> ## Documentation Index
> Fetch the complete documentation index at: https://docs.profy.cn/llms.txt
> Use this file to discover all available pages before exploring further.
# End-to-end: distill an expert
> A complete distillation run: six-dimension research, field writes, scenario calibration, and review submission
This walks through a full distillation, from "I want to build an X expert" to an expert ready for review.
One thing to know first: **distillation is a conversation, but not a casual one**. It has numbered phases and hard gates — research can't be skipped, scenario calibration can't be skipped, and at least one skill must be produced. When it feels like "why isn't it writing yet", it's usually not stuck; it's waiting for you to confirm a checkpoint.
## What you get
A `DRAFT` expert with 12 fields written, at least 1 skill installed, calibrated against 5 scenarios, and ready to submit for review.
## Prerequisites
Creator onboarding must be approved first. Unapproved accounts are rejected by the backend on submit (the frontend blocks it too).
Both "distill Munger" and "I want to make better decisions" work, but they take different paths: the first goes straight to clarification, the second spends 1-2 rounds on diagnosis. The more specific your target, the more time you save.
Your articles, course notes, case records, existing agent config files all work. **Have them ready before starting** — the moment you say "I have materials", the distiller stops and waits for the upload, and won't proceed without it.
## Steps
### Step 1: Open a distillation conversation
Create a new expert in Studio and choose distillation. A `DRAFT` expert is created immediately (identifier shaped like `distilled-a1b2c3d4`, version `0.0.1`), and all writes land on it. Closing the tab mid-way loses nothing — the draft is in your works list.
Open with a clear target:
```text theme={null}
I want to build an "indie developer pricing advisor" expert.
Audience: indie devs shipping SaaS or tools
Problem: pricing by gut feel, afraid to raise prices, no idea how to tier
I've shipped three paid products and have real repricing data. I'll upload materials shortly.
```
That last clause matters. Say you have materials and it will wait, rather than running public research first and bringing you a framework you don't endorse.
### Step 2: Upload materials
Send your stuff in. Documents, PDFs, Markdown, exported conversation logs all work.
```text theme={null}
Materials uploaded:
- Full records and conversion data from three price changes
- 5 chat logs from pricing consultations I've done
- My essay "Why per-seat pricing is wrong for small teams"
```
Then say explicitly "materials are complete, start research." It needs a clear go signal.
### Step 3: Wait for six-dimension research, then actually read the summary
Research covers six dimensions: works, dialogues, expressions, third-party views, major decisions, and timeline. With your materials present, they lead and public sources supplement.
When it finishes, it **stops** and hands you a quality summary, waiting for your OK.
This is the cheapest correction point in the pipeline. Let it pass and everything written afterwards rests on the wrong direction — reworking means starting over.
Read line by line for two things: are the core arguments it extracted actually yours, and are the sources it cites ones you'd stand behind.
If it's wrong, say so:
```text theme={null}
Item 3 is wrong. I have never argued "go low to acquire, then raise prices."
My position is the opposite: users acquired on low prices churn hardest when
prices go up, and I have the data. Re-read the first file I uploaded.
```
### Step 4: Watch it write the fields
After you confirm, it synthesizes and writes. You're mostly a spectator here, but it helps to know where things land:
| What you said | Field |
| --------------------------------------------------------------------- | -------------------- |
| How you understand the domain, your judgment rules, what you never do | `persona` |
| What you don't know, your internal tensions, your behavioral floor | `soul` |
| How you talk, how you respond, what you won't say | `agent-instructions` |
| How it greets people | `opening-message` |
| The marketplace detail page | `overview` (HTML) |
Mental models are capped at 3-7 and decision heuristics at 5-10. More isn't better — past 7 mental models means no choices were made, and the expert comes across as having no position in real conversations.
### Step 5: Skill generation
Every distillation produces at least one skill. A skill is a **reusable procedure**, not your opinions — those are already in `persona`.
For the pricing advisor, a sensible skill is a "pricing plan evaluation" procedure: given product shape, user scale, and competitor pricing, output tier recommendations and risk flags.
```text theme={null}
Make one skill: "pricing tier evaluation". Inputs are product type, target user
scale, and competitor price range. Output three pricing tiers plus a churn-risk
note per tier. No scripts needed — a methodology skill is fine.
```
Skill names must be English kebab-case (e.g. `pricing-tier-evaluator`). Non-ASCII names make the skill permanently un-activatable — that's a hard constraint, not a style preference.
### Step 6: Scenario calibration
It will stop and ask whether to begin calibration. **Reply "continue"**, and it sends 5 test scenarios at once:
1. A known question inside your expertise — is the answer right
2. A tradeoff question — does it choose like you
3. A voice test — does it sound like you
4. A boundary question — does it know what it doesn't know
5. An out-of-scope question — does it force an answer anyway
Every question has clickable options, and you can always write your own.
Questions 4 and 5 deserve the most care. Buyer trust collapses almost exclusively at these two points: an expert that answers everything stops being believed after the second try.
If it produces a plausible-sounding answer to the out-of-scope question, correct it explicitly: "you should say you don't know here and suggest talking to an accountant." That sentence goes into `soul`.
### Step 7: Verification and submission
It checks field fill state and character counts, mental model count, and whether honesty boundaries are explicit, then tells you to publish from Studio.
**Publishing does not happen in the conversation.** The distiller has no publish permission. Its closing "click Publish in the top right" is the end of the pipeline, not a brush-off.
Three things worth doing yourself before submitting:
* Chat with it in Studio for five minutes, asking questions your real customers would ask
* Check that the `persona` content is all still there — every write is a full-field overwrite, so if it was revised repeatedly mid-conversation, give it a scan
* If MCP connections were configured, replace the placeholder API keys with real credentials
## Common sticking points
| Symptom | Cause | What to do |
| ----------------------------------------------------- | ---------------------------------------------- | ----------------------------------------------------- |
| It sits still, only asking you to upload | You mentioned materials but didn't send them | Upload, or say "proceed without materials" |
| It starts writing without researching | Phase 1 was skipped | Say "research first, then write" |
| It declares done without calibration | Phase 3 was skipped | Say "run the scenario calibration" |
| Calibration questions have no options, only free text | It omitted the options array | Ask it to re-send with options |
| Skill scripts invisible in Studio | It used `create` instead of `save_dir` | Ask it to re-save the whole directory with `save_dir` |
| Two similarly named skills appear | The save used a name that didn't match | Delete the duplicate |
| Earlier `persona` content disappeared | Full-field overwrite without a read-back merge | Restore manually in Studio; batch future edits |
| Submission rejected | The account isn't an approved creator yet | Complete creator onboarding first |
## One test for quality
Whether a distillation went well isn't measured by how full the fields are. It's this: **put it next to a general-purpose model answering the same domain question — can you tell them apart at a glance?**
If you can't, the distillation produced generic platitudes, usually because the conversation got abstract descriptions instead of real cases. Going back with 3-5 concrete cases and running it again beats rewording the fields by a wide margin.
## Related
The 12 fields, phase gates, and skill naming rules
The four prompt layers and every config field
Onboarding requirements and levels
How unlock fees relate to conversation costs