dedup-proposals

Exploration Deduplication and Proposal Curation

Context

Raw explorations in .elmer/proposals/ are unvetted AI ideation — not project documents. This skill consolidates them into curated PRO-NNN entries in PROPOSALS.md, the project's proposal registry. See ADR-098 for the three-tier maturity model (explorations → proposals → decisions).

Mode Selection

No arguments (/dedup-proposals): Triage mode — scan all explorations, report clusters, recommend PRO-NNN entries.

With argument (/dedup-proposals <PRO-NNN> or /dedup-proposals <filename>): Synthesis mode — find all related explorations of the named topic, merge them into one PRO-NNN entry.

---

Triage Mode (no arguments)

Read PROPOSALS.md to determine the current max PRO number and existing entries.

Read all .md files in .elmer/proposals/. For each file, extract:

• The elmer:archive metadata block (topic, id, archetype, status)
• The first 200 words of the Summary section

Cluster explorations by topic similarity:

1. Exact match: Same topic field in elmer metadata
2. Near match: Topics sharing 3+ significant words (excluding stopwords)
3. Content match: Summaries with >40% semantic overlap

For each cluster, report:

• File names and sizes
• Overlap percentage (structural alignment of sections)
• Which file is more developed (word count, section completeness)
• Unique ideas in each file not present in the other(s)
• Contradictions between files (if any)
• Recommended PRO-NNN entry: title, type (Feature/Theme/Policy/Enhancement), and which exploration to use as canonical base
• Cross-references to existing ADR/DES identifiers mentioned in the explorations

For non-clustered explorations, report as standalone candidates for individual PRO-NNN entries.

Classify each cluster/standalone by proposal type:

• Feature — New capability (word graph, read-to-me mode, focused reader)
• Theme — Content theme for taxonomy integration (self-worth, vibration/AUM, extra-solar life)
• Policy — Governance, legal, or process change (copyright, AI automation)
• Enhancement — Improvement to existing feature (visual design, reader mode)
• Retrospective — Design review findings (cognitive load, resonance analysis) — these may not need PRO-NNN entries; they inform design revisions

Present the full report. Offer three actions:

1. Auto-resolve all clusters and create PRO-NNN entries in PROPOSALS.md
2. Resolve cluster-by-cluster with approval
3. Show side-by-side comparison for a specific cluster

---

Synthesis Mode (with argument)

Step 1: Find related explorations

If argument is a PRO-NNN, read its entry in PROPOSALS.md and identify the topic and origin files.

If argument is a filename, read the named file from .elmer/proposals/. Extract its topic field from the elmer:archive metadata block.

Scan all other .md files in .elmer/proposals/:

• Check for exact topic match (same elmer topic string)
• Check for near topic match (3+ shared significant words in topic)
• Check for content overlap (>40% of referenced ADRs/DES sections are the same)

Report what was found:

Found N related exploration(s):
  → filename.md (topic match: exact|near|content, overlap: NN%)

If no related explorations found, report and proceed to create a standalone PRO-NNN entry.

Step 2: Structural alignment

For each section present in any source file (Summary, Analysis, Proposed Changes, Open Questions, What's Not Being Asked), compare across all sources:

• Shared ideas: Present in 2+ sources (strengthened signal)
• Unique to source A: Ideas only in one exploration
• Unique to source B: Ideas only in the other
• Contradictions: Claims or recommendations that conflict

Present the alignment table to the user.

Step 3: Synthesize and create PRO entry

On approval, produce:

1. A PRO-NNN entry in PROPOSALS.md with:

   • Next available PRO number (after current max in PROPOSALS.md index)
   • Status: Proposed
   • Type: Feature/Theme/Policy/Enhancement (as classified)
   • Governing Refs: ADR/DES identifiers mentioned in the explorations
   • A curated summary (scheduling-focused, not raw exploration prose)
   • Origin: list of source exploration filenames

2. Update the PROPOSALS.md index table with the new entry.

3. Update ROADMAP.md § Unscheduled Features — add a row to the "Proposed — Awaiting Evaluation" table with the PRO-NNN reference.

4. Archive source explorations — move superseded files to .elmer/proposals/archived/ (create directory if needed). Add a synthesis note to the archived files:

<!-- Consolidated into PRO-NNN in PROPOSALS.md on [date] -->

Step 4: Report

Report the synthesis: PRO-NNN assigned, word counts, unique ideas preserved, contradictions flagged, files archived.

---

Quality Standards

• Lossless on ideas, lossy on phrasing. Every distinct insight from every source must appear in the PRO entry or be noted as a cross-reference. Redundant prose is collapsed.
• Parallax is signal. When two explorations reach the same conclusion independently, that's stronger evidence than either alone. Mark it.
• Don't resolve tensions. If Source A says "defer Neptune" and Source B says "keep Neptune but simplify," present both. Synthesis combines perspectives — it doesn't make decisions.
• Preserve the "What's Not Being Asked" sections fully. These contain the highest-value insights and are the most likely to differ between explorations.
• PRO entries are scheduling-focused. The PRO body in PROPOSALS.md captures: what is proposed, what type it is, what it depends on, what ADR/DES it relates to, and when to re-evaluate. It does not duplicate the full exploration analysis.

Output Management

Present the structural alignment before writing anything. The user approves, edits, or rejects the synthesis plan before PRO entries are created.

If multiple clusters need deduplication, segment into groups. Present each cluster's synthesis plan for approval. After each approved synthesis is executed, proceed immediately to present the next cluster. Continue until all clusters are processed. State the total count when complete.