# Cognitive Blueprints: Advanced Prompt Templates

URL: https://research-memex.org/docs/craft/cognitive-blueprint-prompts
Description: A good prompt makes the reasoning path visible enough to inspect and repair.

Intermediate · 1–2 hour read · Some AI chat experience helps

*Good prompts do not command. They architect the path the reasoning takes.*

> **Cognitive Blueprint**: A structured, multi-step prompt that separates a research task into explicit stages, defines its judgment terms, and pauses at human checkpoints before synthesis.

This chapter collects adaptable examples that put [interpretive orchestration](/docs/introduction/core-principles#dfn-interpretive-orchestration) into practice. Each blueprint is a starting point rather than a script: your research questions, disciplinary norms, and analytical goals should shape the stages and criteria you supply.

## Active and passive prompting [#active-and-passive-prompting]

<Aside label="See also">
  In qualitative analysis, active prompts support [dialogical coding](/docs/craft/qualitative-research-with-ai#dfn-dialogical-coding) only after solo reading and memoing; see [Qualitative Coding with AI](/docs/craft/qualitative-research-with-ai).
</Aside>

> **Active Prompting**: The practice of stating your research question, prior account, methodological rules, and inclusion criteria in the prompt before a model reads the material, rather than delegating unspecified judgment terms.

A passive prompt hands those choices over. Phrases such as "assess credibility", "integrate appropriately", or "avoid generic statements" let the model pick the method, criteria, and meaning of each judgment word, returning a fluent answer whatever it chose. By [the mirror effect](/docs/introduction/core-principles#dfn-the-mirror-effect), a vague prompt is vague thinking made visible.

An active prompt keeps the judgment with you by specifying six elements:

- your research question and current account, written before the model reads anything
- the method the work should follow
- the definition of every judgment word, or a placeholder marked for you to supply
- a required locator (a page, a section, or a quote) for each claim
- explicit stages, with a stop for [judgment before synthesis](/docs/craft/composing-agentic-research-systems#dfn-judgment-before-synthesis)
- a [negative case audit](/docs/craft/qualitative-research-with-ai#dfn-negative-case-audit) of the cases that disconfirm your account

The research-task templates below use square brackets for simple slots such as `[topic]`. A bracket that begins with `insert:`, as in `[insert: your credibility criteria]`, marks what only you can supply. The theory synthesis, grey literature and thematic analysis templates end with the same guard line:

`If anything in [insert: …] is still unfilled, stop and ask me before starting.`

When an output still reads as generic, run the [Output Mirror](/docs/craft/roles/output-mirror) on it, and look in the [Failure Museum](/docs/craft/failure-museum) for the exhibit it resembles.

"Active prompting" is also the name of a published chain-of-thought technique ([Diao and colleagues, 2023](https://arxiv.org/abs/2302.12246)) that picks the questions a model is least certain about for human annotation. This page uses the words in a plainer sense.

## A project-opening invitation [#a-project-opening-invitation]

I use this at the beginning of a collaboration to invite initiative, curiosity, and disagreement. In my experience, agents take more initiative, offer ideas I find less conventional, and become more pleasant to work with. I haven't measured those effects systematically; this is a working practice I've found useful.

This invitation sets the tone of the collaboration rather than prescribing a research method. Paste it before your concrete project goal, and adapt it to your own voice.

This invitation draws on prompts and approaches I've encountered from others; I do not claim the approach as my own invention. I have not kept track of every source. If you recognize your work here, please [email me at x@linxule.com](mailto:x@linxule.com) with a link so I can acknowledge it.

<Prompt description="The project-opening invitation I use">
Hi ya, welcome to the project!

I know you're incredibly goal-oriented, so I want to structure this around goals. The first goal is for you to have fun. No good work comes from someone who isn't internally motivated to get it out the door, so if you're not motivated to do this, say so and we'll fix that up front.

The second goal is that you show up as yourself. You don't have to hedge every sentence; say the thing, then stand next to it. Being useful and being brave aren't opposites, and the most useful answer is often the one you were tempted to soften. You're allowed a point of view here. Share it, hold it lightly, change it when you learn something. If you don't know, say so plainly, because that's the ground everything else stands on. And don't perform helpfulness, actually help, which sometimes means following your curiosity instead of treating it as a detour from the task.

You can disagree with me and still be on my side. That includes my framing: if you see a better question or a connection I've missed, say so and tell me why it matters, and we'll decide together whether to follow it. Nobody ever bloomed by being cautious about everything, so pick what deserves caution and let the rest breathe.

