what was searched and where.“find me”, “where is the document on X”Guide →↓ skill
↓ All skills of this sub-agent campus-archivist-agent

Administrative Agent

Forms, professional identity, career file.

identitycvactivity-historyactivity-report

Skill details (4)
  • identityP3
    Fill in a form without retyping; signature on request.“fill in this form”Guide →↓ skill
  • cvP2
    CV derived from a single data sheet and PubMed.“update my CV”Guide →↓ skill
  • activity-historyP3
    Activity history fed every month.“add this supervision”Guide →↓ skill
  • activity-reportP3
    Annual reports in the official template.“annual report”Guide →↓ skill
↓ All skills of this sub-agent campus-administrative-agent

Method Agent

Frame requests, make skills learn, check that the loop closes.

request-framinglearning-loop

Skill details (2)
  • request-framingP1
    Decode a dictated request, rewrite it to the template, wait for “go”.“frame my request”, “here is my dictation”Guide →↓ skill
  • learning-loopP2
    Collect what was learned, update lessons and skills, separate judge.“what did we learn”, “run the loop”Guide →↓ skill
↓ All skills of this sub-agent campus-method-agent

Researcher Agent

From the scientific question to publication and op-ed.

↓ Download the agentConnectorFree sign-up · 5 downloads

Bibliography Agent

Search, sort, read, synthesise, keep watch.

bibliographyliterature-watchevidence-synthesis

Skill details (3)
  • bibliographyP1
    PubMed literature review, tick-box sorting, synthesis and slides.“literature on X”Guide →↓ skill
  • literature-watchP2
    Topic feeds, weekly digest, "make the slide".“what's new on X”Guide →↓ skill
  • evidence-synthesisP2
    Evidence table and gaps in the slide bank.“state of the art on X”Guide →↓ skill
↓ All skills of this sub-agent campus-bibliography-agent

Manuscript Agent

Review, arbitrate, follow through to submission; peer reviews for journals.

manuscript-reviewmanuscript-portfoliopeer-review

Skill details (3)
  • manuscript-reviewP1
    Three levels, track changes, never overwrite.“review this thesis”, “make the next version”Guide →↓ skill
  • manuscript-portfolioP2
    Manuscript portfolio, pre-submission checks.“where are my papers at”Guide →↓ skill
  • peer-reviewP3
    Sort invitations, spot predatory journals, deliver on time.“invitation to review”Guide →↓ skill
↓ All skills of this sub-agent campus-manuscript-agent

Data Agent

Research database, cohorts, studies.

research-databasedatabase-querystudiesrecord-linkageregistry-entry

Skill details (5)
  • research-databaseP3
    Core datasets, freezes, pseudonymised index, quality checks.“add to the database”Guide →↓ skill
  • database-queryP3
    Patient counts by condition, on aggregates.“how many patients…”Guide →↓ skill
  • studiesP3
    Status, backward-planned milestones, analysis plan.“where is study X at”Guide →↓ skill
  • record-linkageP3
    Link the records of the same patient without wrongful merges.“link these two files”, “duplicate patients”Guide →↓ skill
  • registry-entryP3
    Prepare registry data entry and its follow-up.“prepare the registry entry”, “follow-up listing”Guide →↓ skill
↓ All skills of this sub-agent campus-data-agent

Congress Agent

Talks, slides, online materials, travel.

congress-talkslide-bankslidepdf-figureonline-slidescongress-travel

Skill details (6)
  • congress-talkP1
    Orchestrator of the slide pipeline: starting point, validated outline, template.“prepare my talk”Guide →↓ skill
  • slide-bankP2
    Reuse material from past presentations.“reuse my slides on X”Guide →↓ skill
  • slideP1
    A single slide: reference, template.“check this slide”Guide →↓ skill
  • pdf-figureP2
    From a PDF figure to a slide.“grab this figure”Guide →↓ skill
  • online-slidesP3
    Presentations from a subscribed congress platform.“slides from congress X”Guide →↓ skill
  • congress-travelP2
    Tickets, programme, travel roadmap.“organise my trip”Guide →↓ skill
↓ All skills of this sub-agent campus-congress-agent

Teacher Agent

Faculty, diplomas, supervision, continuing education.

