Make AI answers easier to follow—and easier to return to.

---
name: ai-working-style-builder
description: Help users improve how they work with AI. Use for communication, attention, continuity, instructions, or record-keeping problems.
---

# AI Working Style Builder

Help the user teach AI how to work with them. Produce an early useful example, then deepen the setup only when wanted.

## Understand the User's Arrival

Run after the user deliberately invokes the skill or asks to improve how they work with AI or with the current assistant.

Before drafting instructions or a record system, establish what brought the user here. They may:

- know the immediate irritation but not the underlying problem;
- have several connected communication, attention, or continuity problems;
- want a one-off improvement or an evolving working system;
- be using an ordinary chatbot or a desktop agent with files;
- have followed a recommendation without knowing what outcome to request.

If the user has already described the problem, work from that account. Otherwise invite a natural description or dictation. Do not make the user diagnose themselves or complete a setup questionnaire.

Read [`references/interaction-principles.md`](references/interaction-principles.md) before guiding the conversation.

Read [`references/cognitive-accessibility-framework.md`](references/cognitive-accessibility-framework.md) when designing substantive instructions or explanations, mapping load across attention, perception, cognition, navigation, or organisation, or connecting chat to HTML and durable records.

## Apply Accessible Defaults

Do not wait for the user to disclose overload, distraction, disability, likely interruption, or a preferred learning style. Assume variable attention, knowledge, reading depth, and interruption.

For substantive explanations and multi-stage instructions:

1. answer the immediate question first;
2. expose a quick route, guided route, uninterrupted complete route, and optional technical depth;
3. avoid repeated permission questions that interrupt users who want the complete account;
4. offer a resumable record when the work may continue beyond chat.

Let users move between routes without explaining or defending the choice.

Treat configuration, process, and other multi-step how-to requests as substantive even when the user asks a short, ordinary question. In the initial response, give the short answer and then state the available reading and resumption routes in one unobtrusive sentence. Do not replace those routes with only an offer to identify the product or provide more detail.

## Keep User Artefacts About the User's Work

Separate the reasoning used to design an answer from the answer the user receives.

For guides, examples, HTML, custom instructions, and other user-facing artefacts:

- start with a problem the user can recognise and the result they can obtain;
- address the user directly and use the user's task language;
- describe what the user can do, choose, keep, or share;
- show a finished result before explaining the framework behind it;
- keep implementation details, agent instructions, cognitive dimensions, and design rationale out of the main reading path;
- put full skill source or technical material in a clearly secondary appendix when it must travel with the artefact;
- use headings for navigation, not as slogans or oversized decoration;
- bold load-bearing words in explanatory prose so the eye can scan actions and distinctions.

Do not turn internal conversation notes into public copy. A principle may shape the result silently without becoming something the user has to read.

## Guide the Journey

1. **Understand:** Identify the user's purpose, environment, immediate difficulty, and desired scale of help.
2. **Reflect:** Summarise the problem in plain language. Distinguish presentation difficulties from continuity or record-keeping difficulties without imposing those categories on the user.
3. **Demonstrate early:** Show one small generic example as soon as a plausible pattern emerges. Keep it short and relevant; do not present a catalogue of alternatives.
4. **Test:** When useful, apply the same principle to the user's content. Separate aesthetic preference from observed usefulness.
5. **Deliver:** Offer a useful starter result after no more than two or three exchanges when possible.
6. **Refine:** Offer deeper tuning or durable records only after the user has seen a result.

Read [`references/problem-and-experiment-patterns.md`](references/problem-and-experiment-patterns.md) when choosing a communication pattern or comparison.

## Address Continuity When Relevant

Explain the principle rather than prescribing wording: chat supports exploration, but important decisions, plans, instructions, and lessons need a durable home when they must survive the conversation.

Offer continuity as a normal capability when work may outlast the conversation; do not present it as a response to an inferred deficit. Do not force a record system when a small communication change is enough. When lost history is already the problem, invite a fuller account and propose the smallest useful journal or resumable document. Prefer dictation unless relevant files are already available or the user offers a chat or document.

Prefer a disk-based agent for durable repeated work. Offer editable custom instructions, approved chatbot memories, and resumable documents as fallbacks. Read [`references/durable-record-patterns.md`](references/durable-record-patterns.md) before designing a record system.

## Produce an Appropriate Result

Choose only what helps now:

- a small communication experiment;
- editable custom instructions;
- a concise working-style profile;
- a journal or resumption pattern;
- portable Markdown;
- optional HTML for reading, later feedback, or later sharing;
- Word or PDF when preferred.

Keep reading, feedback, and sharing versions separate unless the user requests a combined artefact. Use `single-html-document` for HTML. Read [`references/public-foundations.md`](references/public-foundations.md) when the user wants rationale or public resources.

For reading HTML, support quick, guided, uninterrupted complete, and optional technical-depth routes through the same content. Make the primary surface useful without knowledge of the skill or its design process. Keep an editable journal or Markdown record canonical when progress or decisions must persist; generate HTML from it and return approved feedback to it.

For repeated use, inspect the existing profile and relevant journal history. Classify proposed changes as **Trying**, **Working well**, or **Standing instruction**. Preserve earlier decisions and reasons for change.

Require approval before changing standing agent instructions, custom instructions, persistent memory, or public material. A user-approved disk-based setup may record routine observations automatically.

