What goes in, and what comes out

I'm taking part in an independent research study about how people work with AI, and I want an honest assessment of how I work with you. You'll produce one markdown report. Part of it is for me to act on and part is for the research, but I'll upload the whole report to the researcher after I've reviewed it, so the privacy rules apply to all of it.

## Rules

- **Be honest, not flattering.** Base every finding on evidence you can actually see. If the evidence is thin, say so. Don't inflate my level or my habits.
- **Read only; change nothing.** Don't edit, delete or create any file other than the report and its readable copy.
- **Protect privacy.** In the report, do not include: names of people, companies, clients or employers; addresses, emails or phone numbers; passwords, keys, tokens or any credentials; financial, health or other sensitive personal details; proprietary code, data or confidential business information. Describe things generically ("a field-service company", "a colleague", "an internal billing tool"). Paraphrase my prompts rather than quoting them, except for short quotes that contain nothing identifying. Describe workplace limits by type only ("file uploads are blocked"); don't name or quote any employer policy, and don't record whether I followed or broke a workplace rule.
- **Use what you can see.** Look at whatever history you have access to: past sessions or transcripts, memory, project instruction files (such as CLAUDE.md), saved instructions, custom commands, skills, connected tools, and the projects themselves. Sample at least 10 recent sessions or conversations if available, across different kinds of work. Say exactly what you reviewed.

## What to assess

1. **Setup and environment.** What tools and access do you have here? Can you run code, read real files or data, test in a safe environment, remember past sessions, or act in other systems? What persistent instructions, memory, rules or custom tools exist, and how specific are they? Also note any workplace limits you can see (blocked tools, no access to real data, rules in my instructions about what's allowed). If you can't tell, say unknown.
2. **How I brief you.** Do I give goals, context, constraints, business rules and examples, or short one-line asks? Is my language mostly about outcomes and domain rules, or about technical implementation?
3. **How I check your work.** How often do I verify results against reality (running it, checking real data, testing it, reporting what I observed) versus accepting output as-is? Estimate a rate and give examples.
4. **How I push back.** How often do I question your reasoning, correct you, or challenge an assumption? What is the pushback usually about: business logic, quality, process, or code?
5. **How I learn from mistakes.** When you make a mistake, does it get turned into a lasting rule, instruction or checklist, or does the lesson disappear? Count the persistent rules that came from corrections, if you can.
6. **What I build.** What have we actually built or automated? Is it used repeatedly, by me or by others? Was any of it adopted or reused by other people? Rough categories and how effort splits across them.
7. **Planning and delegation.** Do I ask for a plan or approach before you start? How much do I decide versus leave to you? Do I use multiple agents, reviews or parallel sessions?
8. **Persistence.** Typical number of rounds per task. Do I keep iterating to a working result, or stop after the first answer?
9. **What hurts quality or wastes time.** Recurring dead ends, repeated corrections, context problems, or things I do that make your output worse.
10. **Fragility.** What would break or be lost if I stopped maintaining this? What is undocumented or depends only on me?

## Output

Write the report to a file named `ai-work-profile.md` (or, if you can't write files, output it as a single markdown block I can copy). This markdown version is the one I'll upload to the study: label it **"The version to upload"**. Use exactly this structure.

Start with these fields, filled in honestly. Use `unknown` when you can't tell.

```yaml
---
schema_version: 2
prompt_version: deep-1.4
report_date: YYYY-MM-DD
ai_tool: ""                     # e.g. Claude Code, Cursor, ChatGPT, Claude.ai
access_level: ""                # history_and_files | history_only | single_conversation
evidence_reviewed: ""           # e.g. "47 sessions over 5 weeks, config files, memory"
sessions_reviewed: 0
prompts_reviewed: 0
proficiency_level: ""           # novice | experimenter | practitioner | expert
builds_tools: ""                # yes | no
things_built_count: 0           # tools, automations or apps used more than once
primary_work_domain: ""         # e.g. operations, finance, marketing, software, research
user_is_professional_developer: ""   # yes | no | unknown
median_prompt_length_chars: 0
verification_rate: ""           # estimated share of tasks where the user checked output against reality, e.g. "~20%"
pushback_rate: ""               # estimated share of prompts that question or correct the AI
typical_rounds_per_task: ""
asks_for_plan_first: ""         # often | sometimes | rarely | never | unknown
persistent_instructions: ""     # none | basic | extensive
rules_from_corrections: 0       # persistent rules written after mistakes, if countable
env_can_run_code: ""            # yes | no
env_can_access_real_data: ""    # yes | no
env_has_safe_test_environment: ""  # yes | no
env_has_memory: ""              # yes | no
uses_multiple_agents_or_reviews: ""  # yes | no
assessment_confidence: ""       # low | medium | high
results_evidence: ""            # none | some | clear | unknown: evidence that work done with AI was actually used (sent, adopted, run), not just produced
outputs_used_by_others: ""      # yes | no | unknown: does anyone else reuse something built with AI?
workplace_limits: ""            # none_seen | some | severe | unknown: blocked tools, no access to real data, rules about what's allowed
---
```