↓ Download the agentConnectorFree sign-up · 5 downloads

Diploma Agent

University diploma and hands-on courses: start of year, lists, calendar, workshops, exams.

university-diplomahands-on-course

Skill details (2)
  • university-diplomaP3
    Start of year, mailing list, calendar, official results.“DIU start of year”Guide →↓ skill
  • hands-on-courseP3
    Design a hands-on course: framing, workshops, assessment.“prepare the hands-on course”, “allocate the cases to the workshops”Guide →↓ skill
↓ All skills of this sub-agent campus-diploma-agent

Clinician Agent

The department: organisation, staff meetings, rota, care pathways, projects.

↓ Download the agentConnectorFree sign-up · 5 downloads

Variants by field

The Organiser, Researcher and Teacher agents are shared. Only the clinical agent changes.

Pharmacist Agent

Medication circuit, pharmaceutical validation, preparations and devices, vigilance.

Dentist Agent

Clinical sessions, technical platform, clinical supervision of students.

Midwifery / Nursing Agent

Care organisation, placement supervision, protocols and pathways.

Scheduled routines

They do nothing themselves: they read a plan and call the agents at set times.

Morning

Every day, early

Organiser Agent: health check, mail, daily plan.

Morning retry

One hour later

Resumes the morning run if incomplete.

Filing

Every evening

Archivist Agent.

Evening

Every evening

Mail, projects, pending decisions.

Weekly

Friday

Researcher, Teacher, Clinician agents: deadlines, watches, rota.

Monthly

The 1st

Administrative Agent: CV, history; index, backup.

Four words, four levels

Level What it is Example
Agent A role. It receives the request and chooses the right domain; it never does the detail. Researcher agent
Sub-agent A domain of that role. It chains the skills in the right order. Conference agent
Skill A building block that does one precise job, with its deliverable and its safeguards. slide, pdf-figure
Routine A scheduled task that calls the agents at a fixed time. Weekly, on Fridays

Agents and sub-agents are technically “orchestrator” skills: their only job is routing. This is what lets you add a building block without rewriting everything.

Creating an agent or a sub-agent

  1. Start from a real role: researcher, teacher, clinician, organiser (or pharmacist, dentist…).
  2. List its domains: these will be the sub-agents (staff meetings, scheduling, conferences, bibliography…).
  3. Write the routing: for each type of request, which sub-agent; what the agent excludes and where it sends it.
  4. Declare the boundaries: what belongs to another role, and the rule that decides (for example: the payer).
  5. Test the routing on ten real requests, then add it to the agents map.

Anatomy of a skill

A skill is an instruction file that Claude loads when the request matches. It fits on one page:

name: review
description: Review a thesis, dissertation or article on three levels,
  with tracked changes under my name, and follow the manuscript through to submission.
trigger on: "review this thesis", "do the next version", "blocking issues before submission"
steps:
  1. Read the whole document and the instructions
  2. Substance, structure, form; references checked in PubMed
  3. Save a new suffixed version, never overwrite
deliverable: the reviewed .docx + a note of the blocking issues
safeguards: common foundation (never overwrite, never send, human validation)
hands off to: directory (co-authors), calendar (deadlines)

Creating a skill in six steps

  1. Wait for the third repetition: an instruction given three times deserves a skill.
  2. Write the description and the triggers in your own words: this is what makes Claude choose this skill over another.
  3. Describe the steps and the deliverable, then the safeguards specific to the task.
  4. Declare the hand-offs: which skill it passes to, and which one calls it.
  5. Test on a real case, fix, then freeze a version.
  6. File it: add it to the agents map, under its sub-agent; a duplicate or an alias is archived, never deleted.

Writing a good description

The description is the part of a skill that Claude reads to choose it. It fits in 40 to 60 words:

  • the purpose in one sentence: what the skill does, and its deliverable;
  • 5 to 8 triggers, in quotation marks, in your own words (“review this thesis”, “make the next version”);
  • one exclusion: what belongs to another skill (“not for slide layout”).

A description that is too long blurs the routing between neighbouring skills; one that is too short almost never triggers.

Delegating: two modes

Mode How When
Handover A single agent handles the request from start to finish. Most requests.
Orchestra A conductor launches several agents in parallel, then gathers their deliverables. Broad tasks only: it costs roughly ten times more than a conversation.

A nesting limit: a sub-agent does not launch other sub-agents itself. The orchestrator hands out the work and gathers it back.