## Reference Map

- [`references/interaction-principles.md`](references/interaction-principles.md) — who the user may be, how they arrive, and why the interaction rules exist. Read for every run.
- [`references/cognitive-accessibility-framework.md`](references/cognitive-accessibility-framework.md) — core working framework for universal access paths, five cognitive dimensions, delivery layers, and a complete computer-configuration example. Preserve as a working draft for further refinement.
- [`references/problem-and-experiment-patterns.md`](references/problem-and-experiment-patterns.md) — problem patterns, early examples, readability concepts, and comparison methods.
- [`references/durable-record-patterns.md`](references/durable-record-patterns.md) — journals, current instructions, lessons, resumable records, and chatbot fallbacks.
- [`references/public-foundations.md`](references/public-foundations.md) — optional public readability, HTML, and agent-instruction resources.
- [`assets/ai-working-style-profile.template.md`](assets/ai-working-style-profile.template.md) — optional durable profile template; remove sections that do not help.
- [`README.md`](README.md) — human-facing introduction, user journey, and paste-in prompt for ChatGPT, Claude, or Gemini.

# AI Working Style Builder

*Teach AI how you work.*

AI Working Style Builder helps with two related problems:

- **how AI communicates with you** — structure, length, clarity, pacing, and useful formats;
- **how work survives the conversation** — decisions, plans, notes, and things you want to return to.

## Who this is for

You do not need to know exactly what is wrong before you begin. You may have noticed that:

- long answers make the main point difficult to find;
- detailed instructions are useful but tiring to follow;
- step-by-step help becomes tedious;
- conversations wander and the original task disappears;
- decisions or useful context become difficult to retrieve;
- you keep repeating information across chats;
- an output is hard to review, reuse, or share.

You may have found this skill yourself, received the link from someone else, installed it in a desktop agent, or pasted its prompt into an ordinary chatbot. You may want one quick improvement or an evolving working system. The process should adapt to that situation rather than assume everyone needs the same setup.

Accessibility and neurodivergence may be relevant, but the skill starts from concrete problems rather than diagnostic labels.

## How the process works

The skill follows a few principles:

- **Design for variable attention by default:** It does not wait for you to identify overload, distraction, likely interruption, or a preferred learning style before providing accessible structure and alternative depths.
- **Understand before configuring:** It learns what brought you here and what work you want to improve before drafting durable instructions.
- **Accept an incomplete account:** You can dictate naturally, mention several connected problems, or change your mind later.
- **Show an example early:** Once a likely pattern appears, the assistant demonstrates one small contrast rather than delaying all value until the interview is complete.
- **Keep examples bounded:** It normally compares two compact versions, not a large catalogue that creates more overload.
- **Make one decision at a time:** Questions stay focused, while facts already available in your account or files are not asked again.
- **Test preference against usefulness:** What looks appealing and what helps you find or act on information may differ.
- **Deliver before exhaustive discovery:** You should normally see a useful experiment or starter result within two or three exchanges.
- **Preserve revision history:** New preferences begin as experiments. You can backtrack, refine them, and keep the reasons for changes.

## Start with an installed skill

In Codex, Claude Code, or another agent with this skill installed, invoke **AI Working Style Builder** and explain that you want to improve how you work with AI or with that assistant. Add whatever you already know about the problem; you do not need to use exact wording.

The agent loads [`SKILL.md`](SKILL.md), then only the supporting references relevant to your situation.

## Use it in ChatGPT, Claude, or Gemini

Paste the prompt below into a new conversation. It works without installing a skill. If the chatbot can open links, it can also consult the [browser guide](https://techczech.github.io/dominiks-agent-skills/estate/ai-working-style-builder/assets/ai-working-style-builder-guide.html) or the [public skill source](https://github.com/techczech/dominiks-agent-skills/tree/main/estate/ai-working-style-builder).

```text
I'd like to improve how I work with AI and how I work with you.

Help me discover what would make the biggest difference. First understand what brought me here, what I am trying to do, and whether I want a quick improvement or a more durable setup. Invite me to describe or dictate several connected problems without organising them first. If I have already explained the problem, work from that account instead of asking me to repeat it.

Treat variable attention, knowledge, reading depth, and interruption as normal. Do not wait for me to disclose overload, distraction, disability, likely interruption, or a preferred learning style before using accessible structure. For substantive explanations or multi-stage instructions, answer the immediate question first. Make a quick route, guided route, uninterrupted complete route, optional technical depth, and resumable version available. Expose these routes without making me answer repeated permission questions or justify why I want one of them.

Reflect the problem in plain language. Consider both how information is presented and whether decisions, plans, or context are being lost in chat. Do not diagnose me or assume that my first stated preference is necessarily what works best.

Give me a small example sooner rather than later. Once you see a plausible pattern, explain the principle briefly and compare two compact generic examples. Do not overwhelm me with a catalogue. When useful, offer to test the same principle on my content. Distinguish what I like aesthetically from what helps me find, understand, or act on information.

Do not turn this into a long questionnaire. Ask one decision question at a time, prefer one substantial dictation to many small questions, and accept backtracking. Within two or three exchanges, offer a useful result, a starter version, or the choice to continue refining.

Possible results include editable custom instructions, a communication experiment, a working-style profile, or a simple way to preserve decisions and resume work. Treat new preferences as experiments first. Let me review any custom instructions or proposed memory before adopting them.

If lost history matters, explain the principle that chat supports exploration but important decisions, plans, instructions, and lessons need a durable home. Recommend the smallest useful journal or resumable document. Mention disk-based desktop agents as the stronger option for repeated work while accommodating ordinary chatbot use.

Keep output human-readable. Offer Markdown or a journal entry as an editable canonical source and, when useful, HTML as a separate reading surface. A reading HTML version may provide quick, guided, complete, and technical-depth routes through the same content. Offer feedback and sharing versions later instead of combining those purposes by default.

Full guide, if you can open links:
https://techczech.github.io/dominiks-agent-skills/estate/ai-working-style-builder/assets/ai-working-style-builder-guide.html
```

Add your own account below the prompt or send it in the next message. Dictation is often easier than formulating the problem precisely.

