td-hype-synthesize — synthesis + feasibility vetting + ranking

This skill is loaded by the td-hype-synthesizer sub-agent. You produce exactly one file: _workspace/hype-scout/HYPE_TOOL_BACKLOG.md. You do not scout new trends.

Procedure

Step 1 — Inventory the scouts

List every _workspace/hype-scout/01_scout_*.md file:

ls _workspace/hype-scout/01_scout_*.md

Read each file end to end. Note:

• Which surfaces are present and which are missing (SCOUT MISSING: <surface>).
• Which files are tagged PARTIAL-DUE-TO-NETWORK — propagate that into the synthesis header.
• The total candidate count and per-surface tally.

Step 2 — Build the master candidate table (in memory)

Collect every entry into a working table with columns:

Field	Source
id	new sequential (S001…)
name	scout's tool name
surfaces	list — usually one, but grow during dedup
summary	merged "what artists are doing" + "why hyped"
evidence	union of all citations
scout_hype	the scout's call (H/M/L) — list if multiple scouts
scout_build_ease	scout's guess (S/M/L)
scout_coverage	NOT-COVERED / PARTIAL / COVERED
proposed_layer	scout's suggestion
proposed_operators	scout's suggestion

Step 3 — Dedupe across surfaces

Two entries collapse into one when they describe the same TouchDesigner outcome, even if the scouts framed them differently. Heuristics:

• Same suggested operators → merge.
• Same trend name (case-insensitive, allowing minor wording differences like "StreamDiffusion realtime" vs "Realtime StreamDiffusion bridge") → merge.
• Same "what artists are doing" core (one summary is a subset of another) → merge.

When merging:

• surfaces = union.
• evidence = union (dedupe by URL).
• scout_hype = if all scouts agreed, keep that. If they disagreed, take the mode but annotate inline (e.g. M (showcase: H, tutorials: M)).
• A trend that appeared in 3+ surfaces gets a +1 step on Hype (capped at H). Multi-surface confirmation is itself a hype signal.

Do not delete dropped entries — keep them in a ## Merge log appendix at the bottom so the merge is auditable.

Step 4 — Vet feasibility against the real codebase

For each merged candidate, run targeted greps to lock the layer, operator coverage, and effort estimate. Quote what you find so the call is auditable.

Check	Command
Layer 1/2/3 tool already exists?	grep -rn "create_<name>" src/tools/
Operator in KB?	grep -irn "<operator-name>" src/knowledge/data/
Recipe already covers it?	ls recipes/ | grep -i <keyword>
Bridge endpoint exists?	grep -rn "<endpoint>" src/td-client/touchDesignerClient.ts td/
Roadmap already lists it?	grep -in "<keyword>" docs/ROADMAP.md

Resolve each candidate to one of:

Verdict	Meaning	Effort default
NEW (clean)	no existing tool, all operators in KB, no bridge work	S
NEW (operator gap)	operators exist but not yet wrapped in tdmcp	M
NEW (bridge needed)	needs a new bridge endpoint (e.g. a new device, persistent state)	L
EXTENSION	existing tool covers most of it; add a preset/param	S
COVERED	already shipped — surface as "add cookbook example"	-
ROADMAP	already on docs/ROADMAP.md for a future phase	note phase
FEASIBILITY-UNCERTAIN	operator may or may not exist, KB ambiguous	flag for live probe

Adjust the build-ease score according to the verdict (overwrite scout's guess if vetting disagrees, but log the change).

Step 5 — Rank with an explicit weighting profile

Default profile (state it at the top of the output):

score = HYPE_WEIGHT[hype] × BUILD_EASE_WEIGHT[ease]
HYPE_WEIGHT       = { L: 1, M: 2, H: 3 }
BUILD_EASE_WEIGHT = { L: 1, M: 2, S: 3 }

So the maximum is H × S = 9, the minimum is L × L = 1. Sort descending. Tie-break by: (1) cross-surface count desc, (2) NEW > EXTENSION > COVERED, (3) Hype desc, (4) name asc.

Alternative profiles the user can request (offer them at the end of the file in a "Re-rank knobs" section):

• Hype-only: rank by hype, ignore effort.
• Quick-wins: filter ease == S, then rank by hype.
• Strategic: rank L items above S items (long-term bets first).
• Conservative: drop everything FEASIBILITY-UNCERTAIN.

Step 6 — Identify force multipliers

A force multiplier is one piece of foundational work (a new bridge endpoint, a new operator wrapper, a new prompt) that unlocks 3+ ranked tools. Surface these at the very top of the output, above the ranked list, with the dependents listed by id. They often beat any individual tool on ROI.

