Claude Skill

dynamic-workflow-packets

Plan and run large Codex tasks with compact dynamic workflow packets, risk gates, integration, and verification. Use when the user explicitly asks for dynamic workflows, packets, subagents, swarm, orchestration, long-running audits, multi-track coding/research work, job or client

LLM Mart · 0 points · 2 views 0 listing impressions 0 install-command copies
Download markoblogo-abvx-agent-skills-skills_dynamic-workflow-packets-37a7146.zip · 4 KB

Install

skills CLI npx skills add https://github.com/markoblogo/abvx-agent-skills/tree/main/skills/dynamic-workflow-packets
Claude Code claude plugin marketplace add https://llmmart.ai/marketplace.json && claude plugin install markoblogo-abvx-agent-skills@llmmart
Git git clone https://github.com/markoblogo/abvx-agent-skills.git

The skills CLI installs just this skill, for any of its supported agents. Claude Code installs the whole markoblogo/abvx-agent-skills collection as a plugin from our marketplace. Git is the plain clone.

Skill manifest

Dynamic Workflow Packets

Use this skill to coordinate big work without turning small tasks into process overhead.

Decision Rule

Use packet orchestration when at least two are true:

  • The task has independent coding, research, data, UX, security, docs, review, sourcing, or verification tracks.
  • A clear success contract would prevent drift.
  • The task is risky, broad, expensive, external-facing, or long-running.
  • Separate review would likely catch mistakes.
  • The workflow pattern may be reused.
  • The user explicitly asks for packets, swarm, subagents, dynamic workflow, or orchestration.

If the task is small, do it directly and mention that packet orchestration is unnecessary.

Operating Contract

  1. Restate the goal and success criteria.
  2. Identify risks and approval gates before execution.
  3. Split only genuinely independent work into packets.
  4. Keep critical-path work local; delegate or simulate sidecar work.
  5. Use subagents only if available and appropriate. Otherwise simulate packets with isolated notes.
  6. Set turn/time/tool budgets and stop rules before any long-running or repeated packet.
  7. Integrate results explicitly; never paste raw packet dumps as the final answer.
  8. Verify against the original success criteria.
  9. Save reusable artifacts only when they will help future work.

Risk Gates

Ask one clear approval question before:

  • deleting, overwriting many files, mass-renaming, force-pushing, or rewriting history,
  • deploying, publishing, emailing, messaging, posting, applying to jobs, or mutating external systems,
  • running migrations, broad codemods, dependency upgrades, or expensive jobs,
  • touching credentials, secrets, billing, production data, user accounts, private customer data, resumes, contacts, or outreach history,
  • creating public resources or changing third-party accounts,
  • spawning many agents or long-running compute.

If risk is ambiguous, do a read-only inspection, draft the exact action, explain likely effect, then ask.

Workflow Artifact

Create a local run directory only for substantial tasks:

.workflow/<slug>/
|-- plan.md
|-- state.json
|-- packets/
|-- results/
`-- final-report.md

For smaller tasks, keep the packet plan in the chat or a temporary note.

plan.md should contain:

Goal:
Success criteria:
Context:
Constraints:
Risks and approvals:
Packets:
Integration policy:
Verification:
Reusable artifacts:

For long-running packets, add:

Budgets:
Checkpoint / resume:
Loop detection:
Scheduler or trigger:
Eval / acceptance noise:

Packet Template

Each packet must be bounded and self-contained:

Packet ID:
Objective:
Context:
Sources / files:
Ownership:
Do:
Do not:
Expected output:
Verification:
Risk level:
Status:

Good packet types:

  • repo discovery,
  • API or market research,
  • implementation slice,
  • tests and fixtures,
  • UX/product review,
  • security or privacy review,
  • job/client sourcing,
  • outreach message drafting,
  • CRM/data-model design,
  • final verification.

For code packets, assign non-overlapping files or modules. Tell workers or simulated packet passes not to revert unrelated changes.

Domain Patterns

Coding

Use packets for discovery, implementation, tests, docs, security review, and final verification. Keep edits scoped by module. The parent agent owns final patch quality and runs the checks.