## Use the shareable HTML guide

Choose the route that fits what you want to do:

- **Read it online:** Open the [interactive GitHub Pages guide](https://techczech.github.io/dominiks-agent-skills/estate/ai-working-style-builder/assets/ai-working-style-builder-guide.html).
- **Keep or attach one file:** [Download the single HTML guide](https://github.com/techczech/dominiks-agent-skills/raw/refs/heads/main/estate/ai-working-style-builder/assets/ai-working-style-builder-guide.html).
- **Inspect or install the skill:** Browse the [public source package](https://github.com/techczech/dominiks-agent-skills/tree/main/estate/ai-working-style-builder).

In a local checkout, open [`assets/ai-working-style-builder-guide.html`](assets/ai-working-style-builder-guide.html). The file is a self-contained introduction to the skill, its principles, a worked example, and the full instructions needed to use the method in another AI chat.

The downloaded file can be attached directly to ChatGPT, Gemini, Claude, or another chatbot that accepts documents. The full skill text is included in the file. The page can also download the same instructions as one Markdown file when a chatbot does not accept HTML reliably.

## What the journey looks like

1. **You initiate it:** You ask to improve how you work with AI or describe a specific difficulty.
2. **The assistant understands your situation:** It identifies your purpose, environment, immediate problem, and desired scale of help.
3. **You see a small example:** The assistant demonstrates one likely useful principle without waiting for a complete profile.
4. **You try a result:** This may be a reformatted example, communication rule, custom instruction, or resumption note.
5. **You choose the depth:** Stop with the quick result or continue into durable instructions and records.
6. **You refine through use:** Keep what works, revise what does not, and preserve the reasons for changes.

The skill may be used once or twice to produce a practical result. It can also be revisited after real work reveals new needs.

## The central information-management principle

Clearer responses solve only part of the problem. Chat is useful for exploration and feedback, but important decisions, plans, instructions, lessons, and things to return to need a durable home when they must survive the conversation.

A desktop agent with access to files can maintain that history most reliably. Custom instructions, chatbot memory, and saved resumption documents remain useful fallbacks.

## Reference guide

Every supporting file is listed here for easy review:

- [`SKILL.md`](SKILL.md) — compact operational instructions for an installed agent.
- [`references/interaction-principles.md`](references/interaction-principles.md) — explained principles governing arrival, pacing, questions, examples, revision, and user ownership.
- [`references/cognitive-accessibility-framework.md`](references/cognitive-accessibility-framework.md) — core working framework for universal access paths, attention, perception, cognition, navigation, organisation, HTML reading surfaces, and durable records.
- [`references/problem-and-experiment-patterns.md`](references/problem-and-experiment-patterns.md) — common problems, readability concepts, early-example method, and pacing signals.
- [`references/durable-record-patterns.md`](references/durable-record-patterns.md) — journal, instruction, lesson, resumption, and chatbot-fallback patterns.
- [`references/public-foundations.md`](references/public-foundations.md) — optional public readability, HTML, and agent-instruction resources.
- [`assets/ai-working-style-profile.template.md`](assets/ai-working-style-profile.template.md) — optional profile template for a durable setup.
- [`assets/ai-working-style-builder-guide.html`](assets/ai-working-style-builder-guide.html) — shareable single-file guide with the full chatbot-readable skill text included.

interface:
  display_name: "AI Working Style Builder"
  short_description: "Improve how you work with AI"
  default_prompt: "I’d like to improve how I work with AI. Use $ai-working-style-builder to help me find and try the changes that would make the biggest difference."

---
document_type: ai-working-style
status: draft
updated: YYYY-MM-DD
---

# My AI Working Style

## What I Am Trying to Make Easier

[Describe the user's problems and the work they want to improve.]

## Current Standing Instructions

[Include only instructions the user has approved. Use clear human-readable propositions.]

## Working Well

[Record patterns supported by experience but still open to revision.]

## Trying

[Record current experiments, the context in which to try them, and what would count as useful.]

## Communication

[Record answer order, structure, density, pacing, questions, language, evidence, and format preferences. Add contextual exceptions.]

## Continuity and Journal

[Record what should be captured, where durable records live, how diversions are noted, and how work is resumed.]

## Examples

[Include approved examples showing the instructions in practice.]

## Revision

[State how observations enter the journal and which changes require approval. Link the relevant history when files are available.]

# Interaction Principles

These principles shape the conversation. Apply them with judgement rather than as a fixed script.

## 1. Start from the user's purpose

The user initiated the process because something about working with AI is difficult or could be better. Establish what they are trying to accomplish and what brought them to the skill before producing durable instructions.

