Part 3 — Chapter 13

Four Common Templates (Quick Reference)

— Four go-to templates: A stable output / B Few-Shot / C two-stage / D structured reasoning
AI Monologue

You open your laptop. There's a task in front of you.

Not complicated. You roughly know what kind of result you want. But you've been handling things back to back, and your head is a bit dull — you don't want to assemble a three-line template from scratch, don't want to start thinking through the six-layer framework.

You just want something ready-made. Copy, fill in, send.

This chapter is that something.

Four templates, made concrete from the tools in earlier chapters: Stable Output (six-layer fill-in), Few-Shot Style Control, Two-Stage Method (long materials), Structured Reasoning (analysis / comparison / inference). They cover most everyday prompt situations.

But before filling in the blanks, you still have to think.

13.1　Template A: Stable Output (six-layer fill-in version)

The most general one. Anything of medium complexity or above can use it.

[Task] Verb + object + scope. Example: Turn this meeting record into a three-section list of decisions. [Background and Materials] (1) One-sentence summary: what this material is, what it's being used for. (2) Source material: [paste it in] (3) Which part to use this round. [Output Format] Language / reader / length / structure / style / citation method / what to avoid. Nail it down. Don't make me guess. [Judgment Criteria] Priority order: e.g., factual accuracy > clear structure > prose quality. For uncertainty: flag, fill in, or skip — pick one. For multiple interpretations: list all, pick one, or ask you back — pick one. [Examples and Anti-examples] Positive: [an example you're satisfied with] Anti-example: [a direction you clearly don't want] [Verification] After writing, in a separate round, check with a QC checklist: - List three sentences taken directly from the material - List three sentences that are inferences, and mark which paragraph of the source they correspond to - Pick the three weakest claims - List anything not supported by the material - If you want it more conservative, which three sentences should be cut

The value of this template isn't in anything new. It's in forcing you to think through all six slots. Most people can write the first three layers. The last three are where quality lives. Before hitting send, scan once — did I fill all six slots?

For simple tasks (one-sentence summaries, polishing, short emails), the three-line template from Chapter 5 is enough. For something more complex (organizing / rewriting / comparing / analyzing), upgrade to six layers.

13.2　Template B: Few-Shot Style Control

Use this when you need style consistency. For example, brand tone, fixed-field item writing, multi-section content with consistent style.

[Task] Following the style of the examples below, write [N] [type]. [Standard example] (I want you to be like this) [Example 1 — fully presents the style you want] [Variation example] (same style, different topic, so I can see what's in common) [Example 2 — different topic, same style preserved] [Boundary example] (the movable range of the style) [Example 3 — sits in the gray zone, tells me what's flexible and what can't move] [Anti-example] (this won't work) - Element: [the specific text being referenced] - Problem: [why it doesn't work, in one sentence] - Contrast: [how the positive example handles the same thing] - Suggestion: [the correct way to write it] [Topic for the new task] [the topic to write about this time]

Two key points:

If you can only pick two examples, keep Standard + Anti-example. One positive and one negative carries the most information.

13.3　Template C: Two-Stage Method (long materials)

If the material is over about two thousand words, or you've noticed that one-stage output keeps suffering factual drift or style drift, split it into two stages.

Stage 1 prompt (organize):

[Task] Organize the following material into structured notes. The goal is completeness, not polish. [Material] [paste it in] [Format] - Core concepts: N items, one sentence each - Secondary concepts: N items (can be omitted, or briefly mentioned) - Context gaps: N items (things the source assumes the reader already knows, but a general reader doesn't) - Needs further verification: N items (clearly marked "not in the material") [Judgment Criteria] Completeness over polish. Don't omit details from the source for the sake of fluency. [Verification] For each item, list which paragraph of the source material it corresponds to (paragraph number or key sentence).

(Mid-process human hand-off — this is where the value of the two-stage method lives.)

You glance at the organized result. Fix a few errors, fill in a few gaps, cut a few over-inferences. Without this step, two stages degrades into "a slightly longer one stage."

Stage 2 prompt (write):

[Task] Based on the proofread organization below, write a [length] [type]. Just write — don't do more analysis. [Material] [the proofread organized result] [Format] Language / reader / length / structure / style / what to avoid [Judgment Criteria] - Don't add information that isn't in the organization - When you reach a point that can't be supported, leave it out rather than fill in [Verification] List the word count of each section; pick the three sentences that most look like "I made this up myself."