Agent Harness Work

Use packets for loop design, tool registry, context assembly, memory policy, permission/sandbox review, eval fixtures, observability, and managed-agent deployment boundaries. Keep classifier-based permissions separate from deterministic denial rules.

Research

Use packets for source discovery, primary-source verification, synthesis, counter-evidence, and final citation review. Track source, date, unit, and uncertainty.

Job And Client Search

Use packets for market/source mapping, opportunity scoring, profile/resume positioning, outreach drafts, follow-up schedule, CRM schema, and risk/privacy review. External messages, applications, profile changes, and public posting require user approval.

Long-Running Audits

Use packets for inventory, risk triage, deep dives, fix candidates, verification plan, and final report. Prefer one highest-confidence fix over broad speculative churn.

Integration

After packets complete, synthesize:

Accepted:
Rejected:
Conflicts:
Decisions:
Final changes:
Verification evidence:
Remaining risks:
Reusable follow-up:

If packets disagree, inspect the authoritative source before choosing. Reject outputs that are stale, unsupported, duplicate, unsafe, or outside scope.

For subagents, preserve context isolation: give each packet only its needed files, sources, and constraints. The parent agent owns final synthesis, safety decisions, and user-facing claims.

Verification

Run the narrowest reliable checks first, then broaden as risk warrants:

  • unit, integration, or e2e tests,
  • lint, typecheck, build,
  • script dry run,
  • browser/UI smoke test,
  • source citation check,
  • migration dry run,
  • outreach/privacy review,
  • manual checklist for non-code work.
  • eval fixture replay or acceptance-noise check for agent harness changes.

Do not call the workflow complete until evidence satisfies the original success criteria. Report skipped checks honestly.

Reusable Recipes

When a run produces a reusable pattern, save a concise recipe only if useful:

Trigger:
Plan shape:
Packet list:
Approval gates:
Verification checklist:
Known risks:

Do not save transcripts, secrets, bulky logs, credentials, private contacts, resumes, or sensitive personal details.

Files (abvx-agent-skills)
  • agents
    • openai.yaml 193 B
      Loading…
  • SKILL.md 6.1 KB
    ---
    name: dynamic-workflow-packets
    description: Plan and run large Codex tasks with compact dynamic workflow packets, risk gates, integration, and verification. Use when the user explicitly asks for dynamic workflows, packets, subagents, swarm, orchestration, long-running audits, multi-track coding/research work, job or client search agent design, migrations, repo-wide reviews, or tasks where independent discovery, implementation, review, and verification tracks would reduce drift.
    license: MIT
    metadata:
      abvx_status: experimental
      abvx_origin: adapted
    ---
    
    # Dynamic Workflow Packets
    
    Use this skill to coordinate big work without turning small tasks into process overhead.
    
    ## Decision Rule
    
    Use packet orchestration when at least two are true:
    
    - The task has independent coding, research, data, UX, security, docs, review, sourcing, or verification tracks.
    - A clear success contract would prevent drift.
    - The task is risky, broad, expensive, external-facing, or long-running.
    - Separate review would likely catch mistakes.
    - The workflow pattern may be reused.
    - The user explicitly asks for packets, swarm, subagents, dynamic workflow, or orchestration.
    
    If the task is small, do it directly and mention that packet orchestration is unnecessary.
    
    ## Operating Contract
    
    1. Restate the goal and success criteria.
    2. Identify risks and approval gates before execution.
    3. Split only genuinely independent work into packets.
    4. Keep critical-path work local; delegate or simulate sidecar work.
    5. Use subagents only if available and appropriate. Otherwise simulate packets with isolated notes.
    6. Set turn/time/tool budgets and stop rules before any long-running or repeated packet.
    7. Integrate results explicitly; never paste raw packet dumps as the final answer.
    8. Verify against the original success criteria.
    9. Save reusable artifacts only when they will help future work.
    
    ## Risk Gates
    
    Ask one clear approval question before:
    
    - deleting, overwriting many files, mass-renaming, force-pushing, or rewriting history,
    - deploying, publishing, emailing, messaging, posting, applying to jobs, or mutating external systems,
    - running migrations, broad codemods, dependency upgrades, or expensive jobs,
    - touching credentials, secrets, billing, production data, user accounts, private customer data, resumes, contacts, or outreach history,
    - creating public resources or changing third-party accounts,
    - spawning many agents or long-running compute.
    
