Distillation isn’t a form or a wizard API — it’s a conversation. But it’s a conversation with hard phases and gates: research can’t be skipped, scenario calibration can’t be skipped, and at least one skill must be produced. These are written into the distillation contract, not offered as advice.
What it actually writes
The distiller uses thedistill_manage tool to write 12 fields on your new expert:
Five cognitive layers land in three content fields, not one field each:
The full pipeline
Distillation creates aDRAFT expert up front (identifier shaped like distilled-xxxxxxxx, version 0.0.1), and every write lands on that row. So an interrupted session isn’t wasted — the draft stays in your works list.
1
Phase 0 · Entry routing
A clear target (“distill Munger”, “build a Feynman expert”) takes the direct path; a vague need (“I want to make better decisions”) takes the diagnostic path, which narrows over 1-2 rounds and then offers 2-3 candidates to choose from.
2
Phase 0A/0B · Clarification and materials
Confirms identity, focus, and purpose, and asks whether you have local materials.If you say you have materials but haven’t uploaded them, it must stop and wait. It won’t run research first — that’s a hard rule, and it exists so the expert isn’t built on a public-information framework you never endorsed.
3
Phase 1 · Six-dimension research
Research cannot be skipped. Even when the target isn’t a specific person (e.g. “design a general methodology”), Phase 1 still runs.
4
Phase 1.5 · Research review
Pauses, presents a research-quality summary, waits for your OK. Filling gaps here is far cheaper than reworking after the persona is written.
5
Phase 2 · Synthesis and write
Extracts 3-7 mental models and 5-10 decision heuristics into
persona; boundaries and tensions into soul; expression style into agent-instructions; then generates the greeting, name, description, and HTML overview.6
Phase 2.7a · Configuration assembly
Analyzes whether the domain needs external services (MCP connections like GitHub or Jira).This step is mandatory but may produce nothing — if none are needed, none are written. Any API key it writes is a placeholder; you fill in real credentials in Studio afterwards.Tool allowlists/denylists are deliberately left empty: all builtin tools are available by default, and casually restricting them switches off capabilities you’d expect to have.
tools-deny is only written when you explicitly ask for restrictions.7
Phase 2.7b · Skill generation
Every distillation produces at least 1 skill — this is a hard requirement. The type follows the domain:
8
Phase 3 · Scenario calibration
Mandatory, and only starts after you explicitly say to continue. It sends 5 test scenarios in a single
Ask, covering five cases:- A known question (does it answer correctly)
- A tradeoff (does it judge like you)
- Voice (does it sound like you)
- Boundary (does it know what it doesn’t know)
- Out of scope (does it force an answer anyway)
This step calibrates the expert behavior that was already written, not your input. So even an exhaustively detailed brief doesn’t let it skip — being clear yourself doesn’t mean it wrote the right thing.
9
Phase 4-5 · Verification and close
Checks that fields are filled (with character counts), that mental models land in the 3-7 range, that honesty boundaries are explicit, and that expression identity holds. Then it tells you to go publish in Studio.
Skill naming traps
Skills with scripts must run successfully in the sandbox (exit code 0) before upload, with at most 3 fix attempts before falling back to a Markdown-only skill. So “the distiller said the skill works but it errors out” shouldn’t happen on the normal path — if it does, the fallback didn’t fire and it’s worth reporting.
Four packages are blocked by the platform:
python-pptx, reportlab, fpdf, pptxgenjs. Skills needing PPT or PDF output must delegate to builtin plugins (builtin/presentation, builtin/kami, …) instead of installing their own libraries.How to talk so something comes out
Bring 3-5 real cases
Bring 3-5 real cases
Concrete cases beat abstract description by a wide margin. Abstractions extract generic platitudes; cases contain the judgment rules you don’t know you’re using.
Expose the reasoning chain, not just the conclusion
Expose the reasoning chain, not just the conclusion
Not “I’d recommend X” but “I look at A and B first; if A meets C I pick X, otherwise I consider Y.” The first yields one conclusion; the second yields a decision heuristic.
Say plainly what you don't know
Say plainly what you don't know
Honesty boundaries go into
soul and directly determine whether the expert bluffs. This is the quality difference buyers notice fastest — an expert that answers everything stops being believed after the second try.Give counter-examples
Give counter-examples
“The common mistake is…” / “If someone tells you X, they’re usually wrong.” Anti-pattern knowledge separates novices from experts, and it’s the part of
persona that public sources can least fill in.Actually read the research summary at Phase 1.5
Actually read the research summary at Phase 1.5
This is the cheapest correction point in the whole pipeline. Let a wrong research direction through and the entire persona has to be redone.
Failure modes and troubleshooting
After distillation
Distillation produces a working first version, not a final one. Next:- Chat with it in Studio and check whether it answers like you
- Hand-tune
persona/soul/agent-instructions— what distillation wrote are ordinary fields you can edit any time - Install more skills from the marketplace
- Set pricing and submit for review
You can distill repeatedly. After publishing, the self-evolution system also proposes improvements from real usage, which you accept or ignore.
Related
End-to-end: distill an expert
What a complete distillation actually looks like
Configure your expert
The four prompt layers and every config field
Self-evolution
Continuous improvement after publishing
Pricing
How unlock fees relate to conversation costs