Pipeline Idea Generator

---
name: pipeline-idea-generator
description: "Stage 10 of the observable pipeline. Generates fresh high-leverage ideas to feed the pipeline queue when it runs low. Queries multiple sources (jonathan-check2 oracle, leverage-predictor, predict-next-action, agent_find for errors/complaints, compaction incomplete items), scores each candidate via governer methodology, formats as pipeline-ready items, and appends to PIPELINE-TRACKER.md after deduplicating against existing runs/."
triggers:
  - "pipeline idea generator"
  - "feed the pipeline"
  - "queue is low"
  - "generate pipeline ideas"
  - "pipeline stage 10"
  - "refill pipeline"
---

# Pipeline Idea Generator — Stage 10

## Purpose

The pipeline only processes what it knows about. This skill is the mechanism by which it learns about new things.

In v1, the pipeline only consumed items pre-generated by leverage-predictor. It never asked: "What else should we look at?" This skill closes that gap. When the queue drops below 5 actionable items, this skill runs, queries multiple institutional knowledge sources, scores each candidate, deduplicates against existing work, and appends new items to the pipeline queue.

**Pipeline tracker location:**
```
"$(alignment-harness records pipeline)/PIPELINE-TRACKER.md"
```

If `PIPELINE-TRACKER.md` does not exist in that folder, create it with an empty `## Proposed/ Queue` table.

---

## When to Run

Trigger automatically or on demand when:

1. The PIPELINE-TRACKER.md Proposed/ Queue has fewer than 5 `queued` items
2. Explicitly invoked: "generate pipeline ideas", "feed the pipeline", "Stage 10"
3. After a major batch of pipeline items reaches `done` status

Check queue depth first:

```bash
# Count queued items in PIPELINE-TRACKER
grep -c '| queued' "$(alignment-harness records pipeline)/PIPELINE-TRACKER.md"
```

If count >= 5, print: "Queue has {N} queued items — no refill needed." and stop.

---

## Step 1: Gather Candidates from All Sources

Run these 5 source queries. Do them in parallel where possible. Collect raw candidates — do not score yet.

### Source A — Jonathan Check2 Oracle

Ask the NotebookLM oracle what has not been proposed yet:

Setup: this source needs a NotebookLM notebook of your own. If you don't have one, make one at https://notebooklm.google.com and load it with your priorities, patterns and corrections, because the oracle can only predict what you would want from what is in that notebook. Put its ID in place of `<your-notebook-id>` below.

```bash
nlm notebook query <your-notebook-id> \
  "Based on everything you know about the user's priorities, patterns, and corrections, what are the 3 highest-leverage things that haven't been proposed or investigated yet? Focus on user-facing pain, conversion gaps, and coaching quality gaps that are unaddressed." \
  --profile <your-profile>
```

Or via MCP:
```
mcp__notebooklm__query_notebook({
  notebook_id: "<your-notebook-id>",
  profile: "<your-profile>",
  question: "Based on everything you know about the user's priorities, patterns, and corrections, what are the 3 highest-leverage things that haven't been proposed or investigated yet?"
})
```

Extract: up to 3 candidate ideas, each with the oracle's reasoning.

### Source B — Leverage Predictor

Invoke `/leverage-predictor` to generate new LP proposals based on current codebase state. Ask the skill to produce candidates (not yet submit them) with scores >= 60.

The leverage-predictor skill workflow produces scored candidates with user impact (40%), business value (30%), effort efficiency (20%), risk reduction (10%) breakdowns. Capture those candidates here.

### Source C — Predict Next Action

Invoke `/predict-next-action` for session context signal. The output will be a UX possibility + highest-leverage next move. Treat this as one candidate if it is not already in the pipeline.

### Source D — agent_find: Recent Errors, Complaints, System Patterns

```bash
# Query for recent unaddressed errors and user pain signals
node "${CLAUDE_PLUGIN_ROOT}/scripts/memory-search.js" \
  "recent error logs user complaints system failures unaddressed" --limit 8
```

Also query the API directly for recent high-severity errors not yet in the pipeline:

```bash
export DEVELOPER_API_KEY=$(grep -E '^DEVELOPER_API_KEY=' .env | cut -d'=' -f2)

curl -s "<your-api-url>/<your-error-log-endpoint>?level=error&limit=20&includeDismissed=false&includeSilenced=false" \
  -H "X-API-Key: $DEVELOPER_API_KEY" | \
  node -e "
    const d = JSON.parse(require('fs').readFileSync('/dev/stdin','utf8'));
    const logs = d.logs || d.data || d;
    (Array.isArray(logs) ? logs : []).slice(0,5).forEach(l =>
      console.log(l.severity, l.message?.slice(0,120))
    );
  "
```

Extract up to 3 candidates from distinct error patterns not already tracked in PIPELINE-TRACKER.md.