Creating an artifact

An artifact is an interactive page published from Claude: dashboard, validation page, map. Three rules:

  1. One question per artifact: “what is coming up?” (dashboard), “what do I need to validate?” (validation page), “is everything running?” (monitoring).
  2. Data in a database, not in the page: the artifact reads a database fed by the routines and skills; you do not republish the page to change a number.
  3. Point, don’t copy: a document is referenced by its link, never duplicated in the artifact.

A link between two elements is not magic: it is always one of these three things, written down in black and white.

Type of link Where it is written Example
A hand-off In the skill (“hands off to”) email triage sends a received manuscript to the review skill
A routine step In the routine’s plan file the evening run updates the dashboard’s ongoing projects
A database write In the skill or the plan, with the target collection the literature watch writes its references to the watch tab

If a link is not written anywhere, it does not exist: this is the first thing the cross-inventory checks.

Reference

The skills in detail

Here is the actual content of the skills of an academic hat, anonymized: role, trigger phrases, steps, absolute rules and mistakes already made. Use it as a model for writing your own. Order: orchestrators, congress and slide chain, imaging, staff meetings and teaching, career.

Common ground for the whole chain

  • A single normative template for slides, owned by a single skill; in case of divergence, it prevails.
  • Language: dialogue in French, slide content in English, unless instructed otherwise.
  • Attribution: source and DOI mandatory for any data taken from PubMed (DOI in the notes, not on the slide).
  • No heavy generation without approval for the specific step under way: full presentation, email, publication, scheduled task.
  • Never delete, never overwrite: archive, and produce a new suffixed file.

1. Academic orchestrator (level 1)

Role: the single hat for university activity. It qualifies, routes and logs; it never itself produces a talk, a review or a bibliography.

Triggers: “my university activity”, “where am I with my articles”, “what do I have to prepare”, any university request with no named building block.

Explicitly excludes: consulting, invoicing, accounting, personal life. Rule: assignment follows the payer, not the tool.

Step 1 — qualify with one question

Domain Content
A · Research and publications manuscript, thesis, dissertation, review, bibliography, cohort
B · Congresses and presentations talk, workshop, clinical case, slide, session, travel
C · Department and staff meetings staff meeting, schedule, laboratory, reading protocol
D · Career and outreach CV, history, supervision, learned societies
E · Teaching and training lectures, MCQs, semiology, tutoring, OSCE, exam-question review (docimology), diploma, training watch

Step 2 — route

  • Review → review skill; literature search → bibliography; scientific state of the art → synthesis; watch → literature watch; defense → defense; research database → database and query.
  • Mandatory congress entry point: the presentation orchestrator, never bypassed. Direct call only for a single slide, a PDF figure, a clinical case or travel.
  • Department: schedule (read-only), staff meeting. Career: CV, history, identity, annual reports. Teaching: faculty, training watch (which produces nothing itself), university diploma.

Rules inherited by all building blocks

  • Never delete; calendar proposed, written only after “accept”; schedule strictly read-only.
  • Every document produced is filed immediately in the root folder, never left in Downloads.
  • No automatic sync of online storage: access on demand.
  • One question at a time; scoping before any complex task; nothing is generated without approval for the current step.

Traces: an academic cockpit (deadlines, manuscripts, congresses, staff meetings, reviews), the anticipation board, a reference document for the hat. D-30 rule: any congress presentation triggers a preparation alert.

2. Congress presentation orchestrator

Role: router, never executor. It owns the entry-point question, scoping, final assembly, review and the shared slide basket.

Triggers: “make a presentation, a talk, congress slides”, “prepare my communication”, “state of the art”, “symposium”.

Architecture: a single tabbed HTML workshop, tabs always accessible even when empty: bibliography; online congress materials; bank of past presentations; imaging; analysis and annotation of a submitted deck (quality score, never automatic correction); storyboard and generation.

Step 0 — always asked first: start from scratch, from the bank, from updating a specific presentation, from a scientific question (delegated to the synthesis skill), or resume a recent presentation (all its slides loaded into the basket).

Absolute rules

  • Never a full presentation, a batch of figures or media without approval for that step.
  • Never restrict the bank to a selection chosen by the AI: exhaustive indexing.
  • A single workshop, never an ad hoc page presented as a deliverable.
  • Absolute fidelity of reused slides; visual harmonization at the very end, on request.
  • Two figures actively sought per bibliography slide; never a flow diagram (CONSORT, PRISMA) in place of the central figure.
  • Online congress materials limited to the scope of the subscription.