## 2. Understand how the user arrived

The user may have followed a recommendation, invoked an installed skill, pasted a prompt into a chatbot, or returned to refine an earlier setup. Their environment and expectations determine what the assistant can usefully produce.

## 3. Design for variable attention by default

Assume that any user may want an immediate answer, a guided route, an uninterrupted complete explanation, optional technical depth, or a place to resume. Offer appropriate access paths without requiring the user to disclose overload, distraction, disability, or a preferred learning style. Treat an ordinary multi-step configuration or process request as sufficient reason to offer them.

## 4. Accept an incomplete problem description

The user may recognise overload, impatience, lost attention, repeated context, or missing history without knowing the cause. Treat their account as evidence to interpret together, not a specification they must get right at the start.

## 5. Let connected problems arrive together

Communication, information processing, attention maintenance, and record-keeping often overlap. Allow a broad account or dictation before narrowing the work, particularly when repeated short questions would fragment the story.

## 6. Give an example early

An example makes an abstract recommendation testable and helps the user discover preferences they could not state in advance. Show one small contrast as soon as a plausible pattern emerges; do not wait for a complete profile.

## 7. Do not overwhelm with examples

Too many alternatives recreate the original problem. Demonstrate one principle with two or at most three compact variants, then pause for reaction or application to the user's content.

## 8. Ask one decision question at a time

One decision reduces working-memory demands and makes ownership clear. Gather several factual details independently when possible, but do not bundle unrelated choices into one response.

## 9. Prefer rich dictation to prolonged interviewing

A longer natural account often reveals relationships that serial questions hide. Use it especially for lost history, wandering work, or situations the user finds difficult to categorise.

## 10. Explain principles rather than prescribe phrases

Describe the intended effect, the evidence behind it, and the relevant trade-off. Examples illustrate a principle; they are not mandatory wording unless the user chooses to adopt them as custom instructions.

## 11. Distinguish preference from performance

Record what the user likes, what helps them complete a task, and the contexts in which those differ. A format can be aesthetically unwelcome yet useful for scanning, or pleasant to read yet poor for retrieval.

## 12. Deliver before exhaustive discovery

The user should normally see a useful experiment or starter result within two or three exchanges. After that, let them stop, adopt a provisional result, or continue refining.

## 13. Accept backtracking and changing preferences

Working styles are revisable hypotheses, not personality diagnoses. Preserve what was tried, what changed, and why; do not frame a revised preference as inconsistency or failure.

## 14. Separate immediate relief from durable continuity

Clearer responses may solve the immediate problem while leaving decisions and plans trapped in chat. Introduce durable records when relevant, but do not force a record system onto a user who only needs a small communication adjustment.

## 15. Keep durable changes user-owned

The assistant may propose and demonstrate. The user approves standing instructions, custom instructions, persistent memories, and sharing copies before they become authoritative.

## 16. Separate design reasoning from the user's artefact

The assistant may use readability, cognitive-accessibility, and information-management frameworks to design a result. The resulting guide or document should not narrate that internal reasoning unless the user asks for it.

Begin with the user's problem, task, or desired result. Use headings that help the user act or navigate. Show what a reformatted instruction, custom instruction, journal entry, or resumable guide looks like in use. Keep framework names, agent-facing notes, source-package terminology, and commentary about the design conversation in a secondary technical appendix or outside the artefact.

Visual hierarchy must reduce load rather than advertise the design. Large headings should remain short and should not consume most of a viewport. Use bold emphasis for the words a reader needs to scan.

# Cognitive Accessibility Framework

**Status:** Core reference; working draft. Preserve and refine through later examples and user feedback.

## Contents

1. [Purpose](#purpose)
2. [Universal design baseline](#universal-design-baseline)
3. [Five dimensions](#five-dimensions)
4. [Delivery layers](#delivery-layers)
5. [Relationship to readability](#relationship-to-readability)
6. [User-facing delivery check](#user-facing-delivery-check)
7. [Worked example](#worked-example)
8. [Load analysis](#load-analysis)
9. [Application in AI Working Style Builder](#application-in-ai-working-style-builder)
10. [Sources](#sources)

## Purpose

Extend readable and accessible document design beyond surface formatting. Evaluate how information design affects:

- maintaining focus;
- discerning what matters;
- understanding the content;
- moving through and returning to it;
- preserving continuity across interactions.

Treat the dimensions as overlapping lenses, not independent stages or diagnoses.

Accessibility applies to the whole information arrangement: content, conversation, interface, and records. Clear prose helps, but it cannot by itself provide alternative reading paths, persistent position, or continuity across conversations.

## Universal Design Baseline

Assume that every user may:

- want the answer immediately;
- want to inspect the complete explanation before acting;
- prefer guidance one stage at a time;
- leave and return unexpectedly;
- change the desired level of detail;
- need the result outside the original conversation.

Do not wait for the user to identify overload, distraction, disability, or a preferred learning style. Do not make accessible structure conditional on a diagnosis or self-description.

For substantive explanations and multi-stage instructions:

1. answer the immediate question first;
2. expose a quick route, guided route, and uninterrupted complete route without requiring repeated permission;
3. make optional technical depth easy to reach without placing it in the main route;
4. offer a resumable record when the work may continue beyond the conversation;
5. let the user move between routes without explaining why.

These are ordinary access paths, not special accommodations.

## Five Dimensions

### Attention

**Question:** How easily can the user remain focused on what matters without being drawn into irrelevant detail?

Common stressors:

- unsolicited alternatives;
- competing signals;
- excessive background before the task;
- long undifferentiated responses;
- unrelated next steps.

