Session Compounding Workflow
Implements PROTO-ORG-5 from
organon/protocols/PROTOCOLS.md. Converts session learnings into durable improvements.
When to Use This Skill
Use this skill when:
- •Ending a significant work session — substantial changes were made
- •After completing a multi-step task — new patterns may have emerged
- •After encountering friction — something was harder than it should have been
- •Periodically — even smooth sessions may reveal optimization opportunities
Purpose: The methodology warns: "Without explicit time allocation, improvement never happens." This workflow makes the COMPOUND step actionable.
Context Loading
- •Load project constraints:
- •Read
CLAUDE.md(project-level guidance and decision heuristics)
- •Read
- •Load pattern references:
- •Read
book-llms/patterns.md(Recursive Collaboration section — the pattern this workflow implements) - •Read
book-llms/workflow-authoring.md(workflow quality attributes for evaluating workflow improvements)
- •Read
Steps
Step 1: Review Session Work
Examine what happened during the session:
git diff --stat HEAD~N # where N = number of commits in this session git log --oneline -N # recent commits
List:
- •Files created or modified
- •Types of changes (organon content, tooling, skills, documentation)
- •Patterns in the work (what was repeated, what was manual)
Step 2: Detect Patterns
Look for these improvement signals:
| Signal | Example | Indicates |
|---|---|---|
| Repeated manual steps | "I ran organon validate then organon verify then organon health every time" | Tool candidate (composite command) |
| Unclear workflow | "I wasn't sure which skill to use for this task" | Workflow gap or unclear naming |
| New heuristic discovered | "I learned that scopes.md always needs updating after ETHOS.md changes" | Heuristic addition to CLAUDE.md or ETHOS.md |
| Awkward tool usage | "I had to run the same command 3 different ways to get what I needed" | Tool improvement (better defaults, new flag) |
| Missing context | "I needed to read 5 files before starting but the skill only listed 3" | Workflow context update |
| Terminology confusion | "I kept saying 'skill' when I meant 'workflow'" | Terminology cleanup needed |
| Error without guidance | "The tool failed but I didn't know how to fix it" | Error recovery table addition |
Step 3: Classify Improvements
For each finding, categorize it:
| Category | Description | Where to Implement |
|---|---|---|
| Tool candidate | Repeated operation that could be automated | packages/tools/ — new command or flag |
| Protocol update | Procedure followed but not documented | organon/protocols/PROTOCOLS.md — new or updated protocol |
| Heuristic addition | Decision made repeatedly in same way | CLAUDE.md or organon/ETHOS.md — new heuristic row |
| Workflow refinement | Existing workflow that was awkward or incomplete | .claude/skills/<name>/SKILL.md — update steps, context, or error recovery |
| Documentation gap | Information needed but not findable | book-llms/ or book-humans/ — new content |
Step 4: Prioritize
Rank improvements by frequency x impact:
Priority = How often does this recur? × How much friction does it cause?
| Frequency | Impact | Priority |
|---|---|---|
| Every session | Blocks work | Highest — do now |
| Weekly | Slows work | High — do this session |
| Monthly | Annoying | Medium — create RFC or TODO |
| Rare | Minor | Low — note for future |
Step 5: Generate Improvement Plan
For the top-priority improvement, draft:
- •What — specific change to make
- •Where — exact file(s) to modify
- •Why — what problem this solves
- •How to verify — how to confirm the improvement works
Step 6: Execute Improvement
With user confirmation, implement the highest-priority improvement:
- •If tool candidate → create RFC (use
domain-feature-designskill) - •If protocol update → edit PROTOCOLS.md
- •If heuristic addition → edit CLAUDE.md or organon/ETHOS.md
- •If workflow refinement → edit the skill file
- •If documentation gap → create or update the relevant file
Step 7: Check for Stale Terminology
Search ALL files for terminology that may have become inconsistent during the session:
Use the Grep tool to search all files for the term:
Grep pattern="<term-to-check>" path="." glob="*.md"
Also check organon files by name:
cd packages/tools && npx organon find --name "<term-to-check>"
Check common drift points:
- •
CLAUDE.mdvsorganon/ETHOS.md(must stay in sync) - •Skill descriptions vs actual skill content
- •Protocol names vs workflow names
Step 8: Run Verification
cd packages/tools && npx organon verify cd packages/tools && npx organon health
Confirm no regressions from the improvement.
Verification
- • At least one improvement identified and classified
- • Highest-priority improvement either executed or documented
- •
organon verifypasses after any changes - •
organon healthscore has not decreased - • No stale terminology found
Error Recovery
| Failure | Recovery Action |
|---|---|
| No patterns detected | Session may have been too small or too routine. Note for next session. Consider: was the session too short, or is the workflow already well-optimized? |
| Improvement breaks verification | Revert the improvement. Re-analyze the approach — the improvement may need a different implementation strategy. |
| User declines execution | Document the improvement as a TODO comment in the relevant file, or create an RFC for larger improvements. |
| Too many improvements found | Don't try to do everything. Execute only the highest-priority one. Document the rest for future sessions. |
| Improvement requires RFC | Use domain-feature-design skill to create the RFC. Don't implement without proper design for significant changes. |