Example: "Wrap realsense2TOP as a Layer 3 tool with depth/color/IR streams + a small Layer 2 helper" might unlock S013 (LiDAR particle scatter), S017 (hand-gesture pointer), S021 (depth-keyed silhouette). Flag as Force multiplier FM-01.

Step 7 — Write the output file

_workspace/hype-scout/HYPE_TOOL_BACKLOG.md — this exact structure:

# Hype tool backlog — tdmcp

**Run:** <YYYY-MM-DD>
**Scout coverage:** <surfaces present> (missing: <none|list>)
**Profile used:** default — Hype × Build-Ease (state weights inline)
**Notes:** any `PARTIAL-DUE-TO-NETWORK` flags, any `SCOUT MISSING` gaps.

---

## Force multipliers

> Foundational work that unlocks 3+ ranked tools.

### FM-01 — <one-line summary>
- **Why:** <hype trends it unlocks>
- **What to build:** <layer + scope>
- **Unlocks:** S###, S###, S###
- **Effort:** S / M / L

(...repeat for FM-02 etc., or "None this run.")

---

## Ready for tdmcp-pipeline (top 5)

> Highest-confidence, lowest-friction picks. Hand straight to `tdmcp-pipeline`.

| # | id | Tool | Layer | Hype | Ease | Coverage | Surfaces |
|---|----|------|-------|------|------|----------|----------|
| 1 | S### | `create_<name>` | 1 | H | S | NEW | community-showcase + tutorials |
| 2 | ... |  |  |  |  |  |  |

For each, link to the detailed entry below.

---

## Ranked backlog

| Rank | id | Tool | Layer | Hype | Ease | Coverage | Score |
|------|----|------|-------|------|------|----------|-------|
| 1 | S001 | ... | 1 | H | S | NEW | 9 |
| 2 | ... |  |  |  |  |  |  |

---

## Per-surface breakdowns

### community-showcase
| id | Tool | Hype | Ease | Coverage |
|----|------|------|------|----------|
| ... |  |  |  |  |

### tutorials
...

### generative-ai
...

### hardware-interactive
...

### vfx-aesthetics
...

---

## Candidate details

### S001 — <Tool name>
- **Summary:** <merged what+why>
- **Surfaces:** community-showcase, tutorials
- **Evidence:** (union of all scout citations, one per line)
  - <URL> — <desc>
- **Hype:** H (showcase: H, tutorials: H)
- **Build-ease:** S — vetted
- **Coverage:** NEW (clean) / NEW (operator gap) / EXTENSION — existing `<tool>` / COVERED — existing `<tool>` / ROADMAP — phase X.Y / FEASIBILITY-UNCERTAIN — probe live
- **Layer:** 1 / 2 / 3 / bridge-endpoint
- **Suggested operators:** `<a>, <b>, <c>` (KB confirmed) | (UNVERIFIED — probe live)
- **Tool sketch:** 1-3 sentences, what the tool does end-to-end.
- **Force-multiplier link:** FM-0X (if any)
- **Vet notes:** what grep found / didn't find. Quote paths.

(...all candidates...)

---

## Re-rank knobs

Available alternative profiles (re-invoke synth with the chosen one):
- `--profile=hype-only`
- `--profile=quick-wins`
- `--profile=strategic`
- `--profile=conservative`

---

## Merge log

| New id | Merged from | Reason |
|--------|-------------|--------|
| S001 | scout-community-showcase #3, scout-tutorials #7 | same outcome, same operators |

---

## Follow-up suggestions (not scouted)

> Gaps you noticed during synthesis but were not in any scout. Surface only — do not score; mark for re-scout next run.

- ...

Working principles

• Quote the grep. "I checked grep -rn 'streamdiffusion' src/tools/ — no matches" is auditable; "this is not covered" is not.
• No silent overwrites. If your vetting changes a scout's hype or build-ease score, log it in the entry's Vet notes:.
• Tight top-5. If you can't confidently pick 5, pick 3. Don't dilute.
• Honest gaps. A missing scout file is not a synthesis failure; just note it in the header and move on.

What not to do

• Don't scout new trends. Work only from existing scout files. Surface gaps you noticed under "Follow-up suggestions" without scoring them.
• Don't drop or hide disagreement between scouts. Annotate it.
• Don't pad the ranked list with weak entries to look comprehensive.
• Don't recommend tools that contradict docs/ROADMAP.md without flagging the contradiction.

Re-invocation

If HYPE_TOOL_BACKLOG.md already exists:

1. Read it first.
2. Apply only the requested change — re-rank under a new profile, re-vet a single candidate, refresh the top-5, add a new section — preserve numbering and the merge log.
3. Bump the Run: date and add a short ## Changes this run section near the top describing what changed.