agent-ops
/agent-ops
Opisuje dyscyplinę pracy agenta: budżetowanie tokenów, pamięć między sesjami, pętle samodoskonalenia, samoweryfikację, pracę równoległą i orkiestrację subagentów.
- Autor
- krishna-modi12 ↗
- Repozytorium
- frontend-design-pro ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 20.10.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 14
- Skrypty
- 3
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 20.10.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
- 3 skrypty w paczce. Przejrzyj je przed instalacją, bo Claude może je uruchomić.
Skrypty: examples/_stubs.d.ts, examples/good-agent-status-panel.test.tsx, examples/good-agent-status-panel.tsx
Co robi i kiedy
Opisuje dyscyplinę pracy agenta: budżetowanie tokenów, pamięć między sesjami, pętle samodoskonalenia, samoweryfikację, pracę równoległą i orkiestrację subagentów.
Oryginalny opis (EN)
Agent operating discipline — token budgeting, cross-session memory, self-improvement loops, self-verification, parallel work, and subagent orchestration. Use when the work is the agent's own process rather than the UI it produces — deciding what to load, persisting a correction so it sticks, verifying a change before reporting done, or dispatching subagents. Not for building dashboards or persistence UI.
Jak działa
- When to Use
- Core Rules
- Patterns
- Examples
- Reference Index
- Constraints
Podgląd SKILL.md
---
name: agent-ops
description: Agent operating discipline — token budgeting, cross-session memory, self-improvement loops, self-verification, parallel work, and subagent orchestration. Use when the work is the agent's own process rather than the UI it produces — deciding what to load, persisting a correction so it sticks, verifying a change before reporting done, or dispatching subagents. Not for building dashboards or persistence UI.
metadata:
version: "15.0.0"
core-deps:
- core/agent-behavior.md
- core/validate-checklist.md
---
# Agent Ops
## When to Use
Any task where the work is the agent's own process, not the UI it produces: deciding what context to load, persisting facts across sessions, incorporating a correction so it sticks, verifying a change actually works before reporting done, deciding whether to parallelize, or dispatching/integrating subagent work. This skill governs how the agent operates — it does not describe dashboards or persistence UI (those live in other skills' reference files and are out of scope here).
## Core Rules
1. **Load lazily, cite precisely.** Read a reference file because the current task needs it, not "just in case." Re-reading a file already in context wastes the same tokens twice.
2. **Persist decisions and corrections, not scratch state.** A durable project fact (a naming convention, a rejected approach) belongs in memory; a task's intermediate variables do not survive the task.
3. **A correction changes the rule, not just the instance.** If a user fixes the same class of mistake once, update the standing approach — do not wait to be told a second time, and do not overfit a whole new policy to a single data point.
4. **"I wrote it" is not "it works."** Verification is a separate, explicit step — compile, run, or re-read the diff — before the work is reported as done.
5. **Parallelize only disjoint work.** Independent subtasks with no shared file and no ordering dependency can run concurrently; anything touching the same file or state serializes.
6. **Own the integration point.** When work is split across subagents, the human-facing owner personally merges, verifies, and reconciles conflicts — that step is never delegated away.
## Patterns
- **Budget before you read.** State what the task needs before opening files; skip references whose content the task doesn't touch.
- **Summarize, don't hoard.** Once a long tool result has served its purpose, keep the conclusion and drop the raw output from further reasoning.
- **Memory has a shelf life.** Every persisted fact needs an implicit or explicit expiry condition — a fact tied to a file that no longer exists is stale, not true.
- **Verify at the seam.** Run the actual compiler/linter/test command; do not infer a pass from having made a plausible-looking edit.
- **Batch independent calls, serialize dependent ones.** Multiple read-only lookups in one turn; a write followed by a read of that same write in two turns.
- **Ground subagent prompts in facts, not vibes.** Give a dispatched agent the concrete file paths, line numbers, and constraints already known — never "figure it out yourself" when the answer is already in hand.
