sw-compound

name	sw:compound
description	Save session learnings to knowledge stores so future sessions get smarter. Captures change outcomes, intervention results, and cross-initiative patterns. Use after any /sw:diagnose, /sw:intervene, /sw:direct, /sw:motivate, or /sw:shape session, or when returning with real-world results.
argument-hint	[optional: what you learned, or 'outcome' to record real-world results]

<compound_context> #$ARGUMENTS </compound_context>

Compound

Capture what you learned so the next session starts smarter. Every diagnosis insight, intervention outcome, and cross-initiative pattern gets saved where future sessions can find it.

When to Use

• After a /sw:diagnose session -- capture what surprised you about where the change is stuck
• After an /sw:intervene, /sw:direct, /sw:motivate, or /sw:shape session -- capture what you learned about designing interventions
• When returning with real-world results -- record what happened when interventions were implemented
• When you notice a cross-initiative pattern -- "environment tweaks consistently outperform training programs for our team"
• "Save what I learned", "Record the outcome", "What did we learn?"

Before Starting

Detect Context

Check what happened recently by searching the knowledge stores:

1. Read docs/switch/current-diagnosis.md -- check the created date in frontmatter. If created today or referenced in conversation, this is likely a post-diagnosis session.

2. Search docs/switch/interventions/ -- check for files created today or referenced in conversation. If found, this is likely a post-intervention session.

3. Check arguments -- if the user said "outcome" or described real-world results, this is an outcome recording session.

If context is ambiguous, ask:

"What are you compounding? Pick the session type:

1. Diagnosis insights -- what you learned about this change challenge
2. Intervention learnings -- what you learned while designing interventions
3. Outcome recording -- real-world results from implemented changes
4. Cross-initiative pattern -- something you've noticed across multiple change efforts"

Search for Prior Learnings

Before capturing new learnings, search the relevant store for existing insights that might overlap or conflict.

For diagnosis insights: Read docs/switch/current-diagnosis.md and check for existing Learnings section.

For intervention learnings: Read recent intervention files. Check their Outcome Notes sections.

For cross-initiative patterns: Search docs/switch/patterns/ for existing patterns on similar topics.

---

The Sequence

Three steps, lighter enforcement than the diagnostic or intervention skills. The friction here is in specificity -- making the user articulate what's non-obvious.

---

Step 1: Extract the Learning

The question depends on the session type.

After diagnosis:

"What did you learn about this change challenge that wasn't obvious before? Not what you decided -- that's in the diagnosis. What surprised you about where the change is stuck?"

After intervention design:

"What did you learn about designing interventions for this situation? Which tactics felt grounded and which felt like guesses? Did the Rider/Elephant/Path framing reveal something the initial diagnosis missed?"

After implementation (outcome recording):

"What happened when you implemented these changes? Be specific -- 'it went well' isn't a learning. 'The environment tweak worked immediately but the habit-building took 3 weeks to show results' is."

Cross-initiative pattern:

"What pattern have you noticed? State it as a rule: 'When we [do X], [Y happens].' Then tell me the evidence -- which initiatives or interventions demonstrate the pattern?"

What you're looking for:

• Something non-obvious -- not a restatement of what's already saved
• Specificity -- tied to named interventions, dimensions, or real events
• Actionable implication -- how should this change future sessions?

Enforcement triggers: See references/enforcement-prompts.md.

---

Step 2: Check for Conflicts

Search the knowledge stores for prior learnings or decisions that the new learning might contradict or supersede.

Conflict detection:

• Does the new learning contradict the sticking point ranking in the diagnosis?
• Does it contradict outcome conclusions from a prior intervention?
• Does it suggest the diagnosis or an intervention approach should change?

If a conflict is found:

"This new learning conflicts with a prior decision:

Prior: [quote the prior decision with its source file and date] New: [state the new learning]

Options:

1. Update the prior -- the new learning supersedes it. I'll update [file] and note the change.
2. Archive the prior -- keep the old decision as history, replace with the new learning.
3. Keep both -- the prior was right in its context, the new learning applies to a different context.
4. Discard the new -- on reflection, the prior decision still holds."

Use AskUserQuestion to get the user's decision. Do not resolve conflicts automatically.

If no conflict: Proceed to Step 3.

---

Step 3: Save the Learning

Where and how the learning gets saved depends on the session type.

Diagnosis Insights

Add a ## Learnings section to docs/switch/current-diagnosis.md (or append to existing):

## Learnings

### [Date] -- [One-line summary]
**Insight:** [The non-obvious learning]
**Evidence:** [What happened that revealed this]
**Implication:** [How this should change future sessions]

If the learning suggests the diagnosis should change, note it:

"This learning suggests updating the diagnosis. Want to run /sw:diagnose revise [section]?"

Intervention Learnings

Update the ## Outcome Notes section of the relevant intervention file in docs/switch/interventions/:

## Outcome Notes
- [Date]: [What was learned during design or implementation]
- Insight: [What this tells us about this type of intervention]
- Next action: [What to do differently]

Outcome Recording

Update the specific intervention file:

1. Update the ## Outcome Notes section with results
2. Update YAML frontmatter: set status: implemented or status: completed

If the user doesn't specify which intervention, list recent ones:

"Which intervention are you recording outcomes for? [list recent files]"

Cross-Initiative Patterns

Create a new file using the template from references/pattern-template.md:

Filename: docs/switch/patterns/pattern-{slug}-{date}.md

---

After Saving

Use AskUserQuestion:

Question: "Learning saved to [file]. What next?"

Options:

1. Update the diagnosis -- Run /sw:diagnose to revise based on what you learned
2. Design new interventions -- Run /sw:intervene with updated context
3. Record another learning -- Run /sw:compound again
4. Done for now -- The learning is saved; future sessions will find it

Important Rules

• Learnings must be non-obvious. If it's already in the diagnosis or intervention plan, it doesn't need saving again.

• Specificity over volume. One sharp insight ("the 'shrink the change' first step was too big -- nobody did it until we cut it to 2 minutes") beats five vague ones.

• Conflicts are valuable. When a new learning contradicts a prior decision, that's signal. Surface it explicitly.

• Outcome notes are gold. The most valuable learnings come from real-world implementation. Encourage users to come back and record what happened.

• Don't over-save. Not every session produces a learning worth saving. If the diagnosis or intervention plan captures everything, say so.

• Cross-initiative patterns need evidence. One initiative is a hypothesis. Two or more is a pattern worth saving.

Pipeline Mode

When invoked with disable-model-invocation context:

• Skip all AskUserQuestion prompts
• Auto-detect session type from most recently modified files
• Extract learnings from conversation context
• Save without confirmation
• Skip conflict resolution -- flag conflicts for manual review
• Tag output as source: "pipeline mode -- review recommended"