Session Improvement Compiler

Session Improvement Compiler

Core Principle

Compile evidence, not anecdotes. Raw session events and current source establish truth; Fabric recall locates evidence, while Hindsight recall/reflect proposes patterns. Every durable promotion must remain traceable to authoritative evidence.

When to Use / NOT

• Use when: the user explicitly requests continual improvement, session-to-skill compilation, or improvement from Fabric/Hindsight memory.
• NOT when: merely recalling or reflecting, at every session end, or when no promotion was requested.

Workflow

1. Bound the evidence. Default to the current project and current session. Record named sessions, user-provided transcripts, and exact questions. Cross-project/global evidence requires explicit scope.
2. Triangulate cheaply. Read the smallest sufficient raw event ranges. Use Fabric recall to locate ranges and Hindsight recall/reflect to surface candidate patterns. Projection output is a lead, never independent proof. If a projection provider fails, continue from accessible raw evidence rather than inventing its conclusion.
3. Build an evidence ledger. For each candidate, record supporting events/tool outputs, counter-evidence, recurrence, current-source status, confidence, and scope. Preserve user corrections and failed rationalizations verbatim where useful (references/evidence-contract.md).
4. Classify before writing. Apply leverage-capture: CODE, GATE, SKILL, PROJECT NOTE, FOUNDATION, or NOT WORTH SAVING. Objective regression classes prefer deterministic gates. Repeated operational judgment may earn a skill. Project facts do not become global skills.
5. Find the owner. Search existing skills, prompts, tests, validators, and docs. Improve the narrowest existing owner before creating another skill. Reject additions that create trigger collisions or duplicate truth.
6. Promote deliberately. This invocation is an explicit promotion request. Carry forward confirmed, in-scope defects from earlier passes: fixing an easier neighbor or documenting the gap does not dispose of them. Implement each qualified repair or name a concrete blocker (missing authority, dependency, evidence, or unsafe scope); do not silently reclassify unfinished work as nothing worth saving. Follow writing-skills for skill changes; preserve canonical prompt/skill separation; keep raw transcripts and private identifiers out of operational instructions.
7. Verify end to end. Run affected validators, catalog/collision checks, deterministic gates, prompt installation checks, and behavior pressure tests. RED may come from the selected real session; GREEN must demonstrate that the revised workflow rejects the same rationalization twice.
8. Report disposition. Reconcile the report against every ledger candidate, including prior deferrals; a previous completion report is not closure evidence. List evidence used, files changed, checks, unresolved uncertainty, and material deliberately not saved. Stop when promotions are verified or no candidate qualifies.

Red Flags

• Treating Hindsight or a summary as primary evidence.
• Dumping entire session logs into model context or skill files.
• Creating a new skill before searching for an existing owner.
• Encoding deterministic requirements only as prose.
• Promoting a one-off preference, stale fact, or project inventory.
• Saying “learned” without a verified behavioral or mechanical change.

Verification

The skill validator and catalog pass; prompt references resolve; objective gates catch the observed regression; behavior tests pass twice; every promotion cites authoritative evidence; the not-saved list is explicit.

References

• references/evidence-contract.md, source authority, ledger, classification, and privacy rules.
• references/behavior-tests.md, real RED cases and GREEN acceptance.
• ../leverage-capture/SKILL.md, cheapest durable representation.
• ../writing-skills/SKILL.md, skill authoring and validation.