> ## 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
> From a name to a publishable expert, through six phases and four quality gates
This walks the full distillation: give it a person or a topic, end up with an expert draft that has every core field filled, at least one skill, and passed scenario calibration — ready to enter the publish flow.
First, correct the most common misconception: **distillation is not copying what someone said.** Copying phrasing gives you a bot that recites quotes and falls apart the moment you ask something it hasn't memorized. What distillation extracts is a **thinking framework** — which mental models they view problems through, which heuristics they judge by, what they categorically won't do, and an honest account of what they can't do. The former is WHAT they said; the latter is HOW they think. Only the latter generalizes to questions it has never seen.
## Prerequisites
The distillation studio is creator-only. Without approval you're redirected back to Studio. See [Become a creator](/en/creators/become-a-creator).
"Distill Munger" and "build a Munger-perspective advisor for reviewing my investment decisions" produce very different quality. The latter gives distillation a basis for making trade-offs.
Have primary sources ready if you can (manuscripts, transcripts, your own long-form writing). Material quality sets the ceiling — public search alone can produce a usable expert, but rarely a distinctive one.
## The key mechanic: distillation writes these 12 fields
Every "write" in a distillation conversation lands content in one of the target expert's fields. There are exactly 12 writable ones:
| Field | Carries |
| --------------------------------------------- | --------------------------------------------------------------------------------------- |
| `persona` | Role definition, core mental models, decision heuristics, value ordering, anti-patterns |
| `soul` | Internal tensions, honest boundaries, expressive tendencies |
| `agent-instructions` | Workflow, behavioral rules, output format |
| `opening-message` | Opening line |
| `overview` | Marketplace overview |
| `name` / `description` / `category` | Basic information |
| `tool-configs` / `tools-allow` / `tools-deny` | Tool configuration and allow/deny lists |
| `guard` | Guardrail rules |
Skills aren't among the 12 — they're separate objects created with a different tool and can carry scripts and reference material.
## Six phases
### Phase 0: Entry routing
Open the studio and state your intent. Two paths follow:
| What you say | Path |
| ------------------------------------------------ | -------------------------------------------------------- |
| A concrete name or topic ("distill Munger") | Direct: confirm understanding → ask for materials |
| A vague need ("I want to make better decisions") | Diagnostic: unpack the need first, then choose a subject |
Either path **asks separately whether you have materials**. Answer carefully:
* **You have them** → say so, send everything, then say "that's all." It will stop and wait rather than searching first.
* **Archives** → sending a zip is fine; it unpacks before reading. Do **not** expect it to read inside a zip without unpacking.
* **You don't** → say so and it proceeds to search.
### Phase 1: Multi-source collection
Six dimensions: written work, conversations (podcasts/interviews), expression (social media), outside views (analysis and criticism), decisions (major calls and turning points), and timeline.
Three hard requirements:
1. Note source credibility — primary beats secondary beats inference
2. Distinguish "what they said," "what others said about them," and "what I inferred"
3. **Keep contradictions when you find them** — don't split the difference
Sources that are mostly second-hand commentary are excluded, because someone else's interpretation gets mixed in and read back as the subject's own view.
Collection ends with a research-quality summary and a pause. **This is the first checkpoint.** The right move here isn't "continue" — it's checking two things: is the primary-source share adequate (target >50%), and is any dimension essentially empty. Waving this through means everything synthesized afterwards rests on nothing.
### Phase 2: Framework synthesis
The core of distillation. Output is written into fields incrementally:
Filtered down from 15-30 candidate claims against three tests: **cross-domain recurrence** (they apply it in unrelated fields), **generative power** (it produces new conclusions rather than describing), and **exclusivity** (others don't think this way).
Each model must carry: name, one-line description, ≥2 source situations, how it's applied, and **its limits**. That last one is the most-skipped and the most important — it's what separates a thinking model from a universal platitude. A model with no stated failure conditions is usually a truism.
Shaped as "if X, then Y," each with a real case. These are the intuitive rules the expert applies to novel problems.
At least 2 pairs of internal contradictions (two things they believe that pull against each other) and at least 3 concrete limits (explicitly saying "I'm weak here" or "I may be biased here").
Honest boundaries aren't a disclaimer — they're what makes the expert say "I don't know" instead of fabricating when it's out of its depth. Experts missing this fail most visibly on edge questions.
Frequent words, sentence shapes, analogies, humor. The acceptance test is concrete: **100 words should be enough to recognize who it is.**
Things they categorically don't do. Negative constraints lock in voice more effectively than positive description.
### Phase 2.7: Configuration assembly and skill generation
After the framework, tools are configured and at least one skill is generated.