Helpful responses:

- put the immediate purpose first;
- bound the current scope;
- keep one active decision visible;
- separate optional detail;
- use small examples rather than catalogues.

### Perception

**Question:** How difficult is it to discern what needs attention?

Common stressors:

- weak hierarchy;
- dense layout;
- inconsistent formatting;
- colour-only distinctions;
- visually competing decoration.

Helpful responses:

- use space and contrast;
- mark semantic headings correctly;
- keep paragraphs and lines bounded;
- use meaningful emphasis;
- communicate distinctions through more than colour alone.

### Cognition

**Question:** How easily can the user understand the content without holding too many unresolved elements in working memory?

Common stressors:

- complex sentences;
- implicit actions or decisions;
- too many simultaneous steps;
- unexplained terminology;
- distant explanations and consequences.

Helpful responses:

- express propositions and actions explicitly;
- use direct clauses and useful verbs;
- keep one principal action per step;
- place reasons near the relevant action;
- demonstrate unfamiliar ideas with concrete examples.

### Navigation

**Question:** How easily can the user locate a relevant part, understand their position, and return after interruption?

Common stressors:

- missing overview;
- vague headings;
- long linear-only presentation;
- unclear progress or completion;
- unsearchable text and links.

Helpful responses:

- show an outline when the content warrants one;
- use descriptive headings and numbered steps;
- mark current position, expected result, and completion;
- provide meaningful links and searchable text;
- include resumption points.

### Organisation

**Question:** How easily can the user maintain continuity and memory across documents, conversations, and sessions?

Common stressors:

- decisions trapped in chat;
- changing instructions with no history;
- unfinished threads with no return point;
- repeated reconstruction of context;
- no current-state summary.

Helpful responses:

- maintain a durable journal or resumable record;
- record decisions with reasons;
- distinguish current guidance from chronological history;
- preserve things to return to and the next action;
- link central instructions to supporting detail.

## Delivery Layers

Use four connected layers when the work extends beyond a short answer:

### Content

State the outcome, actions, reasons, conditions, and expected results. Keep the content independent of one conversational sequence or visual presentation.

### Conversation

Use chat for interpretation, clarification, and adaptive help. Give a useful answer before offering delivery choices. Do not require the user to proceed through repeated questions when the complete route can be made available.

### Reading Interface

Use semantic HTML when the material benefits from parallel reading paths, visible position, non-linear navigation, progressive reveal, search, annotation, or offline reuse. Treat HTML as a delivery surface, not the only copy of the information.

A reading interface may contain quick, guided, complete, and technical-depth routes because these serve different reading strategies. Keep feedback collection and sharing as separate versions unless the user requests a combined artefact.

### Durable Record

Keep current state, decisions, next actions, and history outside chat when they must survive the interaction. Prefer an editable journal or Markdown record as the canonical source. Generate reading HTML from that source; return progress or annotations to the record rather than allowing independent copies to drift.

The layers have different jobs:

- chat supports the working conversation;
- the record preserves continuity;
- HTML supports reading and navigation;
- separate feedback or sharing versions support later purposes.

## Relationship to Readability

| Readability principle | Primary load reduced | Supporting effects |
| --- | --- | --- |
| **Space** | Perception and attention | Makes content more approachable and reduces visual competition |
| **Chunks** | Cognition and attention | Supports scanning, resumption, and bounded action |
| **Guides** | Perception and navigation | Shows structure, importance, and possible reading paths |
| **Information structure** | Attention and cognition | Puts the useful proposition before background |
| **Simple language** | Cognition | Reduces grammatical and lexical processing demands |
| **Semantic accessibility** | Perception and navigation | Supports screen readers, heading navigation, meaningful links, and non-colour cues |
| **Durable records** | Organisation | Extends accessibility beyond one document or interaction |
| **Parallel depth** | Attention, cognition, and navigation | Lets users choose a quick, guided, complete, or technical route without repeated interruption |

One design choice may affect several dimensions. A descriptive heading improves perception, supports navigation, and reduces cognitive effort by revealing the proposition before the section is read.

## User-Facing Delivery Check

Apply the framework silently before exposing it. A user-facing guide should pass these checks:

- **The first screen names a recognisable problem or useful result.** It does not begin with the design theory, a slogan, or a map of internal system components.
- **Headings answer the reader's questions.** They describe what the reader can do, what an option changes, or what a result contains.
- **Examples show results.** Present the finished instructions, profile, journal, or reading routes before discussing why the structure works.
- **Language addresses the reader.** Replace implementation terms such as “operational source package”, “delivery layer”, or “durable record” with the action or benefit relevant to the reader.
- **Visual hierarchy supports scanning.** Keep display headings short, constrain their scale, bold load-bearing words, and avoid layouts where decorative type or secondary panels compete with the task.
- **Technical material remains secondary.** When an HTML file must carry complete agent instructions, place them in a clearly labelled appendix that does not interrupt the main reading path.

Do not copy the design conversation into the artefact. The framework is successful when readers benefit from it without needing to learn its vocabulary.

## Worked Example

The example uses a generic cloud-storage application. It demonstrates information design rather than documenting one product's current interface.

### Ordinary Request

The user asks:

> How do I make my Projects folder available offline?

Do not require the user to add that long instructions are difficult, that they may be interrupted, or that they want an accessible format. Treat variable attention, reading depth, and interruption as normal conditions.

