Strategic proposal vault persistence check

Strategic Proposal Vault Persistence Check

✅ PROMOTED — TDD pressure test PASS. RED: 3-line daily-note limbo ("daily-note-centric instead of vault-structure-centric"); GREEN: 4-step persistence check + target paths per output + status markers with date column + privacy boundary from CLAUDE.md considered. Cycle 2 polish: privacy boundary explicit (sensitive-tag list linking), 50%-persisted edge case (extend existing vs parallel), JSON schema example snippet, cross-ref to "Current Truth before Timeline".

The most expensive bug in multi-session work is re-discovery: a previous session figured something substantial out, didn't write it down, and the current session pays for the rebuild. This skill is the systematic check that prevents it.

When to use

• Natural end of a strategic discussion block (decision tree, prioritization, roadmap)
• Session-end, before context-window close
• Mid-session pivot from "thinking through" mode to "doing" mode
• User pushback "is that in the backlog?" or "where should we pick up?"
• When you notice yourself constructing a table, JSON-schema, or multi-step recommendation that took >10 minutes of structuring

When NOT to use

• Chat-only sessions with no structured output
• Mid-debug or mid-implementation — finish the active task first
• When post-session-skill-review is already running on the same daily note — coordinate (run this FIRST, then skill-review)
• One-off solutions to single-instance problems (just log in daily note, no vault-cluster needed)

The check (4 steps, ~5-10 minutes)

Step 1: Scan the session output for "structured strategic content"

Mentally walk back through the session and flag any of:

• Multi-row decision tables (priority × effort × value, candidate × pain-point × reduction)
• JSON/data schemas with concrete keys (API response shapes, config templates)
• Numbered multi-step recommendations ("Stage 1: X, Stage 2: Y, ...")
• N-category frameworks ("3 strategies", "5 categories", "7 anti-patterns")
• Effort-estimated backlogs with time estimates per item
• Pain-anchor + counter-measure pairings ("Pre-state: 25 calls / Post-state: 1 call")

If NONE of these appeared → no proposals to persist, skip this skill.

If 1+ → continue to Step 2.

Step 2: Vault-grep for existing persistence

For each flagged content block, search the vault for whether it's already there:

Grep keywords from the content (proper nouns, table-row identifiers, distinct phrases)
across `Projects/`, `Areas/`, `Resources/`, and today's daily note.

If the content is already in a cluster hub or daily-note section → done, no action needed. If it's only in the conversation buffer → continue to Step 3.

Step 3: Choose the persistence target

Content type	Target
Strategic framework (N categories, multi-domain)	New cluster Projects/<theme>/<theme>.md
Single-topic decision/roadmap	Existing project's hub note + new sub-note
Tooling/method recommendation	Resources/AI & Machine Learning/<topic>.md
User-specific maxim derived from session	Core memories + project CLAUDE.md
Backlog table for one project	Existing project's hub ## Backlog section

If unsure, default to: new cluster note with explicit "Roadmap" + "Backlog" + "Inventory" structure. Single-purpose hub note is recoverable later; missing content is not.

Step 4: Write with status marker per item

Critical: don't just copy-paste the proposal. Add a status column (✅ done / ⚠ partial / ❌ open) per row/item.

Without status: future session reads it as "fresh backlog" and re-evaluates. With status: future session sees "X done, Y open" and continues from where you left off.

For tables:

| # | Item | Effort | Token-Reduction | **Status** |
|---|---|---|---|---|
| 1 | health-aggregator | 3h | 90% | ✅ Phase A-G live |
| 2 | logs-prefilter | 1h | 70% | ✅ integrated as /health/logs |
| 3 | vault-search-helper | 1h | 20% | ❌ a planned later phase |

For step sequences:

| Stage | Action | **Status** |
|---|---|---|
| Immediate (5min) | activate MCP | ✅ done |
| This week (3h) | health-aggregator | ✅ in 50min wallclock |

Quick-Reference Workflow

1. End of session reached (or user: "let's wrap up")
2. Scan: tab. / schema / N-step / N-category? [Yes/No]
3. if Yes: grep vault for 2-3 distinct keywords from each block
4. if not persisted: choose target path + create note
5. Copy content WITH status column per row
6. Daily note: leave short reference "→ [[link]]"
7. Brain-dump / inbox: mark original block as <!-- migrated → [[link]] -->

Anti-patterns

• ❌ "That'll go in the backlog tomorrow" — tomorrow is a new session without context. Persist NOW, not later.
• ❌ Copy-paste without status marker — status is the difference between "living roadmap" and "dead backlog list"
• ❌ In daily note only "Status: worked out" — daily notes are not searched for "what's running?". That belongs in the project cluster.
• ❌ Skipping vault grep and immediately creating new note — leads to double structures (two clusters for the same topic)
• ❌ Persisting too fine-grained — EACH single table entry as its own note creates vault sprawl. A consolidated cluster note with sub-sections is better.
• ❌ Rephrasing original wording — preserve the user's language (especially the user's own maxims + quotes). Re-phrasing loses nuance.

Real-world example: health-aggregator session

Previous session worked out:

• 5-category token-optimization framework (model selection / sub-agents / MCP / tools / disciplines)
• 6-tool candidate table (tool / pain-point / effort / reduction)
• JSON schema for health endpoint
• 4-stage recommendation order

Persisted: nothing. Brain-dump item "token optimization" had only 3 keywords.

Follow-up session:

• User pushback: "the previous session had nothing documented"
• Recovery cost: ~30-40k tokens (user paste-back + cluster creation + schema rebuild + status marker per item)
• Would have cost ~5-10 minutes at end of previous session with THIS skill → 5-10× ROI

Token saving from active application of this skill: ~25-35k per avoided re-discovery.

Connection to other skills

• post-session-skill-review (GA) — complementary; that one scans for SKILL candidates, this one for PROPOSAL persistence. Run THIS one FIRST so proposals are in vault before skill-review scans.
• A vault graph-cleanup pass — after cluster creation, run a sanity check that catches duplicate-cluster scenarios and ambiguous links
• brain-dump-to-phased-roadmap (GA) — if persistence target is a fresh brain-dump-derived cluster, use that skill for the roadmap-phase structure

Promotion notes (DRAFT → GA)

Created after user pushback on missing persistence of previous-session strategic proposals. Promote via skill-tdd-promotion-workflow after:

• 1 RED subagent pressure test: edge cases like "proposal partially persisted (50% in daily note, 50% in conversation buffer)" and "proposal contradicts existing cluster"
• 1 real-world catch where the skill fires correctly at session-end and identifies an un-persisted strategic block
• Cross-link to post-session-skill-review with explicit "run me FIRST" coordination note