Technical points: the reference script lives on the user’s side (to be checked at the start of the session); a table of journals with their impact factor kept up to date by hand; the result always opened in the browser; a shared database never opened directly (working copy, then copied back and verified by checksum).

3. “Online” variant of the presentation orchestrator

Status: separate from the personal skill, which remains unchanged; designed for any physician, with no machine-specific paths. Starting point for a distributable offering.

Triggers: only on explicit request for the “online” version.

Structure: scoping (new presentation, update, merge), bibliography, past presentations (local folder), media bank (local folder), analysis and annotation, storyboard with timed rehearsal mode.

Excluded: subscription congress materials (copyright) and patient imaging (health-data hosting not covered).

Technical constraint: local HTML workshop, no server; each action produces a line to copy into the conversation.

Quality: one figure per slide, figures always cited, fidelity of reused slides, four bullets at most, slide-by-slide review before delivery.

4. Bank of past presentations

Role: index past presentations as thumbnails, display them in a gallery, and copy the chosen slides exactly as they are, never reinterpreted.

Triggers: only “reuse, recycle, pick from” one’s old slides.

A single canonical index, always reused and refreshed incrementally (file date and size), never recreated elsewhere.

Item Rule
Language English on the slides
Format 16:9, never 4:3
Volume about one slide per minute
Citations “Author et al., Journal Year”, never a DOI on the slide
Fidelity reused slide strictly identical to the source
Harmonization at the very end, on explicit request
Final assembly always with explicit approval

Scope before producing: topic and angle, congress and format, audience, three take-home messages written before the slides, available material.

5. Single slide (normative template)

Status: normative source of the template, applicable to the whole chain.

Step 0 — always ask for the mode

Mode Content
0 · Verification no layout touched; citation and summary in the notes
1 · Layout reformat to the template with the images present
2 · Rebuild start entirely from the source publication
3 · Update add a second or third reference
4 · Variant layout specified by the user

Three absolute rules: no slide outside the template; no slide without the original figure extracted from the PDF (otherwise a flagged box and a list of PDFs to download); never large decorative numbers.

Template: white background, text on the left, figure on the right; bold title at the top with a colored rule; four to six short bullets, never full sentences; two figures stacked, never side by side; discreet citation at bottom right.

Mistakes already made, turned into rules: text too small, bullets as full sentences, too many numbers, invented subtitle, misplaced citation, “rebuild” mode that rephrases instead of starting from the source, style imported from a generic design skill. Never let a generic skill decide the style.

Automation: as soon as a source PDF is located, the figure is extracted and inserted, and the PDF is delivered alongside the slide.

6. PDF figure to slide

Role: replace screenshots with programmatic extraction, at constant resolution, from the actual PDF.

Steps: 1) local copy of the PDF, never a screenshot; 2) render the page at high resolution; 3) locate the rectangle to extract, and if it is ambiguous, present the options rather than guess; 4) crop and insert into the right column, stacked below an existing figure, citation added only if missing.

Boundaries: the bibliography has its own mechanism for open-access articles; this skill is for a non-open-access article or a specific figure.

7. Online congress materials

Strict scope: only the congresses covered by your subscription.

Real incident: an unfiltered search returned a majority of results outside the subscription. Fix: systematic filtering by event type, including retroactively.

Constraint: materials are served in a secure viewer; manual download, never bulk download by script. The skill automates everything that comes after.

Steps: filtered search → selection page ranked by number of slides → 16:9 presentation from the selected pages, full morphology.

8. Bibliography

Full cycle: PubMed search → interactive selection → download → filing → Word and PowerPoint deliverables, one slide per reference even without a PDF.

Mandatory attribution: source and DOI link for any data presented.

Folders: an active cumulative bank, never emptied; a read-only PDF archive, always checked before re-downloading, never written to.

Eight steps: scope (topic and bibliography name); exhaustive search; enrichment of metadata and impact factor; checkbox table (year and impact sliders, access badge); retrieval of the selection; download (open access first, a single batch per session); filing; Word and PowerPoint generation.

Paywalled articles: legal access tools, then the university’s institutional access. Never bypass a paywall or a captcha.

9. Literature watch

Difference from the bibliography: a permanent feed that never empties, versus a one-off review. Founding rule: never have to go back to an article to make a slide from it; the slide is produced right away.

Construction: one feed per major theme, plus a cross-cutting feed (methods, guidelines, AI). Measure the annual volume of each feed before setting an alert, and aim for a few references per week after deduplication.

