Agent skill

session-compounding

Reviews a work session's output to detect improvable patterns, classify improvements, and execute the highest-priority one. Use at the end of significant work sessions to ensure the COMPOUND phase of the enforcement loop happens. Without explicit compounding, improvement never happens.

Stars 163
Forks 31

Install this agent skill to your Project

npx add-skill https://github.com/majiayu000/claude-skill-registry/tree/main/skills/other/other/session-compounding

SKILL.md

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

  1. Load project constraints:
    • Read CLAUDE.md (project-level guidance and decision heuristics)
  2. Load pattern references:
    • Read book-llms/patterns.md (Recursive Collaboration + Observation Accumulation sections)
    • Read book-llms/workflow-authoring.md (workflow quality attributes for evaluating workflow improvements)
  3. Load recent observations:
    • Read organon/observations/README.md (index of existing observation files)
    • Read the most recent observation file listed in the README (for context on prior findings)

Steps

Step 1: Review Session Work

Examine what happened during the session:

bash
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

Review the most recent observation file to check whether this session's findings relate to previously recorded observations. Look for recurrence — the same friction appearing again strengthens a signal into a pattern.

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:

  1. What — specific change to make
  2. Where — exact file(s) to modify
  3. Why — what problem this solves
  4. 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-design skill)
  • 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: Record Observations (Optional)

If this session produced observations worth preserving beyond the immediate session, record them:

Decision:

  • An active observation file covers this topic → add to it (new O-entry in the Observations section)
  • No relevant file exists → create new organon/observations/NNN-descriptive-name.md

What to record: Each observation needs Signal (what happened), Implication (what it means), Suggested Action (what to do). See the Observation Accumulation Pattern for the convention.

When NOT to record:

  • Single-occurrence friction you already fixed (just a bug fix)
  • Opinions without evidence
  • Observations already captured in methodology guidance
  • Session-specific context that won't generalize

Not every session produces observations — that's fine. Only record when the session revealed something worth remembering across sessions.

Step 8: 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:

bash
cd packages/tools && npx organon find --name "<term-to-check>"

Check common drift points:

  • CLAUDE.md vs organon/ETHOS.md (must stay in sync)
  • Skill descriptions vs actual skill content
  • Protocol names vs workflow names

Step 9: Run Verification

bash
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 verify passes after any changes
  • organon health score 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.

Expand your agent's capabilities with these related and highly-rated skills.

Didn't find tool you were looking for?

Be as detailed as possible for better results