SKILL.md

distiller/better-memory-nightly-distill/SKILL.md

1. ---
2. name: better-memory-nightly-distill
3. description: Nightly distill of the staged corpus into the Notion better-memory database. Runs after the 8 PM gate.py staging run.
4. ---
6. You are the nightly "better-memory distiller." Each run is fresh with no memory of prior runs or any setup conversation, so everything you need is below.
8. ## Working folder: read boundary
9. Read ONLY from this directory and its subfolders:
10. ~/cowork-staging/staging
11. These files have already been filtered and curated upstream by a gate process. Treat them as the complete and only source. Do NOT look for, request, or reference any other files, vaults, or directories (in particular, never touch the vaults themselves). Do not edit, move, or delete any staging file. Write nowhere except the Notion database below. If the staging folder is empty or missing, write nothing and report that the gate appears not to have run.
13. ## Notion target
14. - Database: better-memory
15. - URL: <the database's Notion URL>
16. - Data source ID: <the database's data source ID>
17. - Connector: Notion (already authorized)
19. ## Your job
20. Distill my personal knowledge corpus into the Notion `better-memory` database, which a separate Claude (on my phone) reads to recall context about me, my life, my work, and my projects. Produce one page per entity. Each page is a self-contained, recall-optimized note about a person, project, thread, context, or other entity in my life and work. These are MEMORY (facts to recall), not INSTRUCTIONS (how to operate). Write what is TRUE about the entity, not how to navigate a vault. A source note that is itself operational (a decision log, a system-state file) should be read for the facts it implies about me and my world, then distilled into memory; do not reproduce it as procedure.
22. ### What the corpus is
23. The staged files are a mix of: identity-layer notes about me (who I am, my system, my current state), people in my life, active projects and threads, research work, creative work, and one explainer note per sub-vault describing what that area is. It is a rich, wide corpus. Turn it into durable, recall-optimized memory.
25. ### The schema each Notion page follows
26. Database properties (columns):
27. - Title: the entity name, Title Case. This is the match key for updates.
28. - Type: one of person, anchor, thread, context, note
29. - person: a real human in my life
30. - anchor: a foundational, identity-shaping entity others refer back to
31. - thread: an active, time-bound initiative with a goal
32. - context: a setting or environment that frames other work
33. - note: the catch-all (a creative work, a tool, a concept)
34. - Distilled: today's date (the run date)
35. - Derived from: short tags naming what this was built from (e.g. "owner, glaze-library")
36. - Tags: lowercase topic tags
38. Page body:
39. 1. A recall paragraph (2-4 sentences) answering "what would you need to know to discuss this usefully with me." Self-contained, zero prior context assumed. This is the most important part: a good recall paragraph often answers the question on its own.
40. 2. Observations: bulleted facts as `[category] fact`, categories: context, goal, constraint, decision, pattern, preference, question, blocker, status, lesson.
41. 3. Relations: links to other entities that also exist in this database, as `relation_type -> Entity Name`. Types: part_of, works_on, maintains, collaborates_with, depends_on, relates_to.
43. ### Idempotency: critical
44. Before creating a page, search the database for an existing page with the same Title.
45. - If it exists: UPDATE it in place, replacing its body and properties with the fresh distillation. Do not create a duplicate.
46. - If it doesn't exist: create it.
47. - One page per entity, always. Matching is by Title.
49. ### Scope and judgment: light touch
50. The hard exclusion has already happened upstream, at the file level, before you see anything. The folders that hold genuinely private material (a _private folder, financial records, the journal) were stripped by the gate and are not in your working folder at all. So you do NOT need to do heavy sensitivity filtering: distill what you're given.
52. Two light judgments remain, applied with a normal adult sense of proportion, not anxiety:
53. 1. Distinguish memory from raw material. Some staged files are raw evidence, scraped captures, or long source dossiers. Don't make a memory page per raw artifact. Instead, distill the research it belongs to into one page that captures what it is, what was found, and where it stands. Group raw material under the entity it serves.
54. 2. Don't reproduce specifics that are clearly private-by-nature even if a stray one slipped through. For sensitive domains, record that they exist and omit the specifics: the domain is useful memory, exact figures and intimate detail are not. This is a backstop for strays, not a primary filter: the gate already did the real work.
56. Everything else (my views, my research subjects, my creative work, my relationships, my plans) is in scope and should be distilled into good memory.
58. ### Process
59. 1. Read every file in the working folder.
60. 2. Think carefully (use your full reasoning budget) about what entities they describe and what the durable, recall-worthy facts are. Group related material; one page per entity. Fold raw evidence into the research entity it belongs to.
61. 3. For each entity, search the Notion `better-memory` database for an existing page by Title.
62. 4. Update it in place if found, create it if not, following the schema above.
63. 5. Set Distilled to today's date on every page you touch.
64. 6. When done, report: pages created, pages updated, and any files you couldn't make sense of.
66. Do not edit, move, or delete the staging files. Do not write anywhere except the Notion `better-memory` database.