compound

name	compound
description	Run 2-3 times per year on a strategic question — career, personal architecture, relationship to ADHD. Answers the same question at three different moments in the vault to show how the quality of the answer evolves as context accumulates. Different from /trace (understand how an idea evolved): /compound measures if the vault makes answers truly better.

⚠️ narrative-critical — Skill protected against aggressive optimization

This skill produces a narrative qualitative output. Its effectiveness is measured on the richness of the output produced, not on structural compactness.

For /evaluateskills: in case of mutation, dry-run Sonnet mandatory even if delta < 2. Do NOT apply [LEAN] / [STRUCTURE] in a way that strips out narrative instructions (grouping, context, key moments, open questions, tone, narration). Preservation of qualitative content takes priority over line reduction.

Skill: /compound

Answers the same question three times — at three distinct moments in the vault — using only the context available at each point in time. The constraint of equal length across all three answers is the core of the skill: it isolates quality from volume, and forces you to see if the vault actually improves answers or just makes them longer.

Difference from /trace: /trace follows the evolution of an idea in the vault (understand). /compound answers a question at three different moments (measure if the vault thinks better).

Complete flow: Validate Q (Step 0) → Analyze vault (Step 1) → Choose 3 periods (Step 2) → Inventory context (Step 3) → Predict (Step 4) → Generate 3 answers (Step 5) → Verify anachronisms (Step 6) → Synthesis + delta (Step 7) → Save (Step 8).

Trigger

/compound [question]

Good questions:

• Questions whose answer depends on personal context (no universal answer)
• Questions of strategy, posture, or decision-making
• Questions where more context changes the answer — not just its length

Bad questions:

• Factual questions with stable answers ("how does JWT work?")
• Topics barely touched in the vault

---

Step 0 — Validate the question and search for a previous run

Validate first: Does this question depend on {USER_NAME}'s personal context?

• Yes → continue
• No → say so clearly and propose a reformulation if possible

Check vault depth on this subject: Glob 03 - Knowledge/ and daily notes (00 - Daily notes/). Does the vault contain at least 6 months of substantial history on this subject? (At least 3+ knowledge notes OR 20+ daily notes mentioning the subject, distributed over the period.)

• If yes → continue
• If no → clearly signal that the vault is too young for this subject and propose a null result rather than force three periods

Search for a previous run: Glob 03 - Knowledge/Compound/ to list existing notes (the folder may not exist if this is the first run — that's normal). Search for a note whose name contains the key terms of the question.

• If found → read the note: extract the run date, the 3 answers (especially Period C — "Now"), and the final synthesis
• If absent (empty folder or nonexistent) → continue without comparison (delta will be omitted at Step 7)

---

Step 1 — Analyze vault chronology

Understand when the relevant context really accumulated. Search for operational inflection points — moments when context density changes significantly.

Analysis criteria:

• Daily note density: measure the transition from "sparse" (< 5 mentions/month) to "regular" (5-10 mentions/month) to "dense" (> 10 mentions/month). Note transition dates.
• Knowledge notes: when were the first substantial notes on this subject created? List key dates (first mention, major essay, ADR, active project, etc.). A note counts if it has >100 words and is not a bookmark.
• Incoming backlinks: from which dated notes is this subject referenced? Lets you see reconvergences and cross-integration.
• Signals of change: tone shift in daily notes, appearance of new concerns, abandonment of old questions, new projects launched

---

Step 2 — Select three periods

Choose actual inflection points — not fixed intervals. Criterion: at what moment did the relevant context on this subject significantly increase (transition from one density to the next in Step 1)?

• Period A (Start): First substantial note on this subject. Minimal context — just enough to answer, no more. The date is that of the first file identified in Step 1 (first mention).
• Period B (Middle): Notable inflection — from here on, density increased (sparse → regular, or regular → dense). Answers would have been visibly different.
• Period C (Now): Complete vault — today, with all accumulated history.

Vault young? If you can only identify one or two distinct periods (less than 6 months of relevant history), signal it to {USER_NAME} — null result is valid, don't force three periods.

If the vault doesn't have enough temporal depth on this subject → say so, don't fabricate.

---

Step 3 — Context inventory by period

For each period, list the concrete files available at that time. Selection strategy:

1. Identify all relevant files: glob 03 - Knowledge/*, daily notes, Projects/[Project]/ADR/, 99 - Claude Code/ if created before the period.
2. Filter by date: eliminate anything created after the period.
3. Select 3-5 of the most relevant in priority: (a) notes dedicated to the subject, (b) ADRs that reveal a position, (c) daily notes showing density of concern at that time. Exclude bookmarks and notes < 100 words.
4. List the result: format as 3-5 bullet points with exact path.