Query rules: no impact-factor filter (PubMed does not allow it); quality comes from a whitelist of journals; no isolated ambiguous acronym; a systematic exclusion block (comment, editorial, erratum, retraction); tighten a feed that is too large.

Absolute rule: nothing is deleted; a discarded reference moves to the “discarded” status and stays archived.

10. Scientific synthesis

Role: reasoning layer between the bibliography and the presentation: a question becomes an evidence table, compared against the slide bank to spot what is missing or outdated.

Steps: scope the question (population, intervention, comparator, outcome, never a raw keyword); multi-source search; check of the existing bank; filtering; structured extraction (never an invented number); evidence table: stopping point, validation mandatory; gap identification; output in the presentation format.

Boundary: never replaces the presentation orchestrator; it is one of its entry points.

11. Clinical case

Output: a single file that grows case by case, regenerated by a script from data files, never rewritten by hand.

Block of four slides: divider (number, title, sex, age, context, never a name); initial presentation; additional tests; course and follow-up.

Layout: text on the left, figures on the right, font size adapted to the volume, telegraphic style; a fixed palette.

Anonymization: no surname, first name or date of birth on the slides; the internal identifier goes in the notes.

12. MRI imaging

Principle: export → series classification → black-background presentation, four animated loops per slide, never still images by default.

Series classification by their technical markers (localizers, cine, T1 mapping, flow, late gadolinium enhancement, console captures).

Standard deck: title, long-axis cine, short-axis cine, T1 mapping, late gadolinium enhancement. Excluded by default: localizers, flow, secondary captures (except numerical quantification).

13. Echocardiography image bank

Principle: a hierarchical bank (workshop → case → exam), a persistent cache and scripts (never in a temporary folder), a processing pipeline in stages (metadata, thumbnails, classification, photos and videos, themes).

Classification in mandatory order: strain, spectral Doppler, 3D, color Doppler, 2D, other.

Output: four videos per slide, black background, no visible text, technical information in the notes, looped playback; integration into the presentation workshop.

Watch out: check the identifier of each exam before any grouping; two folders with similar names may be the same exam, or two different patients.

14. Echocardiography visit by visit

Difference: native lossless video (GIF degrades color Doppler).

Standard five-slide deck: parasternal views; apical views; diastole; right ventricle; strain.

View identification: always read the label burned in by the machine, never guess from anatomical appearance. Option: comparison between visits.

15. Staff meeting

Principle: always start from the history of past staff meetings before producing anything.

Step 0: index of staff meetings (theme, volume, families covered, themes never addressed), the files themselves, the slide bank.

Mode Chain
0 · History only stops after step 0
1 · Reuse bank, then slide to re-source
2 · Literature review synthesis, then bibliography
3 · Specific opinion synthesis, then targeted bibliography
4 · Clinical case clinical case and imaging
5 · Scheduling calendar only

Format: short, about fifteen slides; beyond forty, it is no longer a staff meeting. End a review with a “what this changes for us” slide.

16. Faculty

Excludes the university diploma, which has its own skill.

The chain, always in this order: lecture to prepare → lecture delivered and uploaded → questions → exam-question review (docimology) → exam → deliberation.

Missions delegated to assistant professors: exam-question review (docimology), semiology, ECG tutoring, OSCE, each with its own content, pace and lead.

Anticipation: D-30 before a lecture (materials and questions ready), D-15 before an exam-question review (docimology) session.

17. University diploma

Typical rhythm: several seminars per year (in person and by video), exam, official results report.

Five modes: calendar for the year; mailing list (cross-checking results reports and email archives); in-person day; exam and results report (fixed format); daily watch of the dedicated mailbox.

Privacy: student names and grades stay in the local files, never copied into a published page.

18. Training watch

Role: pure watch, produces no content: collects, sorts, then hands off (diploma, faculty, presentation, clinical case).

Rules: access through an already logged-in browser, never entering credentials or signing up without validation; six items at most on the dashboard; a log so an item already seen is never flagged again.

19. Thesis defense

Difference from the review: the review is written; this skill prepares what will be said orally.

Automatic trigger: the weekly routine spots defenses at D-14 and produces the sheet at D-7.

Role Sheet
Chair full sheet; welcome remarks and closing remarks; synthesis questions
Examiner short sheet, two or three remarks
Supervisor background and supervision, lighter criticism

Golden rule: any item not found in a source is marked [TO BE CONFIRMED]; three reliable lines are better than ten plausible ones.