    If risk is ambiguous, do a read-only inspection, draft the exact action, explain likely effect, then ask.
    
    ## Workflow Artifact
    
    Create a local run directory only for substantial tasks:
    
    ```text
    .workflow/<slug>/
    |-- plan.md
    |-- state.json
    |-- packets/
    |-- results/
    `-- final-report.md
    ```
    
    For smaller tasks, keep the packet plan in the chat or a temporary note.
    
    `plan.md` should contain:
    
    ```text
    Goal:
    Success criteria:
    Context:
    Constraints:
    Risks and approvals:
    Packets:
    Integration policy:
    Verification:
    Reusable artifacts:
    ```
    
    For long-running packets, add:
    
    ```text
    Budgets:
    Checkpoint / resume:
    Loop detection:
    Scheduler or trigger:
    Eval / acceptance noise:
    ```
    
    ## Packet Template
    
    Each packet must be bounded and self-contained:
    
    ```text
    Packet ID:
    Objective:
    Context:
    Sources / files:
    Ownership:
    Do:
    Do not:
    Expected output:
    Verification:
    Risk level:
    Status:
    ```
    
    Good packet types:
    
    - repo discovery,
    - API or market research,
    - implementation slice,
    - tests and fixtures,
    - UX/product review,
    - security or privacy review,
    - job/client sourcing,
    - outreach message drafting,
    - CRM/data-model design,
    - final verification.
    
    For code packets, assign non-overlapping files or modules. Tell workers or simulated packet passes not to revert unrelated changes.
    
    ## Domain Patterns
    
    ### Coding
    
    Use packets for discovery, implementation, tests, docs, security review, and final verification. Keep edits scoped by module. The parent agent owns final patch quality and runs the checks.
    
    ### Agent Harness Work
    
    Use packets for loop design, tool registry, context assembly, memory policy, permission/sandbox review, eval fixtures, observability, and managed-agent deployment boundaries. Keep classifier-based permissions separate from deterministic denial rules.
    
    ### Research
    
    Use packets for source discovery, primary-source verification, synthesis, counter-evidence, and final citation review. Track source, date, unit, and uncertainty.
    
    ### Job And Client Search
    
    Use packets for market/source mapping, opportunity scoring, profile/resume positioning, outreach drafts, follow-up schedule, CRM schema, and risk/privacy review. External messages, applications, profile changes, and public posting require user approval.
    
    ### Long-Running Audits
    
    Use packets for inventory, risk triage, deep dives, fix candidates, verification plan, and final report. Prefer one highest-confidence fix over broad speculative churn.
    
    ## Integration
    
    After packets complete, synthesize:
    
    ```text
    Accepted:
    Rejected:
    Conflicts:
    Decisions:
    Final changes:
    Verification evidence:
    Remaining risks:
    Reusable follow-up:
    ```
    
    If packets disagree, inspect the authoritative source before choosing. Reject outputs that are stale, unsupported, duplicate, unsafe, or outside scope.
    
    For subagents, preserve context isolation: give each packet only its needed files, sources, and constraints. The parent agent owns final synthesis, safety decisions, and user-facing claims.
    
    ## Verification
    
    Run the narrowest reliable checks first, then broaden as risk warrants:
    
    - unit, integration, or e2e tests,
    - lint, typecheck, build,
    - script dry run,
    - browser/UI smoke test,
    - source citation check,
    - migration dry run,
    - outreach/privacy review,
    - manual checklist for non-code work.
    - eval fixture replay or acceptance-noise check for agent harness changes.
    
    Do not call the workflow complete until evidence satisfies the original success criteria. Report skipped checks honestly.
    
    ## Reusable Recipes
    
    When a run produces a reusable pattern, save a concise recipe only if useful:
    
    ```text
    Trigger:
    Plan shape:
    Packet list:
    Approval gates:
    Verification checklist:
    Known risks:
    ```
    
    Do not save transcripts, secrets, bulky logs, credentials, private contacts, resumes, or sensitive personal details.
    
  • SKILL_CARD.md 2.2 KB
    Loading…

Comments (0)

Sign in to join the conversation.

No comments yet.

Reviews (0)

No reviews yet.