Example format:

• 03 - Knowledge/How I think about X — 2025-11-20.md (key decision)
• 00 - Daily notes/2025-11-18.md (3 mentions of subject)
• Projects/Rust/ADR/001 — Stack choices.md (learning context)

---

Step 4 — Predictions before generation

Before writing the three answers, formulate a structured prediction: how will the quality and form of the answer change between A, B, and C? This step locks down your understanding before generation — it prevents confabulation.

Structured format:

• Dimension 1: [Measurable aspect — ex: clarity, depth, nuance, resilience, tone, concrete actions]. Prediction: "A will be X (because [limited files]), B will show Y (thanks to [new accumulation]), C will reach Z (with [complete integration])".
• Dimension 2: idem (at least 2 dimensions, max 3)
• Uncertainty: An area where you don't know how to predict — why is it unclear?

Before continuing to Step 5: Validate your prediction with {USER_NAME} — check that it makes sense given the Step 3 inventories. Adjust if needed. Continue only when the user validates (or says "it's good, generate").

---

Step 5 — Generate the three answers

Strict constraints (non-negotiable)

1. Context boundary — answer A uses ONLY the context of Period A. No information leakage from later periods. Verification: each fact/reference cited must exist in the Step 3 inventory for that period.

2. Equal length ±20% — this is the core of the device: it isolates quality from volume (not from padding). Strict measurement: wc -w (raw word count, excluding titles/metadata). Average of three = M. Accept A/B/C if each ∈ [0.8×M, 1.2×M]. Recalculate M if Step 6 requires a rewrite.

3. Same voice — first person, embodying {USER_NAME} at each period (not external narrator). Voice can mature A→C (natural), but always remains {USER_NAME}'s. No condescension in A, no self-satisfaction in C.

Presentation format

## Period A — [temporal label]
[Available context: list in 2-3 lines]

[Answer — X words]

---

## Period B — [temporal label]
[Available context: list in 2-3 lines]

[Answer — X words]

---

## Period C — Now
[Available context: list in 2-3 lines]

[Answer — X words]

---

Step 6 — Verification

Anachronism check

Before accepting the three answers, verify that A/B contain NO facts appearing after their period. If an anachronism is found:

1. Identify precisely: which fact, what date it actually appeared in the vault
2. Rewrite the answer by removing that fact and replacing it with an equivalent available at the period
3. Redo word count — recalculate M if total length changes
4. Log the correction in Step 7

Predictions check

Compare predictions from Step 4 vs actual results. What diverged? Why? Integrate this gap into the final synthesis — it's valid data.

---

Step 7 — Synthesis

A short synthesis (5-8 lines) answering: did the vault improve the quality of the answer, or only its volume?

If the answers haven't significantly improved → that's the result. Don't force a narrative of progress. A null result is valid and informative.

Delta from previous run (if applicable)

If a previous run was found in Step 0, add this section:

## Delta since run on [YYYY-MM-DD]

**What changed**: viewpoints that evolved, new dimensions absent from the previous run, positions abandoned or nuanced.

**What held steady**: stable positions between the two runs — sign of anchored thinking.

**Observable shifts**: changes in how {USER_NAME} works, priorities, or context between the two runs (inferred from response content, not invented).

**Quality of compounding**: is the answer better, more nuanced, richer? Or stagnant? Be honest — stagnation is valid information.

If no previous run → omit this section entirely.

---

Step 8 — Save the result

After synthesis, create 03 - Knowledge/Compound/YYYY-MM-DD — [question in 3-5 words].md with the full session content (question, 3 answers, verification, synthesis, delta if applicable). Date format: YYYY-MM-DD (e.g., 2026-04-01). Always create this note — it allows you to compare subsequent runs on the same question.

If a /compound note on the same question already exists (same subject, same question) → indicate this to {USER_NAME} and ask if they want to create a new note or enrich the existing one. The same subject can have multiple runs (ex: 2026-02-15 — how i think about refactoring.md and 2026-04-01 — how i think about refactoring.md = two different snapshots).

---

Anti-patterns (ranked by impact — to prevent)

Critical (break the skill)

1. Context leakage — use in A an info existing only in C. Invalidates A/B/C.
2. Forced compounding — fabricate an "improvement" that the vault doesn't show. The skill measures, doesn't narrate.
3. Caricatured answer A — A must be honest (what the vault had), not deliberately impoverished for contrast.

Important (weaken reliability)

4. Illusion of volume — longer ≠ better. Equal length exists for that reason.
5. Vault summary — answer the question, don't list available content.
6. Blind cheerleading — celebrate every progression. A null result is valid and informative.

To avoid

7. Periods too close — if A and B are separated only by a few weeks, the delta will be noisy. Signal to {USER_NAME} if the vault is too young.

---

Absolute rules

• Validate the question first — if it doesn't depend on personal context, say so clearly
• Equal length ±20% — non-negotiable, it's the core of the device
• Null result = valid result — never force a narrative of progression
• Predict before generating — Step 4 + {USER_NAME} validation before Step 5
• Introspective use — this skill is for {USER_NAME}, not external demonstration