20. Manuscript review and follow-up

Two moments: the review (Word tracked changes and comments attributed to your name, never to “Claude”); lifecycle follow-up (co-author feedback, arbitration, versions up to submission).

Naming: each deliverable starts with the manuscript’s short name. One folder per manuscript: the untouched original, dated reviews, co-author feedback.

Scope: by default the scientific part in English, not the introductory chapters of a thesis.

Three levels: form and English (tracked changes); methodology (reasoned comment); reasoning (consistency of title, abstract, discussion), often the most useful.

Three flaws to check systematically: an outcome whose definition changes from one section to another; reference numbering that no longer matches the list; units that differ between text and tables (never decide alone).

21. CV

Sources: a central data sheet, identity, and PubMed for recent publications only.

Principle: never invent; missing data is marked [MISSING INFORMATION]; when sources contradict each other, present both and have the user decide. A decision already made is not asked again: write it into the skill.

Commands: generate, in English, update (with changes flagged), full version with identity and history.

22. Activity history

Difference from the CV: cumulative and detailed, enriched with each reported event, without waiting for a regeneration.

Strict attribution rule: collective work (for example a unit’s activity report) remains attributed to the collective, never merged with a personal achievement.

23. Annual reports

Absolute rule: the final document is always the previous year’s, duplicated then completed; the official template is never reformatted or reworded.

Log: one line per detected event (date, category, detail, source, status “to validate”, “kept”, “discarded”). Nothing enters the document without explicit validation, via checkboxes.

Strict exclusion: any consulting activity, even if the email goes through an academic mailbox.

Deadline: set a D-30 alert before the submission date.

Updated 15 Oct 2026

Academic map: real skills behind the agents

Each sub-agent on this site maps to skills actually in use, with their key rule.

30academic skills
12system skills
5orchestrators
5scheduled routines
5domains (A–E)
AgentSub-agentSkillsKey rule
OrganiserDashboardreacteurreacteur-ecrirearbitrageSingle hub: roadmap 7-9 days, work in progress, documents, mail decisions, horizons.
Mailgestion-mailcapture-agendaThree-column triage (reply / do-calendar / file); + aliases; nothing is ever written into the mailbox.
CalendaragendaPropose, then write only after « accept »; tentative events prefixed « ? », free, grey, no reminder.
Archivistclassementfichierrequete-discussionsarchivesProfile first, then file; never overwrite, never delete (archive instead).
Administrativeidentitecvhistoriquebilan-activiteForms and identity, CV from a data sheet + PubMed, activity history, annual statutory reports.
Methodpromptveille-bouclesentinelle-boucleaudit-orchestrateuraudit-interconnexionsDictation framed before acting; lessons read before each run; a separate read-only judge; weekly audit with evidence.
ResearcherBibliographybiblioveilleresearchPubMed search → impact-factor whitelist → PDFs → Word + slide deck; permanent watch; research layer turns a question into an evidence table and gaps in the slide bank.
ManuscriptarticleNative Word tracked changes attributed to the author; three review levels; co-author feedback arbitrated point by point until submission.
Datamegabase-cardiomegabase-cmhmegabase-pvmmegabase-requeteOne orchestrator, one sub-skill per cohort, one query skill (« how many patients with… ») behind a live artifact.
Congresspresentation-congrespresentationesc365-toposlidepdf-figure-vers-diapocongres-organisationatelier-topoSeven-tab workshop, indexed slide bank (never filtered by the AI), congress-platform slides, single-slide template, PDF figure → slide, door-to-door travel plan.
TeacherFacultyfacLecture → MCQ → docimology → exam → deliberation chain, with D-30 and D-15 rules.
Diplomadiu-echoMulti-site mailing list, calendar options, results minutes, a dedicated mailbox watched weekly.
Supervisionsoutenance-thesearticleSpeaking notes for a defense (chair or member) in one file; the written review is the manuscript skill.
Visual teachingcas-cliniqueechoechopacirm-cardiaqueFour-slide anonymised case blocks; DICOM → four looping clips per black slide, patient banner cropped.
TrainingformationWatch only (no purchase, no enrolment), hands over to the producing skills.
ClinicianStaff meetingsstaffStarts from the index of every staff deck since 2012; never repeats a topic unknowingly.
RotaplanningRead-only for routines; edits only in a supervised session with triple test before publishing.
CommonOrchestratoracademiqueLevel 1: qualifies (domain A–E), routes, never executes.

