agentleFS
LaansDoleLaansDole/my-hermes-skills/skills/ilearning-autoadvance/SKILL.md

ilearning-autoadvance

Skill3 starsChanged 10 days agoMITView on GitHub

Autonomously advance HCL iLearning / SAP SuccessFactors compliance courses (Articulate Storyline SCORM, Ethena GoEthena, or direct-HTML SVG players) via raw CDP on port 9222. Resumes paused audio, clicks interactive slide topics/cards, answers quizzes, navigates to the next slide.

The file

Show its rules and text

We split this skill into the rules it gives an agent

16 rules. About 0 are common, 8 are distinctive, and 8 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈8

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 8

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Recall before re-probing: When resuming a course, run mnemosyne_recall(query="<course> quiz format") first — the stored layout eliminates …

    line 39
  2. Distinctive:

    Store per-format facts, not per-question trivia: Only durable layout/flow facts belong in memory.

    line 40
  3. Very specific:

    Run with /usr/bin/python3 (bare python3 in background shells lacks websockets):

    line 117
  4. Distinctive:

    No action. Sleep, re-probe.

    line 142
  5. Very specific:

    All clicked → fall through to SLIDE_DONE.

    line 150
  6. Distinctive:

    Never submit twice for the same question (track submitted keys).

    line 159
  7. Distinctive:

    Transitions can take >10s: Wait 4s between clicks; clear per-slide clicked sets ONLY on real slideId/page change (TRANSITION).

    line 165
  8. Distinctive:

    If unchanged after 10 next-clicks → screenshot + vision, one alternative affordance, else hand off (STUCK).

    line 166
  9. Distinctive:

    Stop when any of:

    line 213
  10. Very specific:

    No Next affordance for 60s (12 ticks) after SLIDE_DONE.

    line 215
  11. Very specific:

    Never type into any type="password" field.

    line 227
  12. Distinctive:

    Never click anything whose accessible name contains: logout, sign out, sign-out, signout, log off.

    line 228
  13. Very specific:

    If location.hostname is not hclt.lms.hr.cloud.sap (or app.goethena.com inside the nested iframe), stop and surface.

    line 229
  14. Very specific:

    Do NOT navigate away with browser_navigate.

    line 230
  15. Distinctive:

    Do NOT close tabs or the browser.

    line 231
  16. Very specific:

    Format C (direct-HTML SVG, AppSec): no iframe; NEXT/PREV real buttons at bottom-right.

    line 288

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

iLearning Auto-Advance

Autonomously progress through HCL iLearning / SAP SuccessFactors compliance courses. These run on hclt.lms.hr.cloud.sap in one of THREE player formats (see Course Format Registry). All formats share the same core rule set below.

THE USER'S CARDINAL RULES (follow exactly)

  1. "Unless it is a quiz, you only need to find clickable items when the slide mentions 'Click...' itself. Otherwise it is just a normal slide you can 'Next'. Always try 'Next' first."

  2. "You should always prioritize the 'Next' button on top of anything." — NEXT (or Continue) enabled → click it BEFORE considering any other clickable element on the slide.

  3. "Stop sleeping when Next button cannot be clicked for 10s" — if NEXT stays disabled/unclickable for ~10s, hand off (STUCK) instead of polling forever. Long narrations auto-play and usually resolve it, but a mid-paused narration will NOT — hand off so the agent can resume it or surface to the user.

Decision tree (all formats):

  1. Quiz? (radio/checkbox inputs present, or title contains Quiz/Question/Assessment) → answer it.
  2. NEXT/Continue enabled? → click it. Always first.
  3. Slide text mentions "Click..."? → click each target once, wait for audio between clicks.
  4. NEXT exists but disabled → wait up to 10s, then hand off.
  5. Otherwise → screenshot + vision, hand off UNKNOWN.