Proficiency levels: **novice** (lacks the skills to get value from AI), **experimenter** (uses AI often but for one-off tasks), **practitioner** (AI embedded in repeatable workflows), **expert** (builds AI workflows, tools or automations for themselves and others).

Then:

### Part 1: For you

Written to me, in plain language, so I can act on it this week. The privacy rules above apply here too, and to every example: describe my work by type only ("a weekly client report", "an internal billing tool"). No names of people, employers, clients, products, projects or systems; no file paths, URLs, figures or internal terms; and don't copy my real instructions or data into an example. Keep the "Try this now" message generic: I'll add the specifics when I send it. Use these headings:

#### At a glance

- **Your level:** the same level as `proficiency_level` above, shown as the four levels on one line with mine in bold, like: Novice › **Experimenter** › Practitioner › Expert.
- **In one sentence:** the single most useful thing to know about how I work with AI.
- **To reach the next level:** the one shift that would move me up. If I'm already an expert, what would make me more effective.

#### What you do well

Up to five habits that most improve my results, only those with clear evidence; if there are fewer, list fewer. For each: the habit in one line, the evidence, and one tip to get even more from it.

#### What's holding you back

The two or three habits that most hurt quality or waste time. For each: the habit, the evidence, and what it costs me (time, rework or missed results). Be direct. Don't soften a finding to balance the strengths above.

#### Your next three moves

The changes with the biggest payoff (up to three), most impact first. For each:
1. **What to do,** in one or two sentences.
2. **Try it like this:** a concrete example I can copy (a better prompt, a saved instruction, a checklist or a setting), written for my kind of work. Write it as a quote block (lines starting with `>`), never as a code block.
3. **You'll know it's working when:** one observable sign.
4. **Effort:** minutes, an hour, or ongoing.

#### Try this now

A ready-to-paste follow-up message, as a quote block (each line starting with `>`), that I can send you right after this report to start working on move 1. Write it in my voice, mention the report, and ask you to coach me step by step on that move using a real task I describe to you. Keep it under 80 words.

### Part 2: For the research

Part 2 must be as candid as the Rules require, whatever tone Part 1 used. Report the evidence, including evidence against my level.


- One short section for each of the 10 areas above, with the evidence behind each finding (counts, rates and anonymized examples).
- **Results:** evidence that work done with AI was actually used (sent, adopted, run), and how strong that evidence is.
- **What a typical user of a chat-only AI tool would be missing** compared with this setup, if anything.
- **What I couldn't see:** what you didn't have access to, how far back your evidence goes, and which estimates are rough.

### Closing reminder

End the report with this line, exactly:

> Before uploading: read this report and delete anything you don't want to share, including names, employer or client details, and anything confidential.

Aim for 1,200 to 2,200 words: Part 1 about 500 to 800, Part 2 at least 600. Part 1 is what I'll read first, so make it skimmable.

### A copy that's easy to read

After the report, also give me a copy that's easy to read and keep: the same Part 1 and Part 2 text, word for word, without the summary fields at the top. The privacy rules above apply to this copy too; don't add any detail that isn't in the report. Make it:

- If you can create files, make a **PDF or Word document**.
- If you can't make those but can write files, make a **web page file** (`ai-work-profile.html`) that opens in any web browser.
- If you can't create files at all, show Part 1 as normal formatted text in this chat (not inside a code block), so I can read it easily.