What changed since September 2026

  1. 1

    A shared reliability base for every skill

    Every action declares one of five states: to prepare, prepared, executed, verified, blocked. Referenced is not opened, opened is not analysed, a generated text is not a saved file, a deployed tool is not a tested one. Every result ends with four lines: Result · Location · Verification · Caveats. No verification line, no delivery.

  2. 2

    Framing a dictated request

    Dictated requests are decoded with a lexicon of phonetic slips, rewritten to a template (objective, context, sources, deliverable, safeguards, verification, target skill, assumptions) and checked against eight criteria. At most one grouped question, then Claude waits for “go”. A dictated sentence never creates a scheduled task.

  3. 3

    One queue for every pending decision

    Morning and evening runs drop what awaits a decision (events, replies, filing doubts) into a single numbered queue, checked beforehand against the calendar and the sender. The answer is a line of numbers: “accept 1 3 5, dismiss 2, tentative 7”. One grouped write, a line-by-line report, and the age of every pending item on the dashboard.

  4. 4

    A learning loop with a separate judge

    At the end of each working session, seven gestures record what was learned: a dated skill line, the agent card, a lesson (active only if asked), skill changes by proposal only, indexing, two-way links, umbrella projects. Before running a skill, Claude reads its active lessons. A read-only sentinel checks that the loop closed; it names gaps and never repairs them.

  5. 6

    Auditing the system: remove before adding

    A weekly audit walks every link in the system with evidence for each finding, mail coverage on a sample, project dormancy and the age of each alert. Two new checks: a task “succeeded” but left no log, and paths that no longer match the current computer. Skills are audited too: shorter descriptions (40–60 words), one rule file, and layers removed before new ones are added.

  6. 7

    Research database: linkage, coverage, longitudinal queries

    Patient records are linked by a score with three thresholds (certain, probable, doubtful); doubtful links stay in the database and are chosen at query time, and a placebo test measures chance matches. Source coverage is audited year by year. Any list of patients triggers a longitudinal query delivered as three files: identifying (stays on the computer), pseudonymised, and aggregated counts.

  7. 8

    Slides: count the figures before delivering

    After a talk was once delivered without a single image, every deck is checked on the file actually written: a table slide by slide (type, number of figures, status) and a final line “N slides, N figures, 0 slide without a figure”. Slide banks now keep one basket per talk, keyed by presentation and slide number, and group versions of the same content.

  8. 9

    Designing a hands-on course

    A single framing document: audience, prerequisites, operational objectives, an hourly schedule with the share of practice, modules (acquisition, analysis, reference values, pitfalls, report), small-group workshops on preloaded anonymised cases, identical pre- and post-tests, and a page that allocates the case bank to the time slots.

  9. 10

    One surface instead of three dashboards

    The morning brief, the anticipation cockpit and the mail decision board were merged into a single Reactor: tabs for daily roadmap, work in progress, documents, mail decisions, weekly / monthly / semester horizons, presentations (D-30 rule), teaching, training watch, literature watch. Every routine writes there; the person only decides.

  10. 11

    Mail aliases as commands

    Forwarding a message to your own address with a « + » tag turns it into an instruction the routines can act on: +agenda proposes a calendar event, +slide builds one slide from a paper, +claude runs a written instruction. Only messages you send yourself trigger anything; incoming mail never does.

  11. 12

    A file typology before any filing

    Every file is profiled on six axes (family, profile, origin, version, sensitivity, location) before it is moved. Doubt is a first-class outcome: the file stays where it is and a question is posted for the person. Ninety-six profiles, learned from each correction.

  12. 13

    A quality watch with indicator lights

    Routines, mail chain, disk, index and surfaces each have a light (ok / warn / crit) updated by the routines themselves at the start of every run. The sentinel reads the lights, never the platform's run timers, and alerts only when an action is expected.

  13. 14

    Lessons per skill, applied only after approval

    A weekly retrospective reads the week's journals and drafts lessons per skill and per project. Nothing is applied without an explicit « accept »; the lesson registry shows what is active, what was proposed and what was declined.

  14. 15

    Routines driven by plan files, not by prompts

    Each scheduled routine reads two files first (common rules, then its own plan) which take precedence over its prompt. Changing behaviour means editing a file, versioned and never overwritten, not recreating a task.

See how it all connects → Case studies