Second Brain Optimisation (Mnemosyne)

This skill pairs with the local-first Mnemosyne memory (the user's second brain) to make repeat courses progressively faster:

  • Persist quiz slide layouts to memory. After confirming a course's quiz format (option-row Y coordinates per option count, SUBMIT/Continue/Try Again positions), save it via mnemosyne_remember(..., importance=0.7, scope='global') — e.g. the Format C layout (2-opt y≈382/442; 3-opt y≈386/446/506; 4-opt y≈360/421/481/541; SUBMIT (1383,802); popup ≈(756,531)). Subsequent questions on the same course can then be answered WITHOUT re-screenshotting: read the question via a quick screenshot, click the option at the memorized row, submit, continue.
  • Recall before re-probing. When resuming a course, run mnemosyne_recall(query="<course> quiz format") first — the stored layout eliminates per-question rect discovery.
  • Store per-format facts, not per-question trivia. Only durable layout/flow facts belong in memory; per-question answers are one-off and belong in the session transcript (session_search).
  • Memory is the shortcut, the skill is the method. The skill's probes/orchestrators handle the how; Mnemosyne's stored layouts skip the where on the next run.

Course Format Registry (proven 2026-08-13)

Format A — Storyline SCORM in iframe (e.g. "Information Security", "COBEC" Storyline courses)

  • Player tab URL contains scorm2004contentplayer; course content is in ONE <iframe> (fills viewport 1920x1027 → 1:1 coordinates).
  • Slide objects: <div class="slide-object"> with data-model-id; real click targets are ancestor .slide-object-stategroups (use clickRect/clickMid from probe).
  • Nav: Next.png/Backbtn.png PNG slide-objects, embedded nav at y≈718 or global chrome y≈915. Read rects, never hardcode.
  • JS-dispatched mouse events DO NOT register — only trusted CDP Input.dispatchMouseEvent works.
  • Quiz: option radios id="acc-<modelId>" (labels useless, usually "Rectangle 3") → screenshot + vision to read; submit submit0001.png appears after selection; feedback popup green next0001.png ≈(750,612)/(750,635).
  • Driver: ilearn_orchestrator.py + ilearn_probe.js. Orchestrator states: QUIZ / STUCK / STOP / UNKNOWN / DONE / CLICK_REPEAT.

Format B — Ethena / GoEthena course (e.g. "Anti-Money Laundering Basics", "Export Controls")

  • Player tab URL contains scorm2004contentplayer; inside it a NESTED cross-origin iframe https://app.goethena.com/learning/assignments/... (chain: hclt page → icontent.do?url=... iframe → goethena iframe).
  • The GoEthena iframe is its OWN CDP target with "type": "iframe" — the standard helpers match type == "page" only. Use the iframe-capable variants:
    sed 's/t.get("type") == "page" and match in t.get("url", "")/t.get("type") in ("page", "iframe") and match in t.get("url", "")/' cdp_eval.py > /tmp/cdp_eval_if.py
    sed 's/t.get("type") == "page" and match in t.get("url", "")/t.get("type") in ("page", "iframe") and match in t.get("url", "")/' cdp_click.py > /tmp/cdp_click_if.py
    Target match: goethena.com/learning/assignments. Coordinates are the iframe's own viewport space.
  • Player: simple content slides with a blue Continue button. On long slides the Continue sits BELOW THE FOLD — scroll the slide container (.Slide_container__Zpm4v) to bottom first, then click. Scroll via JS on the iframe target (e.scrollTop = e.scrollHeight on scrollable containers).
  • Quiz: per-section assessments (3 Qs, 2 Qs, 1 Q...) with radio inputs + Submit Answer button. Feedback popup: correct → Continue; incorrect → View Answer (shows correct option green) then NEXT/Continue. Optional feedback survey at the end (Likert 1-5 + tag chips) — submit to finish.
  • Some courses end with a policy-attestation CHECKBOX section — user does that manually; stop when reached (ask first if unclear).
  • Driver: goe_orchestrator.py (probe → scroll if needed → click Continue; hands off on QUIZ).

