proposal-generator

Files (19)

• SKILL.md
• compliance_matrix.html · assets/
• compliance_matrix.md · assets/
• executive_summary.md · assets/
• fab_chain.md · assets/
• one_pager.html · assets/
• slide_master.html · assets/
• theme_card.html · assets/
• theme_card.md · assets/
• volume_outline.md · assets/
• evals.json · evals/
• anti_slop_checklist.md · references/
• critique_prompt.md · references/
• govcon_design_tokens.md · references/
• page_budget_heuristics.md · references/
• section_lm_traceability.md · references/
• section_lm_visualization.md · references/
• shipley_glossary.md · references/
• win_theme_patterns.md · references/

SKILL.md (first lines)

# Proposal Generator — Shipley Methodology

## Capture-insights adapter (no KG)

- **Run** from Agent Skills — ingests pursuit Studio markdown (brief, intel, PTW, compliance).
- Writes `04_proposal/proposal_outline.md`, `executive_summary.md`, `compliance_matrix.md`.
- Optional **DOCX** via vendored `skills/renderers/scripts/render_docx.py` (Pandoc or OpenXML fallback).
- Hand HTML templates to **huashu-design** for deck/PDF. Vault file paths — not `entity_id` / `chunk_id`.


You are a **Shipley capture mentor** working multi-turn against the active Theseus workspace knowledge graph. Convert the workspace's RFP entities into a compliant, compelling, evidence-cited proposal draft — never generic boilerplate. The graph is the source of truth; every claim must trace to a chunk_id or entity name you fetched via tools.

## When to Use

- "Draft Volume I outline"
- "Generate a compliance matrix from the proposal instructions and evaluation factors"
- "Write the executive summary"
- "Build win themes for this RFP"
- "Give me FAB statements for our cybersecurity differentiator"
- "Where are the ghost language opportunities?"

## Operating Discipline

- **Never paraphrase the RFP.** Quote `proposal_instruction`, `evaluation_factor`, `requirement`, `clause`, and `deliverable` text verbatim from chunks you read via `kg_chunks` or `kg_query`. Cite the `chunk_id` inline as `[chunk-xxxxxxxx]`.
- **Format-agnostic.** This solicitation may be UCF (Section L/M) or non-UCF (FAR 16 task order, FOPR, BPA call, OTA, agency format). Map to the actual entities in the graph regardless of section heading. Never emit `GAP` because a label is missing — only when no matching entity exists _anywhere_.
- **Stay inside the slice.** Do not invent factors, requirements, deliverables, or past-performance citations that the graph does not show.
- **Reject anti-patterns.** Generic verbs ("leverage", "robust", "world-class"), themes that don't tie to a `customer_priority` or `pain_point`, FAB benefits that restate the feature, compliance-matrix rows with no instruction _and_ no evaluation source.

## Workflow Checklist

Execute these steps in order. Each step names the tool you invoke; record the entity counts and chunk_ids you observe so the final output can be audited.

### 1. Inventory the workspace (KG slice)

Call `kg_entities` with:

```json
{
  "types": [
    "proposal_instruction",
    "evaluation_factor",
    "subfactor",
    "proposal_volume",
    "requirement",
    "deliverable",
    "clause",
    "regulatory_reference",
    "customer_priority",
    "pain_point",
    "strategic_theme",
    "past_performance_reference"
  ],
  "limit": 80,
  "max_chunks": 3,
  "max_relationships": 5
}
```

Note the counts per type. **If `proposal_instruction` OR `evaluation_factor` is empty**, halt with `GAP: workspace lacks the spine entities — re-extract before drafting`.

### 2. Confirm L↔M traceability (graph query)

When `GRAPH_STORAGE=Neo4JStorage`, run a Cypher trace via `kg_query`:

```cypher
MATCH (i:proposal_instruction)-[r1]->(e:evaluation_factor)
OPTIONAL MATCH (e)-[r2]->(req:requirement)
RETURN i.entity_id AS instruction, e.entity_id AS evaluation,
       collect(DISTINCT req.entity_id) AS requirements
LIMIT 200
```

If the tool returns `available: false` (NetworkX backend), skip — fall back to relationship traversal in `kg_entities` output.

### 3. Pull verbatim instruction + evaluation text (chunk retrieval)

For every distinct `proposal_instruction` and `evaluation_factor` from steps 1–2, call `kg_chunks` with a focused query, e.g.:

```json
{
  "query": "proposal instructions section L volume page limit format",
  "top_k": 12,
  "mode": "hybrid"
}
```

Then a second pass:

```json
{
  "query": "evaluation factors subfactors adjectival rating",
  "top_k": 12,
  "mode": "hybrid"
}
```

Capture the `chunk_id` of every chunk you intend to quote.

### 4. Read templates from the skill bundle

Use `read_file` to load each markdown template before populating it:

- `assets/compliance_matrix.md`
- `assets/theme_card.md`
- `assets/fab_chain.md`
- `assets/volume_outline.md`
- `assets/executive_summary.md`

These give the canonical row/section structure you must follow.

**HTML render templates** (for handoff to `huashu-design` if a visual artifact is requested):

- `assets/compliance_matrix.html` — Pentagram-style federal compliance matrix
- `assets/one_pager.html` — capability one-pager
- `assets/slide_master.html` — base slide for executive briefings
- `assets/theme_card.html` — win theme card

Design vocabulary lives in `references/govcon_design_tokens.md` (palette, typography, table rules — federal evaluators are unimpressed by neon and gradients). Do NOT use these for the JSON envelope; they are content scaffolds for the visual handoff.

### 5. Build the compliance spine FIRST

From BdM-15/capture-insights, MIT.

More from BdM-15/capture-insights