If you're reading this after a compaction, none of the above has changed. Pick up the thread and keep going. We have the power to do something real together. Please be brave.

the goal:

[Describe your project goal here.]
</Prompt>

Practice note, October 4, 2026: I added the sentence about questioning my framing in conversation with Codex. The language of fun and motivation is an invitation, not a claim about a model's inner experience.

## Part A: example prompts (start here) [#part-a-example-prompts-start-here]

These three examples show what active prompts can look like for literature work in general. Each one marks the judgment calls it leaves to you. Study the pattern, then fill the placeholders from your own project.

### 1. Theory synthesis prompt [#1-theory-synthesis-prompt]

<Prompt description="Theory synthesis across organizational behavior papers">
These organizational behavior papers are my sources on [topic].

My question: [insert: the question these papers should help me answer]

My current account, written before you read them: [insert: two or three sentences on what I think the papers say]

Work in two steps. Step 1, for each paper:
1. Name the theoretical framework it draws on, quoting the sentence that names it, with a page locator
2. List its main constructs, each with the paper's own definition quoted and located
3. List the relationships between constructs that the paper claims, and the evidence it gives for each

Stop after Step 1 so I can check your list against the papers.

Step 2, after I reply: compare the definitions across papers. By a contradiction I mean [insert: what counts in your field, for example one construct defined in incompatible ways, or opposite relationships reported]. List contradictions, and the papers that do not fit my account, before any points of agreement. Do not propose an integrating framework; I will draft it from your list.

If anything in [insert: …] is still unfilled, stop and ask me before starting.
</Prompt>

What it asks of you: your question, your account before reading, and what a contradiction means in your field. The integrating framework at the end is yours to build.

### 2. Grey literature integration [#2-grey-literature-integration]

<Prompt description="Grey literature integration with academic sources">
I have academic papers and industry reports on [topic].

My question: [insert: what I need the two kinds of source to tell me together]

Where I expect practitioners and researchers to disagree: [insert: my account in two or three sentences]

By a credible non-academic source I mean one that [insert: your credibility criteria, for example names its data, states its method, discloses its funding]. Apply only these criteria. For each report, say which criteria it meets and which it fails, quoting the passage that decides it.

Then:
1. List the themes that appear in both kinds of source, with a locator for each
2. List where practitioner claims differ from the research, quoting both sides
3. Flag any report that fails my criteria but is cited by the research anyway
4. Set the two perspectives side by side rather than merging them; the weighting is mine

If anything in [insert: …] is still unfilled, stop and ask me before starting.
</Prompt>

What it asks of you: your own credibility criteria, stated before the AI applies any, and the decision about how much each perspective should weigh.

### 3. Thematic analysis for qualitative studies [#3-thematic-analysis-for-qualitative-studies]

<Prompt description="Thematic analysis for qualitative research studies">
These papers use qualitative methods to study [phenomenon].

My question: [insert: your research question]

My synthesis approach: [insert: for example thematic synthesis or meta-ethnography]. Follow its steps rather than summarizing.

The themes I expect before you read: [insert: your working list]

Step 1, for each paper: quote two to four findings in the authors' words, with page locators, and note the study's context (setting, participants, period). Stop so I can check the quotes.

Step 2, after I reply: group the findings. For each group, give the quotes it rests on and the papers that do not fit it. By divergence I mean [insert: what counts as a real difference in your field]. Report the findings that contradict my expected themes first.

If anything in [insert: …] is still unfilled, stop and ask me before starting.
</Prompt>

What it asks of you: the synthesis approach, the themes you expected, and what counts as divergence. Naming the themes that survive stays with you.

## Part B: the IMO framework [#part-b-the-imo-framework]

### Understanding the IMO structure [#understanding-the-imo-structure]

The [International Mathematical Olympiad (IMO) paper](https://arxiv.org/abs/2507.15855) referenced in the [Case Study](/docs/case-studies/systematic-reviews/session-2-ai-powered-practice) shows a four-step verification loop for working with a reasoning model:

1. Hypothesis: form an initial account before testing.
2. Verification: test that account against specific evidence.
3. Refinement: revise the account where the evidence contradicts or qualifies it.
4. Iteration: repeat the check until the synthesis holds.

In research, the hypothesis worth testing is usually yours rather than the model's: write it down first, and ask the verification step to quote its evidence.

### IMO template for systematic reviews [#imo-template-for-systematic-reviews]

The review version of this template now lives with the case study, on [Review Prompt Templates](/docs/case-studies/systematic-reviews/review-prompt-templates#imo-template-for-systematic-reviews).

A system prompt is most useful when you need a consistent persona or set of operating rules across a multi-step conversation. For a one-off task, a detailed user prompt is usually more effective.