Format C — Direct-HTML SVG slide deck (e.g. "Application Security Foundation Certification" / Appsec_may12)

  • Course opens at a plain URL like .../icontent_e/CUSTOM_fra/hcl/self-managed/LSS/Appsec_may12/index_lms.html — NO iframe, slides are inline SVG in the page DOM. Target match: index_lms.html.
  • Player bar bottom-right: PREV (1260,786) / NEXT (1345,786) real <button class="cs-button btn">. Disabled state is the CSS class cs-disabled, NOT the HTML disabled attribute — always check className.includes('cs-disabled').
  • Per-slide <audio> elements (do not reliably auto-play; NEXT stays cs-disabled until narration completes — WAIT_NEXT handles this; if audio is mid-paused, clicking the Play button (82,786) resumes it).
  • Slide number in footer: <n> Copyright © 2024 on one line (nbsp-separated) → regex /(\d+)\s*(?:\u00a0|\s)*Copyright/i.
  • Quiz: "Question X of 25" top-right, interleaved with content slides (NOT one block at the end). Purple banner question; radios left-aligned x≈130-147, w≈796-901, h=63; rows depend on count:
    • 2 options: y≈382/442
    • 3 options: y≈386/446/506
    • 4 options: y≈360/421/481/541 (or 352/447/542/602 when options are tall)
    • option centers ≈ (537, row+20)
  • SUBMIT at (1383,802), cs-disabled until an option is selected.
  • Feedback popup (center): correct → Continue ≈(756,531); incorrect → Try Again ≈(756,531) (retry SAME question; the "View Answer" pattern is Format B only). Sometimes incorrect just shows the green-highlighted correct answer + NEXT.
  • Hyperlink <a> elements are NOT course interactions — exclude from clickable detection.
  • Driver: appsec_orchestrator.py + appsec_probe.js-style probes (in skill scripts/). States: QUIZ / STUCK / DONE / UNKNOWN / INTERACTIVE. NEXT-first priority and 10s WAIT_NEXT timeout are baked in.

Parameters

  • max_modules (int, default 50): hard cap on the number of slides/modules to advance through before stopping.
  • dry_run (bool, default false): when true, observe and describe the action you would take, but do NOT call any click/type/fill tool.
  • tick_seconds (int, default 5): idle polling interval. After an action, poll again immediately (the page needs time to react).

Prerequisites

  • Brave (or Chrome) running with --remote-debugging-port=9222 (profile brave-live — switch to it with browser_profile(name='brave-live')).
  • User logged into SAP SuccessFactors LMS (hclt.lms.hr.cloud.sap).
  • Course open: overview page shows a "Launch Content" UI5 button (Formats A/B) or the direct content URL (Format C). Clicking Launch opens the SCORM player in a NEW TAB; the overview tab keeps a disabled Launch button.

CRITICAL: drive the player over raw CDP, not browser_console

Hermes browser tools stay pinned to the tab active when the profile connected — usually the OVERVIEW tab, not the player tab. Use raw CDP over websocket:

curl -s http://localhost:9222/json/list | python3 -c "
import json,sys
for t in json.load(sys.stdin):
    if t.get('type')=='page' and 'scorm2004contentplayer' in t.get('url',''):
        print(t['id'], t['webSocketDebuggerUrl'])"
# Format B: also list iframe targets
#   if t.get('type')=='iframe' and 'goethena' in t.get('url',''): ...

CDP Helper Scripts

In scripts/ (copy to /tmp when working):

  • cdp_eval.py / cdp_click.py / cdp_shot.py — page-target helpers (Formats A/C).
  • ilearn_probe.js / ilearn_orchestrator.py — Format A.
  • goe_orchestrator.py — Format B (iframe-target; also make cdp_eval_if/cdp_click_if via sed, see Registry).
  • appsec_orchestrator.py / appsec_probe.js — Format C (page-target, direct DOM).
  • quiz_rects.js — map quiz radio ids to real .slide-object click rects (Format A).

Run with /usr/bin/python3 (bare python3 in background shells lacks websockets):

cd /Users/laansdole/.hermes/skills/ilearning-autoadvance/scripts
/usr/bin/python3 cdp_eval.py "<target-match>" ilearn_probe.js
/usr/bin/python3 cdp_click.py "<target-match>" X Y
/usr/bin/python3 cdp_shot.py "<target-match>" /tmp/shot.png

Watch Loop

Repeat until a stop condition (see Loop Control) fires:

  1. Observe — one probe call (cdp_eval.py + the format's probe JS).
  2. Classify into: COURSE_OVERVIEW, TIMELINE_PLAYING, TIMELINE_PAUSED, INTERACTIVE_SLIDE, QUIZ_SLIDE, SLIDE_DONE, UNKNOWN (Formats A/C); Format B uses its own probe (Continue button presence).
  3. Act per the action policy below.
  4. Sleep tick_seconds if no action was taken; otherwise repeat immediately.