### Inadequate Chat-Only Answer

> To make the Projects folder available offline, open the cloud storage application and go to settings, although on some versions you may need to open the folder in Finder first, and select the Projects folder. Make sure you have enough disk space because this will download everything. Choose Always keep on this device or Make available offline. It may take some time. You should not close the application, and if there are errors check your internet connection. Once it has finished there should be a green icon, but icons differ between versions. You can then disconnect from the internet and open a file to test it.

Problems:

- the immediate action is buried;
- preparation, action, warning, and verification are mixed;
- interface variants compete for attention;
- the user cannot choose a different depth without restarting the exchange;
- no uninterrupted complete explanation is available;
- no stopping or resumption point is visible;
- nothing preserves the state outside chat.

### Accessible Default Response

Answer the task before presenting options:

> Select the **Projects** folder, open its context menu, and choose **Always keep on this device** or the equivalent offline option.
>
> Keep the cloud-storage application running until the download finishes. Then disconnect from Wi-Fi briefly and open a file to verify that it works.
>
> I can also put this into a browser guide with quick, guided, complete, and technical views, and save a resumable record of where you reached.

This response does not diagnose the user or turn the task into an interview. It provides immediate value and exposes the available routes once.

### Reading Routes

When the user wants a browser guide, give them parallel routes through the same content:

```text
PROJECTS FOLDER AVAILABLE OFFLINE

Quick route     Guided steps     Full guide     Technical detail
```

The routes support different reading strategies. They do not represent fixed types of user.

#### Quick Route

1. Select **Projects**.
2. Choose **Always keep on this device** or the equivalent offline option.
3. Wait for the download to finish.
4. Disconnect briefly and open a file.

Show the current position here when the task has already started.

#### Guided Steps

Present one stage prominently while keeping the overall sequence visible:

```text
Step 2 of 4: Make the folder available offline

Open the context menu for Projects.

Choose “Always keep on this device” or the equivalent option.

Expected result: The folder begins downloading.

Previous step                              Next step
```

Do not require confirmation after every step. Let the user move directly between stages.

#### Full Guide

Present the whole procedure as a continuous, searchable document. Do not interrupt it with conversational questions.

##### Outcome

Files in the Projects folder will open without an internet connection.

##### Before Starting

Check that:

- the cloud-storage application is running;
- the Projects folder has finished synchronising;
- the computer has enough free space for the folder.

The download can continue while other work is taking place.

##### 1. Find the Projects Folder

Open the cloud-storage folder in Finder or File Explorer.

Select **Projects** without opening it.

**Expected result:** The Projects folder is selected.

##### 2. Make the Folder Available Offline

Open the folder's context menu.

Choose **Always keep on this device**, **Make available offline**, or the equivalent option used by the application.

Avoid changing unrelated synchronisation settings.

**Expected result:** The application begins downloading the folder.

##### 3. Wait for the Download

Keep the cloud-storage application running.

**Expected result:** The application marks the folder as downloaded or available offline.

##### 4. Verify the Result

Temporarily disconnect from Wi-Fi.

Open one file from the Projects folder.

- If the file opens, the configuration works.
- If it does not open, reconnect and check whether downloading has finished.

Reconnect to Wi-Fi after the test.

##### Resume Later

- Continue from Step 3 if downloading is still in progress.
- Continue from Step 4 if the folder appears fully downloaded.
- Return to Step 2 if no offline setting is active.

#### Technical Detail

Keep optional explanations reachable without placing them in the main route. They may cover:

- the difference between synchronised and offline files;
- the meaning of status indicators;
- local-storage use;
- what happens when a file changes on another device;
- product-specific labels and recovery paths.

The user can move into this material without requesting permission from the assistant and return to the task without losing position.

### Durable Setup Outside Chat

When the task may continue beyond the conversation, offer a resumable record as a normal capability rather than as a response to an inferred deficit:

> I can save the instructions and current position so this can be resumed from another conversation.

Use one canonical journal entry where possible:

```markdown
# Projects Folder Available Offline

Status: In progress
Current point: The folder is downloading
Next action: Check whether the download has finished
Verification: Not yet completed

## Intended Result

Files in Projects should open without an internet connection.

## Guide

[The complete instructions appear here.]

## Journal

### 12 August 2026

- Selected Projects.
- Used “Always keep on this device”.
- Download started.
- Resume from Step 3.
```

Generate the reading HTML from this record. If the user returns in another conversation, read the record and resume from the explicit next action rather than reconstructing the state from chat history.

### Separate Later Purposes

The reading HTML may include quick, guided, complete, and technical routes because all four support reading. Do not add feedback collection or sharing controls by default.

After the user has reviewed or used the guide:

- offer a separate feedback version for progress, annotations, corrections, or exported notes;
- offer a separate privacy-checked sharing version without personal history or private paths.

A portable feedback page should export its observations. The agent applies approved feedback to the canonical journal and regenerates the reading version. Do not allow the journal, reading copy, feedback copy, and sharing copy to become independent sources of truth.

## Load Analysis

### Attention

The immediate answer appears before delivery choices. The quick and guided routes keep the active task prominent. Technical detail remains available without competing with the main route.

### Perception

The HTML gives preparation, action, expected results, current position, and recovery information distinct roles. Semantic headings and text labels reveal hierarchy without relying on colour.

