agentleFS
VorElzeVorElze/hermes-office-skills/skills/live-agent-office-operations/SKILL.md

live-agent-office-operations

Skill0 starsChanged 20 days agoMITView on GitHub

Use when debugging live agent office dashboards.

The file

We split this skill into the rules it gives an agent

9 rules. About 0 are common, 7 are distinctive, and 2 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈7

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 2

    Very specific

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

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Use this workflow for office dashboards (2D/3D) that visualize agent state, navigation, seating, or work activity.

    line 19
  2. Very specific:

    Never force a fixed height on pixel sprites: Always render with intrinsic aspect ratio (height:auto + width only).

    line 39
  3. Distinctive:

    Persist animated entity positions across periodic re-render: If the dashboard rebuilds avatars on a data refresh (e.g.

    line 41
  4. Distinctive:

    Do not treat a build timestamp, HTTP 200, or loaded asset chunks as proof of a visual fix.

    line 42
  5. Distinctive:

    Stop at the agreed request cap: If the operator set a cap (or complained about request spend), further probes/dispatches need a fresh …

    line 44
  6. Distinctive:

    Do not infer 3D lounge/platform reachability from a 2D map that omits platform geometry or elevation.

    line 45
  7. Distinctive:

    Do not label a dependency-waiting TODO as a stuck worker.

    line 46
  8. Very specific:

    Do not call a task healthy solely because it emits heartbeats.

    line 47
  9. Distinctive:

    Do not invent precise completion percentages from intuition.

    line 48

≈ “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

Live Agent Office Operations

When to Use

Saat membangun atau memelihara dashboard kantor agen 2D/3D, khususnya memeriksa aktivitas live, status agen, navigasi/avatar, atau meminta implementasi ulang memakai aset yang diberikan.

Use this workflow for office dashboards (2D/3D) that visualize agent state, navigation, seating, or work activity. Keep evidence tied to the exact runtime and agent; a healthy page or plausible screenshot alone does not prove correct activity.

Procedure

  1. Identify the exact product before describing or changing it. Inspect the user-provided screenshot, archive, project files, and prior-session records where available; state what is directly verified versus inferred. Do not rename a project or claim its stack from visual resemblance alone. For an asset archive, inspect its README and license and inventory image/sprite formats before proposing implementation.
  2. Agree the request budget first when quota is tight. Give a short cost estimate (main-session probes vs worker dispatches), let the operator approve a cap, prefer cheap read-only probes over browser rounds and worker dispatches, and check the cap before each build/restart or new dispatch. On stop/cancel, halt immediately — stop dispatches and project services, do only the cleanup actually requested, and report verified final state.
  3. Establish the active task graph. Inspect the Kanban root and all decomposed children, their parent/dependency links, assignees, latest heartbeat/event/comment, retry/failure state, and worker process. Distinguish a child waiting on a prerequisite from an idle worker. Check for explicit loop/block events and operator-input waits.
  4. Build the layout yourself; the asset pack's example scene is only reference. When a user supplies a sprite/asset archive, slice it into individual sprites (PIL region detection or manual crops), but compose the office map from scratch — the archive's demo scene must never become the layout. Verify each sliced sprite visually before placing it.
  5. Capture a baseline before editing. Record service/build identity, HTTP response, live agent count/state, and a fresh screenshot. If onboarding obscures the canvas, close its accessible control before capture (in Hermes3D: the [aria-label="Close onboarding"] element — Playwright's role-based lookup misses it because it is a div, so click by attribute). Do not infer positions or poses from HUD counters.
  6. Prove the specific failure path. For pathfinding, probe start and target from the same live agent snapshot. Include cases where start and goal occupy different grid cells, blocked-goal behavior, final waypoint, pixel distance, and the consumer's arrival threshold/state transition. A degenerate same-cell path (findFree snapping both cells to one) does not prove arrival works.
  7. Compare navigation geometry with rendered geometry. Derive furniture bounds after the same pivot/rotation/scale transforms used by rendering (AABB of the actual GLB, not the canvas footprint). Confirm the target lies at the rendered chair and outside desk/monitor bounds. Keep 2D-map reachability separate from 3D platform/elevation/collision evidence.
  8. Instrument narrowly when needed. Capture agent ID, state, current position, target, path length, and distance to target from the live runtime. Prefer temporary probes over speculative code changes; remove debug hooks before the final build.
  9. Change only after root cause is supported by concrete evidence. Preserve working agents and unrelated navigation behavior. If multiple causes are plausible, run discriminating probes first rather than stacking speculative fixes.
  10. Verify in layers. Run typecheck/build, restart the intended service, verify active process/build ID and HTTP response, then inspect a fresh visual capture. Confirm the actual named agent is sitting in the rendered chair facing the workstation, not on furniture or stuck at a platform edge. For 2D-only builds, verify sprite/map rendering and interaction without loading or requiring 3D models. Report incomplete visual checks plainly.
  11. Prove movement by measurement, not screenshot. Sample element positions (e.g. style.left) twice a few hundred ms apart and show the delta; for worker-to-desk behavior check the state attribute (mode=sit) and the badge color against the legend. Verify the busy worker persists in its seated position across the dashboard's periodic re-render (positions must survive the refresh that rebuilds avatars).
  12. Report progress and health honestly. Give timestamp, root/child task IDs and statuses, measurable completed/remaining acceptance criteria, concrete evidence (file:line, numbers, counts), issues/actions, and a conditional ETA range. Before saying nothing is stuck or looping, inspect heartbeat ages, explicit loop events, failures/retries, and blocked-on-input state. Fresh heartbeats prove activity, not correctness or progress. Keep interval reports concise and only fire on real change.