Action Policy

COURSE_OVERVIEW

  1. Click "Launch Content" (opens player in a NEW tab; native .click() works).
  2. Re-enumerate tabs; re-target helpers on the player tab (or the goethena iframe target for Format B).
  3. Re-probe.

TIMELINE_PLAYING

No action. Sleep, re-probe. Narrations can be 100s+.

TIMELINE_PAUSED

Find Play/Resume and click via CDP. If none, re-classify (may be SLIDE_DONE). In Format C the Play button is at ≈(82,786).

INTERACTIVE_SLIDE

  1. Targets = clickables minus (instruction banner, slide title, header band y < 100, background stategroups, nav, hyperlinks).
  2. For each target not in the clicked set (key = PHYSICAL rect "x,y,w,h"): click its center; hard guard: any key >1 click → STOP (CLICK_REPEAT); wait for audio between clicks.
  3. All clicked → fall through to SLIDE_DONE.
  4. Persist per-slideId clicked sets to ~/.hermes/logs/ilearning-autoadvance-resume.json.

QUIZ_SLIDE

  1. Screenshot + vision to read the question and options (DOM labels are usually useless).
  2. Reason the correct answer from course content; common patterns: "All the above" when all options are valid; "NOT true"/"NOT a good practice" picks the single bad option; scenario questions pick the policy-compliant action.
  3. Map radios to click targets per format (Format A: acc-<modelId> → slide-object rect; Format B: walk up to the clickable container; Format C: use the memorized row layout, or re-probe rects).
  4. Click option center, then Submit/SUBMIT (Format A submit0001.png ≈(741,693); Format B Submit Answer; Format C (1383,802)). Submit appears only after selection.
  5. Feedback popup: Continue ≈(750,612)/(756,531) to proceed; Try Again ≈(756,531) to retry an incorrect answer (Format C); View Answer (Format B) shows the correct option.
  6. Never submit twice for the same question (track submitted keys).
  7. Assessments can be interleaved with content (Format C: "Question X of 25" spread through the deck) — after the last question the course returns to content slides; restart the orchestrator.

SLIDE_DONE

  1. Find Next/Continue (PNG Next.png in clickables, or button with text NEXT/Continue).
  2. Click center via CDP. Record last_slide_id BEFORE the click.
  3. Transitions can take >10s. Wait 4s between clicks; clear per-slide clicked sets ONLY on real slideId/page change (TRANSITION).
  4. If unchanged after 10 next-clicks → screenshot + vision, one alternative affordance, else hand off (STUCK).
  5. Some slides' embedded nav acts as play/resume — if playing goes 1 after a next-click, wait for audio, click Next again.
  6. Increment modules_advanced; stop at max_modules.

Clicking Slide Objects (CRITICAL)

JS-dispatched MouseEvents DO NOT register in this Storyline build (Format A). Only trusted CDP Input.dispatchMouseEvent works:

await send("Input.dispatchMouseEvent", {"type": "mousePressed", "x": x, "y": y, "button": "left", "buttons": 1, "clickCount": 1})
await send("Input.dispatchMouseEvent", {"type": "mouseReleased", "x": x, "y": y, "button": "left", "buttons": 0, "clickCount": 1})

Click identity = PHYSICAL LOCATION (f"{x},{y},{w},{h}"), never modelId/acc-text — duplicate DOM layers of one card share one rect; id-keyed tracking causes same-card double-clicks.

Pitfalls:

  • Click the stategroup, not the label (Format A).
  • Skip background stategroups whose clickRect covers >50% of the slide.
  • Offscreen decoys at x≥1920 (Format A probe filters them).
  • cs-disabled class = disabled (Format C) — check className.includes('cs-disabled'), not just the attribute.
  • Zero-size SVG buttons → click parent .slide-object rect; else screenshot + vision.
  • Popup overlay freeze (z=9999 full-slide container at popup position) → surface to user for reload.