### Source E — Compaction Incomplete Items

Query compactions for unfinished items from prior sessions:

```bash
alignment-harness compaction open-items
```

Extract up to 3 candidates from confirmed incomplete items that represent real unfinished work.

---

## Step 2: Deduplicate Against Existing Pipeline

Before scoring, eliminate any candidate that is already in the pipeline.

```bash
# Read the tracker to check existing titles and run dirs
cat "$(alignment-harness records pipeline)/PIPELINE-TRACKER.md"
```

Also check for existing run directories:

```bash
ls "$(alignment-harness records pipeline)" 2>/dev/null
```

**Dedup rule:** If a candidate's core claim substantially overlaps with any existing LP, Seed, or Critical item (regardless of status), it is a duplicate. Drop it. When in doubt — if 60%+ of the core finding is already investigated — drop it.

---

## Step 3: Score Each Remaining Candidate

Apply governer methodology (SILA + touchesUserState) to each deduplicated candidate.

For each candidate, derive:

| Factor | Weight | Questions |
|--------|--------|-----------|
| User Impact | 40% | How many users affected? How visibly? |
| Business Value | 30% | Revenue, retention, conversion delta? |
| Effort Efficiency | 20% | Expected impact per hour of work? |
| Risk Reduction | 10% | Prevents future outages or silent failures? |

**Category overrides** (from governer):
- payment / auth claim → floor 100
- core-coaching claim → floor 90
- onboarding claim → floor 85
- admin-only (no user state) → floor 0, skip

**Minimum threshold:** Only carry forward candidates scoring >= 50. Drop anything below.

Print a quick score table:

```
| Candidate | Source | Raw Score | touchesUserState | Final Score |
|-----------|--------|-----------|-----------------|-------------|
| {title}   | A      | 72        | yes             | 72          |
```

---

## Step 4: Format as Pipeline-Ready Items

For each surviving candidate, produce a pipeline-ready entry in the same format as existing LP entries in PIPELINE-TRACKER.md:

```markdown
| LP-{next} | {score} | {Short descriptive title} | queued | — | — | Source: {A|B|C|D|E} — {1 sentence core claim to verify} |
```

**LP numbering:** Read the highest existing LP number from the tracker and increment. Check both the table rows and any `done` entries.

**Title format:** 2-5 words, noun phrase, describes the capability or gap — not the finding. Matches the convention of existing LP titles (e.g., "Resonance-Gated Paywall", "Fix 209 Swallowed Errors").

**Notes field must include:**
- `Source: {letter}` — which source surfaced it
- One sentence stating the core claim that Stage 3 (verify) will confirm or disprove
- Initial score

---

## Step 5: Append to PIPELINE-TRACKER.md

Read the tracker, find the `## Proposed/ Queue` table, and append the new rows before the blank line that ends the table. Do not reorder existing entries. Do not change any existing row's status.

Update the "Last updated" line at the bottom with today's date and note: "pipeline-idea-generator Stage 10 — {N} new items added."

---

## Step 6: Print Summary

After writing, print:

```
## Stage 10 Complete — Pipeline Refilled

Sources queried: A (jonathan-check2), B (leverage-predictor), C (predict-next-action), D (agent_find errors), E (compaction incomplete)
Raw candidates: {total}
After dedup: {N}
After score filter (>=50): {N}
New items added to queue: {N}

New items:
  LP-{N}: {title} — Score: {score} — Source: {A|B|C|D|E}
  LP-{N}: ...

Queue depth now: {new total queued count}
```

---

## Anti-Patterns

- NEVER add an item already investigated (any status) — check both LP rows AND run directories
- NEVER add candidates scoring < 50 — they waste pipeline capacity
- NEVER fabricate scores — derive from the 4-factor formula or governer category override
- NEVER skip the dedup step — the pipeline tracker is the single source of truth
- NEVER add items from Source A alone without checking against Source D (oracle predictions need system log corroboration to be credible)
- NEVER modify existing tracker rows — only append new ones

---

## Related Skills

- `/leverage-predictor` — generates LP proposals with SILA scoring
- `/predict-next-action` — session-context next move prediction
- `/jonathan-check2` — the user oracle for priority prediction
- `/governer` — leverage scoring methodology
- `/next-priority-tasks` — aggregated ranked task view
- `/how-to-submit-and-track-proposals` — proposal submission after pipeline processing