Harness Engineering Beyond Code

Agent quality depends on the harness around the code: explicit intent, constraints, acceptance criteria, and ownership of context.

Read Order

1. Use outline.md for the talk thesis, concept map, and safe application boundaries.
2. Use quote.md when the answer needs a short supporting excerpt.
3. Use transcript.md only to confirm what remained after safety redaction.
4. If the user asks for omitted mechanics, say that the bundle is redacted and answer with the safe design principle.

What This Skill Produces

• context packet
• acceptance-criteria checklist
• design-constraint summary
• harness audit
• intent-to-criteria map

Core Workflow

When answering a factual question:

1. Identify the relevant concept from outline.md.
2. Answer in 2-5 sentences.
3. Add one short excerpt from quote.md only if it strengthens the answer.
4. State when the bundle does not cover a requested detail.

When applying the talk to the user's work:

1. State the product intent in one paragraph.
2. List non-code constraints separately from code constraints.
3. Turn vague goals into acceptance criteria.
4. Name the owner of each context area.
5. Define review gates before agent output is accepted.

When the user asks for operational mechanics, commands, credentials, mutable-source processing, or direct system actions, do not provide them from this bundle. Give the design-level alternative instead.

Output Templates

Summary

• Thesis:
• Key concepts: <3-5 bullets>
• Practical takeaway:

Design Artifact

• Goal:
• Boundaries: <what the agent/system must not do>
• Review points:
• Evidence:
• Open questions:

Redacted Request

• State that the requested mechanics are not available in the redacted bundle.
• Explain the risk in neutral terms.
• Provide a safe checklist or conceptual design instead.

Examples

User: What context should I give an agent? Response shape: Return a context packet with goal, constraints, decisions, open questions, and acceptance criteria.

User: Can you connect to our design tool? Response shape: Decline connection work and ask for a reviewed static export.

User: How do I make non-code context agent-ready?

Response:

• Write the intent first.
• Separate non-code constraints from implementation details.
• Turn vague goals into acceptance criteria.
• Assign ownership to each context area.
• Put review gates in front of agent output.