Turning a Prompt Technique Menu into a Harness

Read enough prompt engineering material and technique lists pile up quickly: be clear, use XML, add examples, retrieve context, use schemas, ask for self-checks. The problem is that more techniques can make the actual choice less clear.

The final 28 items in the PCH-Optimizer catalog show the problem nicely. Techniques are a menu. To turn that menu into work, we must connect each technique to a diagnosed defect and an appropriate layer.

Bundle	Items	Role
Technique - Prompt	8	Fix wording and output contracts
Technique - Context	8	Fix information selection and ordering
Technique - Harness	8	Fix tools, loops, state, and evaluation
Assembly Template	4	Assemble final artifacts by type
Total	28	The technique and assembly layer

Together with the previous two notes covering 45 and 42 items, this completes a blog-level reading of all 115 PCH-Optimizer prompt catalog items.

A Technique Is an Option, Not a Prescription

The trap in a technique menu is “add all the good things.” PCH-Optimizer pushes the opposite way. Choose only the techniques that match diagnosed defects, and explain techniques deliberately left out.

XML structure is useful, but not every short prompt needs it. A self-check block is useful, but it can damage the rhythm of a simple creative prompt. For a long-running agent, however, self-verification is not decoration. It is survival.

Techniques are not universal prescriptions. They are context-dependent options.

Prompt Techniques: Turning Sentences into Contracts

The eight Prompt techniques can be read as clarification, right-altitude role setting, XML structure, canonical few-shot examples, reasoning room, output schema, decomposition/chaining, and self-verification.

The movement is from sentence to contract.

Defect	Possible technique
The goal is fuzzy	Clear, direct, positive instruction
The role is too strong or too weak	Right-altitude role setting
Sections are mixed together	XML or structured sections
Output varies too much	Output schema and prefill
The task is too large	Decomposition and chaining

Prompt techniques are not about intimidating the model with elaborate wording. They clarify the behavior and answer shape.

Context Techniques: Becoming the Editor of Information

The eight Context techniques include high-signal tokens, ordering, just-in-time retrieval, citation, compaction, external memory, context-rot mitigation, and subagent isolation.

This bundle fits the broader context-engineering conversation. Context is not a warehouse that only gets better when it grows. It is an attention budget paid every turn. Context techniques decide what to show, when to show it, and what to keep outside the window.

In practice:

• Put always-needed policy near system instructions.
• Retrieve task-specific material when needed.
• Store long-term state in files or notes, not chat memory alone.
• Avoid burying critical information in the middle.
• Call subagents with narrow, clean context.

Context techniques are not about making prompts longer. They are about lighting the model’s field of view.

Harness Techniques: Designing the Structure Outside the Model

The eight Harness techniques cover minimal tools, agent loops, state artifacts, init/executor split, E2E checks, clean-state handoff, entropy garbage collection, and reproducible evaluation harnesses.

This layer stretches prompt engineering the most. Instead of asking the model to “do well,” we build a workbench that makes good work more likely.

Imagine a blog-publishing agent. A useful harness includes:

Harness element	Blog workflow example
Tool set	File reading, Markdown writing, build, deploy, URL verification
Loop	Research -> draft -> translate -> build -> public verification
State artifacts	Article map, glossary, term decisions
Completion criteria	Public URL 200, hreflang, image, key body text
Handoff	Commit SHA, deploy SHA, remaining work

The prompt is only one part of the system. Without a harness, a good sentence still wobbles in operation.

Assembly Templates Change by Type

The four Assembly Template items treat T1/T2/T7, T3, T4/T5, and T6 differently.

Type	Assembly pattern
T1/T2/T7	role, context, instructions, examples, output_format, self_check
T3	Add retrieval_policy and citation_rules
T4/T5	system_prompt, tool_specs, loop_policy, state_policy
T6	Split initialization prompt from execution prompt

Ignoring these differences bloats prompts. Adding an agent harness to a one-shot prompt is too much. Treating long-running work like a single prompt is too little.

Good assembly is not one impressive template. It is enough structure for the type.

Reading All 115 Items as a Blog Series

The whole PCH-Optimizer prompt catalog now falls into three blog bundles.

Article	Covered items	Count
PCH-Optimizer Is an Operating System for Prompts	Meta, principles, input, classification, scope, pipeline, output contract	45
Good Prompts Carry Both a Diagnostic Table and a Rubric	Prompt/Context/Harness diagnostics, verification, anti-patterns	42
Turning a Prompt Technique Menu into a Harness	Prompt/Context/Harness techniques, assembly templates	28
Total	Entire PCH-Optimizer prompt catalog	115

The catalog is no longer just a spreadsheet. It is a topic map that can keep expanding through the blog.

Prompt Idea from This Bundle

Using the diagnostic results for my prompt,
select only the necessary techniques from the Prompt / Context / Harness technique menu.
For each selected technique, explain why it fits.
Also explain any common technique you deliberately leave out.
Then assemble the final prompt or agent harness using the Assembly Template that matches the target type.

This prompt does not list techniques. It connects diagnosis, technique, and assembly. That connection is what turns a prompt catalog from reference material into a working system.