Weekly synthesis

weekly-synthesis

Constraints

• Output is NOT a recap; must contain thesis + evidence
• EXACTLY one action, EXACTLY one question (not zero, not many)
• NEVER let old notes dominate; they are calibration only
• NEVER write synthesis note without user confirmation
• If signal is weak: say "conclusion is uncertain" — do NOT force a thesis
• Default time window: 7 days unless user specifies otherwise

Trigger Boundary

Use when: weekly synthesis, what recent notes mean, emerging thesis, one most valuable next action/question, time-window meaning-making.

Do NOT use for:

• Raw material routing → inbox-triage
• Link-only or relationship review → connection-review
• Project status reporting (not synthesis)
• Promoting result into stable note → note-promotion

Boundaries

• Owns: time-window meaning extraction, thesis formation, gap identification
• Delegates to: note-promotion (synthesis ready for stable note), opencode-context-maintenance (thesis reveals drift)
• Never absorbs: inbox routing, link-only review, health diagnosis

Continuations

Condition	Next skill
User wants to preserve synthesis as stable note	note-promotion
Thesis or gaps reveal context drift	opencode-context-maintenance
Insufficient signal; try relationships first	connection-review
Signal too weak to form any thesis	Escalate to human

Companion Skill

All vault operations follow obsidian-mcp. Forbidden: obsidian_patch_note, obsidian_append_to_note. Verify every write through readback.

Procedure

Phase 1: Scope

1. Identify time window (default: 7 days; respect user-specified period)
2. obsidian_list_notes or obsidian_search_notes → notes in window
3. Count and report scope before proceeding

Phase 2: Read

1. obsidian_get_note format:"document-map" → theme distribution
2. Deep-read only notes that support thesis, contradiction, or gap claims
3. Use older notes only for calibration (not primary signal)

Phase 3: Synthesize

When extracting meaning, read references/examples.md for thesis quality calibration.

Six dimensions:

1. Emerging thesis: one claim + supporting evidence (or "uncertain" if weak)
2. Contradictions: only tensions that matter for reasoning
3. Gaps: missing perspective, evidence, decision, or experiment
4. One action: exactly one highest-value next move
5. One question: exactly one question worth preserving
6. Follow-up note suggestions: optional downstream notes

<details> <summary>Weak Signal Handling (expand when theme stability is low)</summary>

• Fewer than 3 notes in window: report insufficient data, suggest expanding window
• Notes all on different topics: report fragmentation, no forced thesis
• Contradictions only surface-level: report "no meaningful contradictions found"
• Acceptable: "Thesis uncertain. Strongest signal is [X] but evidence is thin."
• One action can be "gather more data on [X]" when signal is genuinely weak

</details>

Phase 4: Report

When formatting output, read references/templates.md for the exact synthesis format.

Do not write synthesis note by default. After user approval, create with obsidian_write_note overwrite:false and read back.

Gotchas

Gotcha 1: Recap disguised as synthesis

What happens: Agent lists what happened this week without extracting meaning Why it's wrong: A recap has zero cognitive return; user already knows what happened Correct approach: Force a thesis claim with supporting evidence, even if uncertain

Gotcha 2: Multiple actions and questions

What happens: Agent returns 3-5 actions and 2-3 questions Why it's wrong: Defeats the purpose of synthesis (focus); user loses signal in noise Correct approach: Force exactly one action and one question; pick highest-value

Gotcha 3: Old notes dominate the thesis

What happens: Agent builds thesis primarily from notes older than the time window Why it's wrong: Recent signal disappears; synthesis becomes rehash of established knowledge Correct approach: Use older notes only as calibration; thesis must be grounded in recent evidence

Validators (CI only, not runtime)

• validators/post-check.sh: Verifies exactly one "One action:" and one "One question:" in output

Exit Criteria

• Output is not a recap (has thesis + evidence)
• Exactly one action and exactly one question
• Thesis has evidence; contradictions are not forced
• Any written note has been read back and verified