Pitfalls

  • When a delegated build fails mid-task (provider outage, timeout), verify what actually changed on disk before reporting — check file mtimes against the run window and grep for the requested feature. Prepared assets (sliced sprites) often survive while the main file is untouched; salvage those and finish in the main session rather than re-dispatching the whole task.
  • Never force a fixed height on pixel sprites. Always render with intrinsic aspect ratio (height:auto + width only); forced height stretches/warps objects and reads as "berantakan" (messy) to the user. Place furniture on shared baselines (same top value for a row) to look intentional.
  • Route 2D characters through corridors, not straight lines. A straight start→desk line cuts through walls; send the character to a corridor waypoint on the room boundary first, then to the desk. The sitting target is the chair position, offset below/in front of the desk — not the desk center.
  • Persist animated entity positions across periodic re-render. If the dashboard rebuilds avatars on a data refresh (e.g. every 20s), stash each character's position/mode in a runtime map and restore it on rebuild, or workers teleport and the busy worker stops looking seated.
  • Do not treat a build timestamp, HTTP 200, or loaded asset chunks as proof of a visual fix; they establish deployment/availability only.
  • Trace the live-state path end-to-end before blaming the renderer. Log agent IDs and counts at every boundary (settings/config key → presence request route → gateway/proxy connection → response payload → agent mapping → rendered array → HUD header/avatars). A placeholder/demo roster or a zeroed counter means the data path returned empty or failed upstream of render — a separate probe proving data at the gateway does NOT prove the UI path carries it. Match preferences against the key the runtime actually reads (a stored preference under a differently-normalized URL is dead config), and never accept seeded/random state or demo roster as fallback for a dashboard meant to show real activity.
  • Stop at the agreed request cap. If the operator set a cap (or complained about request spend), further probes/dispatches need a fresh agreement; re-verifying the same finding with new evidence costs more than reporting the existing one honestly.
  • Do not infer 3D lounge/platform reachability from a 2D map that omits platform geometry or elevation.
  • Do not label a dependency-waiting TODO as a stuck worker; verify the task graph and parent completion condition first.
  • Do not call a task healthy solely because it emits heartbeats; detect explicit loop events (block_loop_detected) and waits-for-input as separate states.
  • Do not invent precise completion percentages from intuition. Tie them to acceptance criteria and label estimates as estimates.
  • Screenshot capture on continuously animating WebGL pages can time out — retry with a longer timeout or animations="disabled", falling back to CDP Page.captureScreenshot.

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.