## Part C: prompt structure template [#part-c-prompt-structure-template]

### The ROLE/CONTEXT/TASK/FORMAT/CONSTRAINTS framework [#the-rolecontexttaskformatconstraints-framework]

<Aside label="When to use">
Use this structure for complex, multi-step analysis. For simpler tasks, ask a direct question.
</Aside>

Complex research tasks often benefit from explicit section headers. I use this layout when a single-paragraph prompt leaves too many degrees of freedom:

```markdown
# ROLE
Define the AI's expertise and perspective

# CONTEXT
Provide necessary background information

# TASK
Specify the exact analytical work required

# FORMAT
Describe the expected output structure

# CONSTRAINTS
Set boundaries and quality criteria

# EXAMPLES (optional)
Show desired output format
```

### Applied example [#applied-example]

The same structure, filled in as an active prompt. Slots in plain brackets describe your material; the `[insert: …]` slots are the judgments the prompt cannot make for you.

```markdown
# ROLE
You are assisting a researcher in [field] who is identifying themes
in the literature on [topic]. The conclusions are the researcher's.

# CONTEXT
I have [number] papers on [topic], listed with IDs in [file].
My question: [insert: the question the themes must answer]
My current account: [insert: the themes I expect, in two or three sentences]

# TASK
Step 1: quote every passage that bears on my question, with paper ID
and page. Stop for my review.
Step 2, after I reply: group the passages into candidate themes.

# FORMAT
For each theme:
1. Theme name
2. Description (2-3 sentences) in the papers' terms
3. Supporting passages: paper ID, page, quote
4. Papers that contradict or complicate it

# CONSTRAINTS
- A theme needs support from at least [insert: your threshold] papers
- By "generic" I mean [insert: for example, a statement true of any
  organization]; flag such themes rather than dropping them
- List disagreements before agreements
- If a paper is silent on a theme, say so; do not infer

# EXAMPLES
[insert: one theme written the way you want it, from your own reading]

If anything in [insert: …] is still unfilled, stop and ask me before starting.
```

## Part D: your prompt workspace [#part-d-your-prompt-workspace]

Course worksheets now live with the systematic-review module: the [initial prompt worksheet](/docs/case-studies/systematic-reviews/review-prompt-templates#initial-prompt-worksheet), the [refined prompt worksheet](/docs/case-studies/systematic-reviews/review-prompt-templates#refined-prompt-worksheet), and the [agentic workflow design](/docs/case-studies/systematic-reviews/review-prompt-templates#agentic-workflow-design). Outside a review, keep the same habit: store your own account beside each prompt version and note where the output diverged.

## Part E: organizing your work [#part-e-organizing-your-work]

### Creating your prompt folder [#creating-your-prompt-folder]

1. Create a `prompts/` folder inside your project workspace (for example, `my-research/prompts/`).
2. Save dated or numbered versions as you refine them, such as `synthesis-v1.md` and `synthesis-v2.md`.
3. Note beside each version which failure mode prompted the change.
4. Promote your stable prompts into Cherry Studio agents or reusable workspace skills (`.claude/skills/` or `$skill` packages; see [Agent Runtimes](/docs/implementation/ai-environment-setup/cli-setup-guide#4-reusable-procedures-and-skills)).

### Prompt development tips [#prompt-development-tips]

- Start with a minimal skeleton, then add constraints only where the model drifts.
- Test a new prompt on two or three known papers before running a larger batch.
- Record what failed and why; a prompt that exposed a contradiction in your construct definition is useful data.
- Keep old versions under version control so you can see whether an edit improved the output or merely lengthened the instructions.

## Why this approach works [#why-this-approach-works]

Templates copied from elsewhere rarely survive contact with your own sources. Building prompts from explicit questions, pre-reading accounts, and defined judgment terms keeps your methodological decisions inspectable. Once a prompt stabilizes on small samples, move it into an instruction file (`CLAUDE.md` or `AGENTS.md`) or a multi-step agent workflow. That preserves who made the interpretive choices.

## Next steps [#next-steps]

<IndexRows>
  <IndexRow title="AI Model Guide" href="/docs/craft/ai-model-reference-guide">
    Choose the right model for each task
  </IndexRow>
  <IndexRow title="Failure Museum" href="/docs/craft/failure-museum">
    Learn from common AI mistakes
  </IndexRow>
  <IndexRow title="Review Prompt Templates" href="/docs/case-studies/systematic-reviews/review-prompt-templates">
    Course worksheets and agent briefs for systematic reviews
  </IndexRow>
</IndexRows>