Skills are what let the expert actually do work (a deep-research procedure, a decision-review template) rather than only talk. Two hard rules:
* **Skill names must be English kebab-case** (`decision-analysis`, `data-pipeline`). Non-English names make the skill uncallable, and saving creates duplicates instead of overwriting.
* **Don't turn the cognitive layer into skills.** Mental models, decision heuristics, expression DNA, anti-patterns, and honest boundaries are content fields. Making them skills means they only apply when activated, instead of always.
If a skill carries scripts or reference material (`scripts/`, `references/` directories), it must be saved as a whole directory — otherwise only the skill description text is stored and the scripts are lost.
### Phase 3: Scenario calibration (mandatory)
Once core fields are written and at least one skill exists, **5 test scenarios** are generated for you to confirm, covering five dimensions:
| Dimension | Tests |
| ------------ | ----------------------------------------------------------- |
| Known domain | Whether it answers in-scope questions correctly |
| Trade-offs | Whether its preference on dilemmas matches yours |
| Voice | Whether it sounds right |
| Boundary | How it behaves near the edge of competence |
| Out of scope | Whether it admits ignorance on things it truly doesn't know |
Each scenario offers three options: matches / not quite / skip. **When you pick "not quite," say exactly what's wrong** — it uses that to revise the corresponding field. Clicking "matches" through all five wastes the step, and the fifth scenario deserves the most attention: an expert unwilling to say "I don't know" will produce confident, professional-looking wrong answers forever after launch.
### Phase 4: Quality verification
An automatic pass against the checklist:
| Check | Passing standard |
| --------------------- | ---------------------------------------------------------------------------------- |
| persona role section | Opens with role definition, starts with "you are," names the expert and background |
| Mental model count | 3-7, each with source evidence |
| Limits per model | Explicit failure conditions written out |
| Voice recognizability | 100 words identify the subject |
| Honest boundaries | ≥3 concrete limits |
| Internal tensions | ≥2 contradictions |
| Primary-source share | >50% |
| Skill count | ≥1, successfully created |
Failing sends it back to Phase 2 with the weak areas flagged, **for at most two loops**. Failing twice usually means the material is thin — supply primary sources rather than letting it keep searching.
### Phase 5: Finish and publish
**Publishing does not happen in the conversation.** When distillation finishes it tells you to click the Publish button in the top right, which takes you to the expert editing page. There you review the content, fill in pricing and other details, and go through the formal publish review.
Saying "publish it for me" in the conversation has no effect.
## What a distilled expert looks like
A draft expert is created automatically during distillation, with:
| Property | Value |
| ---------- | ------------------------------------------------ |
| Status | Draft (not listed) |
| Version | `0.0.1` |
| Identifier | Auto-generated, shaped like `distilled-xxxxxxxx` |
| Source | Marked as distillation-created |
The draft is always available from Studio, and distillation conversations can be interrupted and resumed — re-entering checks the current field status and continues from where it stopped rather than starting over.
## Updating an existing expert
Say "there's news about X, update it" for an already-distilled expert and it will:
1. Read the existing content first
2. Research only what's new
3. Update the relevant fields incrementally, **without rewriting the whole profile**
That's deliberate — a full rewrite would wipe out the calibration you already did.
## Migrating from other platforms
Upload project files from another AI platform (recognizable by markers like `CLAUDE.md`, `.cursor/rules/`, `openclaw.json`) and content is routed by meaning:
| Original content | Lands in |
| --------------------------------------------------------- | ------------------------------ |
| Identity, capability description | `persona` |
| Personality, communication style | `soul` or `agent-instructions` |
| Workflow, behavioral rules, output format | `agent-instructions` |
| Skill definitions | Created as separate skills |
| Platform-specific mechanics (heartbeats, session startup) | Discarded |
Migrated content **still needs scenario calibration**. A prompt tuned on another platform won't necessarily behave the same under this runtime.
## Boundaries and failure modes
| Symptom | Cause | Fix |
| -------------------------------------------- | ------------------------------------ | ---------------------------------------------------------------------------------------- |
| Can't open the distillation studio | Creator status not approved | Complete the creator application first |
| It searched without waiting for my materials | You didn't say you had materials | Interrupt and say you have material to send |
| Content inside the zip wasn't read | Archives must be unpacked first | Explicitly ask it to unpack and read each file |
| Skill created but not callable | Skill name isn't in English | Recreate with an English kebab-case name |
| Scripts missing from a skill | Saved as plain text | Save as a whole directory instead |
| Quality verification keeps failing | Insufficient material | Add primary sources rather than searching more |
| Said publish, nothing happened | Publishing isn't in the conversation | Click Publish to go to the editing page |
| Output reads like generic platitudes | Mental models lack exclusivity | Require evidence that others don't think this way; delete the ones that can't produce it |
## Related pages
Distillation overview
Four-layer prompt injection order
What to set before publishing
How the expert keeps improving after launch