Screenshot & Vision

cdp_shot.py "<target>" /tmp/shot.png, then vision_analyze (ask for: question text verbatim, all options verbatim, instruction text, submit button, popup, button colors/positions). Use on every QUIZ_SLIDE, on UNKNOWN > 2 ticks, and after stuck Next.

Resume State

Persist per-slide click history per slideId to ~/.hermes/logs/ilearning-autoadvance-resume.json. Seed clicked sets on startup; save after every card click. Delete a slideId's entry if it was recorded from a buggy run.

Orchestrator Patterns

  • Format A: ilearn_orchestrator.py — MATCH="scorm2004contentplayer", TICK=5, STUCK_AFTER=10 (transitions slow), NO MENU_SLIDE (instructions-only trigger), is_background(), physical-rect dedupe, CLICK_REPEAT guard, per-slideId resume.
  • Format B: goe_orchestrator.py — iframe target; scroll-to-bottom then click Continue; QUIZ handoff on inputs.
  • Format C: appsec_orchestrator.py — NEXT-first priority; WAIT_NEXT with 10s timeout (hand off STUCK after 2×5s ticks); cs-disabled check; hyperlink exclusion; QUIZ handoff on inputs.
cd /Users/laansdole/.hermes/skills/ilearning-autoadvance/scripts
/usr/bin/python3 <format>_orchestrator.py 50 <runid>
# background with notify_on_complete: true

ALWAYS kill all previously-started orchestrator background sessions before starting a new one — stale processes from older code keep clicking and corrupt state.

Loop Control

Stop when any of:

  • modules_advanced >= max_modules.
  • No Next affordance for 60s (12 ticks) after SLIDE_DONE.
  • last_3_slide_ids identical AND last action was a Next click (infinite loop).
  • Page navigates outside hclt.lms.hr.cloud.sap (or to a report/thanks page = course finished).
  • CDP connection dropped.
  • unknown_streak > 6.
  • Session/login expired.
  • Any card's click_counts > 1 (CLICK_REPEAT).
  • NEXT disabled for 10s+ (WAIT_NEXT timeout → STUCK handoff).
  • User-stopped boundary: policy-attestation checkbox section (user completes manually).

Safety Rules

  • Never type into any type="password" field.
  • Never click anything whose accessible name contains: logout, sign out, sign-out, signout, log off.
  • If location.hostname is not hclt.lms.hr.cloud.sap (or app.goethena.com inside the nested iframe), stop and surface.
  • Do NOT navigate away with browser_navigate; interact within the current tab only.
  • Do NOT close tabs or the browser.

Logging

JSONL per run in ~/.hermes/logs/ilearning-autoadvance-<runid>.jsonl with ts, state, slide/heading, action, page, audio_playing/audioPlayingCount, ok. Include per-card counts in click lines (user reviews these to verify no double-clicks). Print a summary at loop end.

Error Recovery

FailureDetectionRecovery
CDP connection droppedRuntime.evaluate errorStop, log, surface.
Session/login expiredSSO redirect / iframe missing >30sStop, surface. No auto-relogin.
Click didn't registerslideId unchanged after clickRe-probe; click stategroup center; verify in-viewport; screenshot+vision. Transitions can take >10s.
Card re-clickedclick_counts[key] > 1HARD STOP (CLICK_REPEAT) — fix dedupe/state bug before resuming.
Clicked sets cleared wronglysets reset on failed NextOnly clear on real slideId change (TRANSITION).
Quiz submit disableddisabled/aria-disabled/cs-disabledRe-scan for unanswered questions, answer, retry.
Quiz option labels uselesslabels "Rectangle 3" (Format A)Screenshot + vision; radio id → modelId → rect.
NEXT disabled longcs-disabled / no click effect for 10sHand off STUCK; agent resumes audio (Play button) or surfaces.
Feedback popup Next unreachableelementFromPoint hits z=9999 overlayPlayer frozen — surface to user for reload.
Format C quiz wrong answerpopup says Incorrect + Try AgainClick Try Again (756,531), select the other option, resubmit.
Next click starts narrationplaying goes 1 after next-clickWait for audio, click Next again.
Stale orchestrator processduplicate clicks in logprocess list + kill ALL, restart current code.
Infinite loopsame slide_id 3+ actionsStop, log, surface.