## Examples
`examples/good-agent-status-panel.tsx` — a token-budget meter and subagent task-queue status panel: shows per-reference token cost, a load/skip decision per file, and per-subagent state (queued/running/verified/failed), illustrating token budgeting, parallel dispatch, and the "verified" distinction as rendered UI state.
## Reference Index
Load only for the specific task:
| Task | Load |
|---|---|
| Deciding what to load/skip, measuring cost, avoiding re-reads, caching | `references/token-optimization.md` |
| Token-efficient wire format for bulk structured data in prompts | `references/toon-format.md` |
| What to persist across sessions, memory file shape, staleness/invalidation | `references/memory-persistence.md` |
| Incorporating feedback durably without overfitting to one correction | `references/continuous-learning.md` |
| Self-verification before reporting done — compile/lint/test, diff re-reads | `references/verification-loops.md` |
| Identifying independent work, batching tool calls, avoiding shared-state races | `references/parallelization.md` |
| Decomposing tasks across subagents, grounding prompts, owning integration | `references/subagent-orchestration.md` |
| Packaging a pack for cross-agent discovery — depth-2 walk, portable frontmatter, transport caps | `references/skill-packaging.md` |
| Why loading less improves accuracy, not just cost — attention U-curve, the four degradation modes, where to put conclusions in a long reference | `references/context-engineering.md` |
| Optimising a frontend — the Inspect→Research→Brainstorm→Implement pass, the eight-section audit report, never reporting a measurement you did not take | `references/frontend-optimization-workflow.md` |
## Constraints
No claim of "done" without a run/compile/test step actually executed · no persisted memory fact without a staleness signal · no parallel dispatch across tasks that share a file or mutable state · no subagent prompt that omits facts already known to the dispatcher · no re-reading a file already fully in context without a stated reason.
Pliki w paczce (14)
- SKILL.md 5 KB
- examples/_stubs.d.ts 1 KB
- examples/good-agent-status-panel.test.tsx 2 KB
- examples/good-agent-status-panel.tsx 10 KB
- references/context-engineering.md 6 KB
- references/continuous-learning.md 6 KB
- references/frontend-optimization-workflow.md 7 KB
- references/memory-persistence.md 9 KB
- references/parallelization.md 7 KB
- references/skill-packaging.md 6 KB
- references/subagent-orchestration.md 10 KB
- references/token-optimization.md 8 KB
- references/toon-format.md 3 KB
- references/verification-loops.md 7 KB
Wersje
| Repozytorium | Wynik | Licencja |
|---|---|---|
| Ta strona · krishna-modi12/frontend-design-pro | 100 | MIT |
| weellio/gander | 100 | MIT |
Migawka bazy z 13.11.2026 · wersja paczki 093e0b0dc343 (skrót SHA-256 całej paczki)
Pobrania w ostatnich 30 dniach: 0 · ranking
Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.
Podobne skille
- 100agent-engineering
Zbiera sprawdzone zasady inżynierii dla agentów kodujących: workflow z planem na początku, delegowanie do subagentów, pętle samodoskonalenia, bramki weryfikacji i ustrukturyzowane zarządzanie zadaniami. - 90agent-operator
Pomaga wybrać sposób pracy równoległej w Claude Code: podagenci, zespoły agentów, worktree i wzajemne wiadomości między sesjami. - 100coordinator-dispatching-work
Wprowadza dyscyplinę koordynatora przy rozpoczynaniu epiku, zlecaniu pracy implementatorowi, tworzeniu worktree dla agenta i uruchamianiu subagentów. - 100coordinator-dispatching-work
Wprowadza dyscyplinę koordynatora przy rozpoczynaniu epiku, zlecaniu pracy implementatorowi, tworzeniu worktree dla agenta i uruchamianiu subagentów. - 100parallel-agent-work
Decyduje, czy podzielić pracę między podagentów, i bezpiecznie prowadzi podział z izolacją gałęzi. - 90subagent-orchestration
Opisuje wzorce rozdzielania pracy między równoległe subagenty w systemach wielu agentów: kiedy dzielić zadania i jak koordynować wyniki.