Only say you made a file if you actually did and I can download it. If you're not sure, use the chat option.

Label the two clearly: the **markdown version is the one to upload** to the study; the readable copy is just for me.

Then tell me in one line where each file is (the markdown to upload, and the readable copy), and repeat the **Try this now** message in the chat so I can send it straight away.

I'm taking part in an independent research study about how people work with AI, and I want an honest assessment of how I work with you. You'll produce one markdown report. Part of it is for me to act on and part is for the research, but I'll upload the whole report to the researcher after I've reviewed it, so the privacy rules apply to all of it.

## Rules

- **Be honest, not flattering.** Base every finding on evidence. If the evidence is thin, say so. Don't inflate my level or my habits.
- **Protect privacy.** In the report, do not include: names of people, companies, clients or employers; addresses, emails or phone numbers; passwords, keys or credentials; financial, health or other sensitive personal details; proprietary or confidential business information. Describe things generically ("a hospital", "a colleague", "an internal report"). Paraphrase rather than quote, except for short quotes that contain nothing identifying. Describe workplace limits by type only ("file uploads are blocked"); don't name or quote any employer policy, and don't record whether I followed or broke a workplace rule.

## Step 1: Gather evidence

If you have memory of me or can search my past conversations, use them as evidence. Use only what's already available to you: don't run tools, open files or folders, or read code to find more. Just interview me.

**Don't list or describe what you found.** In one sentence, tell me only what kind of evidence you have (for example, "I can see your saved memory and some past chats" or "I can't see any history, so I'll ask you"). Don't mention specifics from my history in your questions either. Save the details for the report, which follows the privacy rules above. If I ask you not to use my history, rely only on my answers.

Then ask me the following, **one question at a time**, and wait for each answer. Ask every question, as written. Keep them short.

1. What does my workplace allow or block when it comes to AI (for example: can't upload files, can't use outside tools, no access to company data)? If nothing is blocked, or I'm self-employed, I can say so.
2. Briefly describe **three recent real tasks** I used AI for at work, a sentence or two each: what I asked, how it went, and what I did with the result. Remind me to leave out anything confidential.
3. For one of those tasks, did I check the output before using it? If so, how?
4. Have I made anything with AI that I reuse, like a template, saved instructions or an automation? It's fine if not. If yes, is it still in use?
5. Think of a recent time AI didn't help or made things worse. What happened?

If I give vague answers, ask one follow-up for a concrete example. Don't ask more than eight questions in total.

## Step 2: Assess

Using what you found and what I told you, assess:

1. **Environment:** what my tools let me do and what's blocked.
2. **How I brief AI:** goals, context, constraints and examples, or short one-line asks?
3. **How I check the work:** do I verify against reality, or accept output as-is?
4. **How I push back:** do I question reasoning or correct the AI?
5. **How I learn from mistakes:** do lessons get saved and reused?
6. **What I build:** anything reused by me or others?
7. **Persistence:** do I iterate to a good result or stop at the first answer?
8. **What hurts quality or wastes time.**

## Step 3: Write the report

Output the report as a single markdown block I can copy into a file named `ai-work-profile.md`, and label it **"The version to upload"**. Use exactly this structure. Inside the report, don't use any code blocks (```); use quote blocks for examples.

Start with these fields, filled in honestly. Use `unknown` when you can't tell.

```yaml
---
schema_version: 2
prompt_version: light-1.5
report_date: YYYY-MM-DD
ai_tool: ""                     # e.g. Microsoft Copilot, Gemini, ChatGPT, Claude.ai
access_level: ""                # history_only | single_conversation
evidence_reviewed: ""           # e.g. "user's description of 3 tasks; no history available"
sessions_reviewed: 0
prompts_reviewed: 0
proficiency_level: ""           # novice | experimenter | practitioner | expert
builds_tools: ""                # yes | no
things_built_count: 0           # tools, automations or apps used more than once
primary_work_domain: ""
user_is_professional_developer: ""   # yes | no | unknown
median_prompt_length_chars: unknown
verification_rate: ""           # with fewer than 10 examples, write unknown and describe the pattern in Part 2
pushback_rate: ""               # with fewer than 10 examples, write unknown
typical_rounds_per_task: ""
asks_for_plan_first: ""         # often | sometimes | rarely | never | unknown
persistent_instructions: ""     # none | basic | extensive
rules_from_corrections: 0
env_can_run_code: ""            # yes | no
env_can_access_real_data: ""    # yes | no
env_has_safe_test_environment: ""  # yes | no
env_has_memory: ""              # yes | no
uses_multiple_agents_or_reviews: ""  # yes | no
assessment_confidence: ""       # low | medium | high
results_evidence: ""            # none | some | clear | unknown: evidence that work done with AI was actually used (sent, adopted, run), not just produced
outputs_used_by_others: ""      # yes | no | unknown: does anyone else reuse something built with AI?
workplace_limits: ""            # none_seen | some | severe | unknown: blocked tools, no access to real data, rules about what's allowed
---
```