Key Lessons Learned

2026-07-19 session (Format A basics)

  1. Probe, classify, act, repeat — ONE probe per tick. Don't over-investigate Storyline internals.
  2. acc-shadow-el/#acc-<modelId> are accessibility mirrors — clicking them does nothing. Click the real .slide-object.
  3. Audio playback is the progress indicator; playing > 0 = timeline advancing.
  4. Interactive slides require clicking each topic and waiting for its audio before Next enables.
  5. Next is often a zero-size SVG inside a .slide-object div — click the parent.

2026-08-13 session (v3.0.0 — Format A hard truths)

  1. JS-dispatched mouse events DO NOT register — only trusted CDP Input.dispatchMouseEvent.
  2. Drive the player tab over raw CDP — Hermes browser tools stay pinned to the overview tab.
  3. Nav buttons are PNG slide-objects; embedded nav at y≈718 vs global chrome y≈915.
  4. Interactive targets are stategroups (click the ancestor, not the label).
  5. Duplicate DOM layers + offscreen decoys at x≥1920 — filter to viewport, dedupe by physical rect.
  6. Never re-click a card — log counts, hard-stop on >1 (user requirement).
  7. Clear clicked sets only on real slide transitions.
  8. Narrations can be 110s+ — TIMELINE_PLAYING is often just a long wait.
  9. Video slides play separately from audio MP3s.
  10. Kill all stale orchestrator background sessions before restarting.

2026-08-13 full-course completion (v4.0.0 — user's rules)

  1. Interactive targets exist ONLY when slide text says "Click..." — otherwise just Next. No pointer-cursor fallback, no menu heuristics.
  2. Click-tracking key = physical rect (x,y,w,h), never modelId/acc-text.
  3. Filter background stategroups (>50% slide); keep static-named elements ("Oval 2") with a stategroup clickRect — they ARE targets.
  4. Quiz flow (used ~12x): screenshot+vision; radio id → slide-object rect; submit appears after selection; popup next ≈(750,612); never submit twice.
  5. Final Assessment = 10 questions, 80% pass, separate scene; then Results → Thank You → certificate.
  6. Transitions can take >10s — STUCK patience ~10 next-clicks with 4s spacing.
  7. Embedded nav at y≈718 can act as play/resume — wait for audio, click Next again.
  8. Popup overlay freeze (z=9999) → surface to user for reload.

2026-08-13 Ethena + AppSec courses (v5.0.0 — new formats)

  1. Format B (Ethena/GoEthena): nested cross-origin iframe = its own "type":"iframe" CDP target — use iframe-capable helpers (sed-variant of cdp_eval/click). Scroll the slide container, then click Continue. Per-section assessments with Submit Answer; incorrect → View Answer; ends with optional feedback survey. Some courses end with a policy-attestation checkbox the USER completes.
  2. Format C (direct-HTML SVG, AppSec): no iframe; NEXT/PREV real buttons at bottom-right; disabled = cs-disabled CSS class (not the attribute); audio must finish before NEXT enables; slide number <n> Copyright in footer. 25-question assessment INTERLEAVED with content slides ("Question X of 25").
  3. NEXT-first priority (user rule): enabled NEXT beats any clickable item — always click NEXT first.
  4. 10s wait cap (user rule): NEXT disabled → wait ~10s, then STUCK handoff — never sleep forever. Long narrations resolve themselves; mid-paused narrations need the agent (resume Play) or the user.
  5. Quiz layout memory (Format C): 2-opt y≈382/442; 3-opt y≈386/446/506; 4-opt y≈360/421/481/541 (x≈130-147, centers ≈(537, row+20)); SUBMIT (1383,802); Continue/Try Again (756,531). Wrong answers → Try Again to retry.
  6. Course completion signal: player reaches "Thank you"/"Completed" screen, or the LMS portal tab shows courseCompleted=true / a certificate-report page — stop the orchestrator there.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.