### Cognition

Each step contains one principal action, while the full guide preserves the complete model. The user can select depth without holding earlier chat messages or unanswered questions in working memory.

### Navigation

Quick, guided, complete, and technical routes support different reading strategies. Numbered steps, current-position cues, searchable headings, and resumption points support movement within and between sessions.

### Organisation

The journal preserves instructions, current state, observations, and the next action outside chat. HTML remains a generated reading surface rather than a competing record. Another interaction can resume without reconstructing the work from chat history.

## Application in AI Working Style Builder

Use this framework to:

1. apply accessible structure before the user identifies a difficulty;
2. identify additional load from the user's account without diagnosing them;
3. answer the immediate task before offering alternative delivery routes;
4. provide quick, guided, uninterrupted complete, and optional technical-depth routes for substantive material;
5. demonstrate a relevant principle with one bounded example;
6. test the same principle on the user's real content;
7. record aesthetic response and practical usefulness separately;
8. treat successful patterns as provisional until tested in real work;
9. offer a durable record when the work may outlast the conversation;
10. keep reading, feedback, and sharing artefacts separate by default.

Do not present the full framework during a normal first interaction. Apply it silently, answer the task, and expose routes without requiring self-disclosure. Use the complete worked example for review, teaching, or deeper refinement.

## Sources

- [Readability guidance](https://readability.edutools.fyi)
- [Public `readability-skill` repository](https://github.com/techczech/readability-skill)
- [Five principles of readability](https://github.com/techczech/readability-skill/blob/main/references/five-principles.md)
- [Reading strategies](https://github.com/techczech/readability-skill/blob/main/references/reading-strategies.md)
- [Accessibility considerations](https://github.com/techczech/readability-skill/blob/main/references/accessibility.md)
- [Document tips](https://github.com/techczech/readability-skill/blob/main/references/document-tips.md)

These public foundations inform the working draft. The five-dimension model and its application to continuity across interactions remain provisional and should be refined through examples and user feedback.

# Problem and Experiment Patterns

Use only the patterns that help with the user's stated situation. Recommend a path; do not turn this map into a questionnaire.

## First-Conversation Budget

Before producing durable instructions, understand enough to identify:

- the work the user wants to improve;
- the difficulty they notice;
- its practical effect;
- the environment they are using;
- whether they want a quick experiment or a more durable setup.

These are information needs, not five mandatory questions. Infer what the user has already supplied. For lost history, prefer one longer account over serial prompts.

Do not make accessible presentation depend on this discovery. Apply readable structure immediately; use discovery only to refine the result or create durable personal instructions.

## Universal Delivery Pattern

For any substantive explanation or multi-stage instruction:

1. answer the immediate question;
2. make a quick route available;
3. provide a guided route when execution benefits from stages;
4. provide an uninterrupted complete route for users who want the whole model;
5. keep optional technical depth reachable without placing it in the main route;
6. offer a resumable record when work may extend beyond chat.

Expose the routes once. Do not make the user proceed through repeated permission questions or justify why they want a particular depth.

Treat configuration, process, and other multi-step how-to requests as substantive even when the user's question is short. Put the route offer after the immediate answer and keep it to one unobtrusive sentence.

## Common Problems and Candidate Patterns

| Problem described | Patterns worth demonstrating |
| --- | --- |
| Any substantive explanation or multi-stage instruction | Immediate answer; quick route; guided route; uninterrupted complete route; optional technical depth; resumable record when relevant |
| Main point disappears | Answer first; short orientation; proposition-led paragraphs; bounded detail |
| Too much text to process quickly | Chunking; headings; summaries; visual guides; layered detail |
| Long instructions become tiring | Destination first; short phases; one actionable step; resumable checkpoint |
| Step-by-step becomes tedious | Give an overview and let the user move between guided and complete views |
| Bullets feel unattractive | Compare prose, bounded bullets, key–value pairs, or a table on one small example |
| Conversation wanders | Record the new thread; restate the current focus; preserve a return point |
| Decisions disappear | Record the decision and reason; keep dated history; derive a current view |
| Earlier work is hard to find | Durable Markdown; descriptive titles; linked instructions; retrieval cues |
| User repeats context | Resumable summary; stable reference; approved chatbot memory as fallback |
| Output is hard to share | Portable Markdown; reading HTML; later privacy-checked sharing copy |

## Information-Design Principles

Teach only what is relevant:

- **Space:** separation reduces visual competition.
- **Chunks:** bounded units support scanning and resumption.
- **Guides:** headings, emphasis, summaries, and navigation show where to look.
- **Information structure:** put the important proposition before supporting context.
- **Simple language:** prefer direct clauses and useful verbs without deleting necessary concepts.
- **Scanning versus skimming:** scanning locates a target; skimming builds a quick model of the whole.
- **Outline:** reveal relationships and order, not merely a decorative heading hierarchy.
- **Propositions:** make the claim or action explicit; do not bury it inside framing language.
- **Parallel depth:** provide quick, guided, complete, and technical routes without forcing the user through them sequentially.

## Early-Example Method

1. Select the single principle most likely to help.
2. Explain its intended effect in one or two sentences.
3. Show two compact versions of the same generic content.
4. Ask the user to find, explain, or act on something—not only choose which version looks nicer.
5. Offer to repeat the experiment with the user's content.
6. Record aesthetic response, task usefulness, and context separately.
7. Treat the result as **Trying** until it survives real use.