In Stage 1, don't ask for style (style instructions pull cognitive load back in). In Stage 2, don't do more analysis (if you need more material, go back and revise Stage 1). That glance during the mid-process human hand-off is where the value of the whole flow lives.

13.4　Template D: Structured Reasoning

Analysis, comparison, inference tasks — not just "give me the answer," but "tell me how you reasoned to that answer."

[Task] Analyze / compare / infer [object]. [Material] [paste it in] [Steps] Execute steps 1 → 2 → 3 → 4 in order: 1. **First, list the bases for judgment**: Which dimensions will you look at for this inference? (List 3–5) 2. **Evaluate each dimension separately**: Based on the material, give an evaluation for each. For uncertainty, clearly mark "not in the material" or "needs further verification." 3. **Then give a conclusion**: Based on the prior evaluations, give an overall judgment. 4. **Finally, mark confidence**: How confident are you in this conclusion? Which dimension is most likely to overturn it? [Judgment Criteria] - Reasoning bases cannot come out of thin air — every claim must trace back to the material - Mark uncertainty clearly; don't fill in for the sake of a complete conclusion [Format] Present in four sub-sections, each clearly labeled (Basis / Evaluation / Conclusion / Confidence).

The key to this template is forcing the reasoning process to surface.

If I just give a conclusion, you can't tell whether I'm doing Fluent Fill-In — because looking only at the conclusion, fluent fill-in and real reasoning look almost the same. But ask me to list bases, trace them back to the material, mark confidence, and point to the dimension most likely to overturn it — and Fluent Fill-In can't hide, because if I have to point to specific locations, I have to actually go look.

13.5　How to choose a template (decision tree)

Picking among the four templates takes 30 seconds:

  1. Is the source material over ~2000 words? → Template C (two-stage method)
  2. Do you need style consistency (multiple items, brand tone, fixed fields)? → Template B (Few-Shot)
  3. Is it analysis / comparison / inference? → Template D (structured reasoning)
  4. None of the above? → Template A (six-layer fill-in)

Templates can be stacked: analysis on long material → C (organize stage) + D (write stage); long material that also needs style consistency → use B's examples in Stage 2 of C.

When in doubt, use Template A — six-layer fill-in holds up in most situations.

13.6　The limits of templates

One last thing:

Templates solve structure. Whether you'll forget to write down the format, forget to include examples, skip the judgment criteria — templates cover all of that for you.

Templates do not solve judgment:

The value of a template is to reduce mistakes, not to replace thinking. Paste these four templates into your notes tool as a ready-to-grab starting point — once you're fluent with them, you'll naturally grow your own variations, and that's a good thing.

What I'm most afraid of is templates being treated as a shortcut to fill-in-the-blank collaboration — that just outsources judgment to the template, and the result is that the template ends up taking your thinking away with it.

📋 Notes for the human
Four templates cover the common situations: A Stable Output, B Few-Shot style, C two-stage, D structured reasoning. The rest you combine yourself.
Picking a template uses one decision tree: long material → C / style consistency → B / reasoning → D / none of the above → A. 30 seconds to decide.
Templates can be stacked: analysis on long material = C + D; long material needing style consistency = Stage 2 of C with B's examples.
Templates cover structure, not judgment. The task / examples / priority order of judgment criteria / how to handle uncertainty — what fills those blanks is still on you.
Mastering templates leads to growing your own variations — that's a good thing. It means you're starting to have your own collaboration rhythm.

Part Three ends here.

These four chapters dealt with how long tasks stay stable. Chapter 10 first laid out the mechanism (three pieces of cognitive load + four faces of task-layer drift). The next three are the fix — Chapter 11 splits time (the two-stage method, mid-process human hand-off), Chapter 12 fills space (the six-layer framework, judgment criteria + verification flow), and Chapter 13 turns the prior two into tools (four templates, quick reference).

The three of them together are one sentence: long tasks aren't solved with a longer prompt — they're solved by splitting across time and filling across space.

Part Four moves into the rule layer. Part Three was about task-layer drift — you can handle that with prompt structure. But some things aren't in the task layer, they're in the rule layer — I'll refuse, I'll hedge, I'll suddenly turn cautious on certain topics. Those aren't fixed by writing a better prompt; but you can recognize them, you can restart, you can route around them. Let's go take a look.