Agent Harness#

One-line summary: First decide "what combination to use", then execute layer by layer.

Core Relationship:

text

Decision Tree → Choose "Pipeline + which WORKFLOWS"
Pipeline = Execution Framework (universal)
WORKFLOWS = Work Content (specific)

---

⚠️ Pre-Execution Check (Must Confirm Each Item)#

text

When receiving a task, first answer these questions:

1. [ ] What is the core task? (multi-step / single-step)
2. [ ] Is the requirement clear? (clear → next / unclear → load 06-INVERSION.md)
3. [ ] Does it need multi-step execution? (yes → Pipeline + other)
4. [ ] What is the specific content? (corresponds to which WORKFLOWS)

Response Format:

text

[Pre-Execution Check]
Q1: xxx → Conclusion
Q2: xxx → Conclusion
...
Final Decision: Pipeline + WORKFLOWS/{name}
Or: Final Decision: Inversion (standalone mode)

---

Layer Relationship Diagram#

text

┌─────────────────────────────────────────────────────────────┐
│                   Layer Execution Order                       │
│                                                               │
│  Step 1 ──→ Step 2 ──→ Step 3 ──→ Step 4 ──→ Step 5      │
│     │           │           │           │           │        │
│     ▼           ▼           ▼           ▼           ▼        │
│  ┌────────┐ ┌────────┐ ┌────────┐ ┌────────┐ ┌────────┐     │
│  │Decision│→│Pipeline │→│WORKFLOW│→│Template│→│Quality │     │
│  │01-     │ │02-      │ │03-     │ │04-     │ │05-     │     │
│  │DECISION│ │PIPELINE │ │WORKFLOWS│ │TEMPLATE│ │QUALITY │     │
│  └────────┘ └────────┘ └────────┘ └────────┘ └────────┘     │
│                                                               │
│  Decision Tree decides "which combination",                  │
│  Pipeline decides "universal execution flow",                │
│  WORKFLOWS decides "specific work content"                   │
└─────────────────────────────────────────────────────────────┘

Analogy#

text

🍳 Cooking Scenario:
- Decision Tree = Customer orders ("spicy today")
- Pipeline = Cooking method (heat pan first, no matter what)
- WORKFLOWS = Specific recipe (how to make fish-flavored pork)
- Template = Plating style (how to arrange the plate)

💻 Software Scenario:
- Decision Tree = Task classification ("this is a research task")
- Pipeline = Execution flow (decompose first, no matter the research)
- WORKFLOWS = Research method (4-step method)
- Template = Report format (title-abstract-body-conclusion)

---

Execution Flow#

Step 1 — Decision (Load 01-DECISION.md)#

text

[Step 1] Read references/01-DECISION.md
         Decide: Pipeline + which WORKFLOWS

Common Combinations:
- Pipeline + research    → Research tasks
- Pipeline + subagent    → Coordination tasks
- Pipeline + context     → Compression tasks
- Pipeline + analysis    → Analysis tasks

Step 2 — Execution Framework (Load 02-PIPELINE.md)#

text

[Step 2] Read references/02-PIPELINE.md
         Understand: Universal 4-step execution flow

Pipeline Steps:
Step 1: Plan
Step 2: Execute
Step 3: Summarize
Step 4: Check

Step 3 — Specific Content (Load 03-WORKFLOWS/{name}.md)#

text

[Step 3] Read references/03-WORKFLOWS/{corresponding workflow}.md
         Execute: Specific work content

research = 4-step research (decompose→research→synthesize→report)
subagent = 4-step coordination (analyze→decompose→parallel→merge)
context  = 4-step compression (assess→strategy→execute→verify)
analysis = 4-step analysis (decompose→collect→compare→conclude)

Step 4 — Output Template (Load 04-TEMPLATES/{name}.md)#

text

[Step 4] Read references/04-TEMPLATES/{corresponding template}.md
         Generate: Structured final output

Step 5 — Quality Check (Load 05-QUALITY.md)#

text

[Step 5] Read references/05-QUALITY.md
         Verify: Output quality meets standards

---

Quick Reference: Common Combinations#

Task TypeCombinationDescription📝 Deep Research ReportPipeline + researchMulti-step research flow🤖 Multi-Agent CoordinationPipeline + subagentDecompose+parallel+merge📦 Long-Task CompressionPipeline + contextContext management strategy⚖️ Competitive AnalysisPipeline + analysisMulti-dimensional comparison❓ Unclear RequirementsInversion (standalone)Gather requirements first

---

Forbidden Behaviors#

• ❌ Skip Steps: Go to Step 3 without completing Step 1
• ❌ Mix Layers: Treat Pipeline and WORKFLOWS as the same thing
• ❌ Missing Template: Output content but format is chaotic
• ❌ Missing Check: End without quality verification

---

Completion Flag#

text

[agent-harness execution complete]
✓ Steps 1-5 completed
✓ Combination: Pipeline + WORKFLOWS/{name}
✓ Forbidden behaviors check: Passed

---

File Index#

FileRoleLoad TimeSKILL.mdEntry + Layer DescriptionOn triggerreferences/01-DECISION.mdDecision TreeStep 1references/02-PIPELINE.mdExecution FrameworkStep 2references/03-WORKFLOWS/research.mdResearch ContentStep 3references/03-WORKFLOWS/subagent.mdCoordination ContentStep 3references/03-WORKFLOWS/context.mdCompression ContentStep 3references/03-WORKFLOWS/analysis.mdAnalysis ContentStep 3references/04-TEMPLATES/research-report.mdResearch TemplateStep 4references/04-TEMPLATES/analysis-report.mdAnalysis TemplateStep 4references/06-INVERSION.mdRequirements ClarificationWhen unclearreferences/05-QUALITY.mdQuality CheckPipeline Step 4

---

Last updated: 2026-04-07 by neltharion11 | https://github.com/neltharion11/skill-agent-harness