/predict-next-action2 — Institutional Knowledge-Enriched Next Action Prediction

How This Works

This skill dispatches a Sonnet subagent that:

1. Assembles a context payload describing what was built, what overarching intent it serves, what scope was completed, and what remains
2. If you have a response-predictor oracle set up (see below), queries it with that context to get institutionally-informed recommendations. If not, reasons from the session's own content alone and says so.
3. Cleans up and frames the result as a nuanced, UX-focused prediction

The orchestrator (you) does NOT do the research — the subagent does. You assemble context and dispatch.

Which oracle this uses

This queries the same response-predictor oracle as /jonathan-check2 and /jonathan-check3 — a NotebookLM notebook built from your own history of corrections, decisions, and priorities, if you've built one (/alignment-harness:harness-setup). It is not a separate "next action" oracle you need to build twice; setting up one response-predictor notebook satisfies all three skills. If none is configured, this skill still runs — it just says plainly, in its opening, that the prediction is based only on this session's content, with no institutional history behind it.

Execution Protocol

Step 1: Assemble Context (orchestrator — inline)

Before dispatching, compose a context block containing:

OVERARCHING INTENT: {what the session's work was building toward — the bigger initiative}

SCOPE COMPLETED:
- {UX story 1 — "When X, then Y" format}
- {UX story 2}
- {UX story N}

CRITICAL CONTEXT:
- {Key finding 1 that shapes what comes next}
- {Key finding 2}
- {Data point that constrains options}

OPEN GAPS:
- {Gap 1 — what's broken or missing}
- {Gap 2}

WHAT WAS NOT DONE:
- {Incomplete item 1}
- {Incomplete item 2}

Step 2: Dispatch Sonnet Subagent (background)

Dispatch a Sonnet agent with run_in_background: true using this prompt template:

You are producing a deeply informed next-action prediction for the person you're working with.

## Your Context
{paste the assembled context block from Step 1}

## Your Task

### Part A: Query your institutional-knowledge oracle (if one is configured)

If a response-predictor oracle is set up (check the switches file / oracle registry — see
/alignment-harness:harness-setup), run something like:

```bash
nlm notebook query <your-notebook-id> \
  "Given this session's work: {1-2 sentence summary of what was built}. What does the institutional knowledge suggest is the highest-leverage next move? Consider: what historically drove the biggest improvements, what patterns this person has corrected agents on, what strategic priorities are active, and what open gaps in the system would compound if left unfixed." \
  --profile <your-profile-name>

If no oracle is configured, skip this step and say so plainly in your opening — do not present a session-only guess with the same confidence as an oracle-backed prediction.

Part B: Synthesize into a prediction

Using the oracle's response (if you got one) AND your session context, produce a prediction with this structure:

Opening (MANDATORY — 2 full sentences minimum, before anything else)

The first two sentences must make crystal clear:

1. What the prediction IS — the specific change being proposed
2. Why it matters NOW — in what conditions this change would have its effect, what exact system would cause the intended effect, and the specific UX-framed intent impact within that context

If no oracle was available, the opening must also say so in one clause, e.g. "based only on this session's own content, since no institutional-history oracle is set up yet."

Do NOT compress these sentences. Be nuanced. Be specific about the system, the user, the condition, and the experience change. The person reads these two sentences and either instantly gets it or moves on.

Example of the right density (from a real use of this skill — opt-in illustration of specificity, not a template for content)

"The checkout flow on this product currently shows the pricing page before the person has experienced the core product, which means the subscription ask lands before they've felt what they're paying for — and the institutional knowledge shows a large past conversion lift was specifically driven by reversing this order, showing a personalized pitch only AFTER the person had engaged enough to generate their own extracted context from the conversation. Restoring this 'pitch after investment' timing to the funnel — where the person experiences the product first, hits an engagement threshold, THEN sees their personalized pitch with a discounted first-period offer — would test whether that same mechanic works in this funnel's context, which is the single highest-leverage conversion experiment available because it combines a proven UX trigger with the funnel that has actual traffic behind it."

Body

After the opening, include:

• What the oracle surfaced (if one was queried) — key institutional knowledge that informed this prediction (cite specific patterns)
• The UX journey this creates — walk through what a person experiences step by step if this change ships
• What could go wrong — risk in UX terms, not technical terms
• What this does NOT address — scope boundary, what's explicitly not included
• Confidence and reasoning — your confidence level with the specific reasoning chain

Format Rules

• No bullet-point summaries in the opening — full prose sentences
• Technical terms must be immediately followed by (what this means for the person using it)
• Every claim about "what works" must reference whether it comes from the oracle's institutional knowledge, live data from this session, or your inference
• Do NOT compress nuance — the person values seeing the full reasoning more than a tidy conclusion

### Step 3: Present Result (orchestrator — inline)

When the subagent returns, present its output directly to the person. Do not summarize or compress — the nuance is the product.

### Step 4: If the person asks follow-up questions

Re-dispatch the subagent with the follow-up question added to the context. The subagent can query the oracle again for deeper context on the specific follow-up, if one is configured.

## Model Selection

| Role | Model | Why |
|------|-------|-----|
| Context assembly | Orchestrator (current) | Has the full session context |
| Research + synthesis | Sonnet | Good balance of depth and speed for oracle queries |
| Presentation | Orchestrator | Maintains conversation continuity |

## What This Is NOT

- NOT a replacement for `/predict-next-action` — that's a quick inline prediction. This is the enriched version.
- NOT a planning skill — it predicts ONE action, not a roadmap
- NOT a governer — it doesn't score or route. It recommends.