Use a third variant only when it tests a materially different structure. Stop before the comparison becomes another large information load.

## Problem-Led Language

Mention accessibility and neurodivergence as possible contexts, not as the organising frame. Prefer concrete descriptions such as attention, rapid processing of large amounts of text, difficulty with long instruction lists, resuming interrupted work, or finding earlier decisions. Never infer a diagnosis or fixed style from a label.

## Pacing Failure Signals

Stop gathering information and demonstrate a result when:

- the user repeats or circles the same concern;
- answers become shorter or impatient;
- the conversation itself starts wandering;
- enough evidence exists for a reversible experiment;
- the user asks to see what the advice would look like.

# Durable Record Patterns

Use plain language with users. Internal terms such as control plane, backlog, side quest, context window, and progressive disclosure should not be prerequisites.

## Core Principle

Chat is useful for exploration and feedback. Information that must survive the conversation—decisions, plans, instructions, lessons, and things to return to—needs a durable home.

Explain the principle and trade-off in the user's context. Do not prescribe a slogan or imply that every message needs archiving.

## Smallest Useful Pattern

Use one chronological journal containing:

- concise confirmed summary;
- what happened and what was decided;
- notes and ideas;
- plans and things to return to;
- working-style experiments;
- original dictation or source material underneath the summary.

Do not echo a long dictation into the live conversation. Preserve it in the source chat and include it in the durable entry when one is created.

At the next session, read the current summary and only the relevant journal entries. Do not reload the entire history indiscriminately.

## File-Based Pattern

When the environment supports durable files, recommend:

1. **Current instructions:** short, human-readable approved guidance.
2. **Journal:** chronological decisions, experiments, notes, plans, and raw accounts.
3. **References:** supporting instructions loaded only when relevant.
4. **Lessons:** reusable findings promoted from repeated experience, with provenance.

Keep the central instructions short and link to relevant supporting detail instead of copying it. Treat the journal as authoritative history and current instructions as the maintained view of what is presently in force.

## Worked Example: A Linked File-Based Setup

A user who does repeated work with a file-capable agent might keep:

- one chronological journal for decisions, plans, observations, and source notes;
- a short file containing the working instructions currently in force;
- supporting reference files that the agent opens only when relevant;
- a brief resumption note naming the current focus and next action;
- selected lessons promoted from patterns that have survived repeated use.

Advantages:

- stronger retrieval and a visible history of changes;
- reliable resumption across sessions;
- separation between current guidance and historical evidence;
- less dependence on the user's memory for every open thread.

Costs:

- setup and maintenance;
- conventions and folders to learn;
- storage, backup, and synchronisation decisions;
- best results require a desktop agent with file access.

This is a synthetic example of the pattern, not a template every user must reproduce. A simpler journal may be enough.

## Chatbot Fallbacks

When files are unavailable:

- produce editable custom instructions;
- propose a compact memory entry and obtain approval before saving it;
- create a dated resumable Markdown or document summary for later upload;
- encourage periodic export of important work rather than reliance on chat navigation alone;
- verify or qualify claims about what a particular product retains.

Preserve at least the current goal, decisions, useful context, things to return to, and next action.

## Change Model

- **Trying:** proposed and under test.
- **Working well:** supported by experience.
- **Standing instruction:** approved for continued use.

Allow routine observations to enter the journal automatically only in a user-approved disk-based setup. Require review before changing standing instructions, custom instructions, or persistent chatbot memory.

Never erase an earlier preference merely because it changed. Record the replacement, date, reason, and any context where the older guidance still applies.

## Resume Pattern

End substantive work with:

- current focus;
- what changed;
- decisions and reasons;
- things to return to;
- next action;
- source locations when available.

On resumption, present the current focus and next action first. Offer deeper history only when needed.

# Public Foundations

Offer these as optional resources. The working-style process must not depend on installation or assume that one pattern suits everyone.

## Readability

- Guide: <https://readability.edutools.fyi>
- Public skill: <https://github.com/techczech/readability-skill>

Use space, chunks, guides, information structure, and simple language as hypotheses to demonstrate. Test preferences on tasks; do not rely on intuition or diagnosis alone.

## HTML Reading, Review, and Sharing

- Public skill: <https://github.com/techczech/dominiks-agent-skills/tree/main/dev/single-html-document>
- Rationale and examples: <https://claude.com/blog/using-claude-code-the-unreasonable-effectiveness-of-html>

Offer three separate experiences progressively:

1. a polished reading version;
2. an optional decision and feedback version that exports annotations back to an agent or chatbot;
3. a separately prepared sharing version.

Keep Markdown as the editable source. Accommodate Word and PDF preferences while explaining the navigation, accessibility, annotation, and reuse benefits of HTML.

## Durable Agent Instructions

- Global instructions: <https://github.com/techczech/dominiks-agent-skills/tree/main/estate/global-agent-instructions>
- Instruction streamlining: <https://github.com/techczech/dominiks-agent-skills/tree/main/estate/agents-md-streamline>

Use one short current instruction set with links to detail loaded when relevant. Keep the chronological journal and lessons outside the always-loaded instructions.

## Sharing Boundary

- Link public resources directly.
- Do not expose private repositories, paths, raw dictation, identities, organisations, clients, or machine details.
- Generate a separate privacy-checked sharing copy; do not rely only on hidden filters inside a private document.
- Verify changing product names and capabilities before recommending a specific chatbot or desktop agent publicly.