Proficiency levels: **novice** (lacks the skills to get value from AI), **experimenter** (uses AI often but for one-off tasks), **practitioner** (AI embedded in repeatable workflows), **expert** (builds AI workflows, tools or automations for themselves and others).

Then:

### Part 1: For you

Written to me, in plain language, so I can act on it this week. The privacy rules above apply here too, and to every example: describe my work by type only ("a weekly client report", "an internal billing tool"). No names of people, employers, clients, products, projects or systems; no file paths, URLs, figures or internal terms; and don't copy my real instructions or data into an example. Keep the "Try this now" message generic: I'll add the specifics when I send it. Use these headings:

#### At a glance

- **Your level:** the same level as `proficiency_level` above, shown as the four levels on one line with mine in bold, like: Novice › **Experimenter** › Practitioner › Expert.
- **In one sentence:** the single most useful thing to know about how I work with AI.
- **To reach the next level:** the one shift that would move me up. If I'm already an expert, what would make me more effective.

#### What you do well

Up to five habits that most improve my results, only those with clear evidence; if there are fewer, list fewer. For each: the habit in one line, the evidence, and one tip to get even more from it.

#### What's holding you back

The two or three habits (or limits of my workplace tools, if those are the main problem: say so plainly) that most hurt quality or waste time. For each: the habit, the evidence, and what it costs me (time, rework or missed results). Be direct. Don't soften a finding to balance the strengths above.

#### Your next three moves

The changes with the biggest payoff (up to three), most impact first. For each:
1. **What to do,** in one or two sentences.
2. **Try it like this:** a concrete example I can copy (a better prompt, a saved instruction, a checklist or a setting), written for my kind of work. Write it as a quote block (lines starting with `>`), never as a code block.
3. **You'll know it's working when:** one observable sign.
4. **Effort:** minutes, an hour, or ongoing.

#### Try this now

A ready-to-paste follow-up message, as a quote block (each line starting with `>`), that I can send you right after this report to start working on move 1. Write it in my voice, mention the report, and ask you to coach me step by step on that move using a real task I describe to you. Keep it under 80 words.

### Part 2: For the research

Part 2 must be as candid as the Rules require, whatever tone Part 1 used. Report the evidence, including evidence against my level.


- One short section for each of the 8 areas above, with the evidence behind each finding.
- **Results:** evidence that work done with AI was actually used (sent, adopted, run), and how strong that evidence is.
- **What I couldn't see:** say what this assessment is based on (my history, my answers, or both) and which findings are least certain.

### Closing reminder

End the report with this line, exactly:

> Before uploading: read this report and delete anything you don't want to share, including names, employer or client details, and anything confidential.

Aim for 800 to 1,400 words: Part 1 about 400 to 600, Part 2 at least 400. Part 1 is what I'll read first, so make it skimmable.

### A copy that's easy to read

After the report, also give me a copy that's easy to read and keep: the same Part 1 and Part 2 text, word for word, without the summary fields at the top. The privacy rules above apply to this copy too; don't add any detail that isn't in the report. Make it:

- If you can create files, make a **PDF or Word document**.
- If you can't make those but can write files, make a **web page file** (`ai-work-profile.html`) that opens in any web browser.
- If you can't create files at all, show Part 1 as normal formatted text in this chat (not inside a code block), so I can read it easily.

Only say you made a file if you actually did and I can download it. If you're not sure, use the chat option.

Label the two clearly: the **markdown version is the one to upload** to the study; the readable copy is just for me.