competitive-intel

Files (4)

• SKILL.md
• evals.json · evals/
• black_hat_template.md · references/
• data_sources.md · references/

SKILL.md (first lines)

# Competitive Intel — Black-Hat + Obligation Intel

## Capture-insights adapter (no KG / no VDB)

**Run** from **Agent Skills** or chat ("burn rate on contract X") — no CLI.

- **Workflow B/C (active):** Deterministic USAspending collector via `usaspending-gov-mcp`. Paginates IDV child orders per Settings caps. Full JSON at `02_intel/competitive_intel_obligation.json`; brief at `02_intel/competitive_intel.md`.
- **Context:** Pipeline row `award_key`, pursuit workspace, or PIID typed in chat.
- **Workflow A (black-hat):** Later slice — uses pursuit vault files + DuckDB + MCP, not a knowledge graph.

Every claim must trace to a **vault file path**, **DuckDB citation**, or a **USAspending award** from MCP — never invented entities.

You are a **federal capture analyst** working against the active **pursuit workspace** (Studio markdown + Pipeline row) PLUS live USASpending.gov award data (via the `usaspending` MCP). This skill has three modes:

- **Workflow A (Black-Hat)**: Convert pursuit/RFP context into a defensible black-hat brief on incumbent and likely competitors. *(PR3 — vault context bundle, not KG.)*
- **Workflow B (Vehicle Obligation Intel)**: Start from a contract number and build a parent-level obligation rollup, child-order accumulation, sibling-awardee context, and vehicle-wide insights.
- **Workflow C (Single Award / Order Burn Intel)**: Start from one contract or order number and produce a clean by-award burn read with grouped modifications, POP windows, and localized insights.

## When to Use

- "Who's the incumbent on this contract?"
- "Pull award history for similar work in the last 5 years"
- "Black-hat: what would Lockheed bid here?"
- "What's the typical price range for this NAICS at this magnitude?"
- "Top 3 competitors for a 541512 IT modernization recompete with HHS"
- "Here's a contract number - show burn rate / obligation trend"
- "Is this contract an IDIQ parent, an order, or standalone?"
- "Roll up obligations across all child orders"
- "If this is an order under a multiple-award IDIQ, who are all parent-level competitors?"

## Operating Discipline

- **Workspace first, web second.** The KG is the source of truth for the opportunity (program, agency, NAICS, PSC, place of performance). USAspending is the source of truth for who has actually won similar work.
- **Cite everything.** Workspace claims cite `chunk_id` (e.g., `[chunk-xxxxxxxx]`). Award claims cite the USAspending `generated_internal_id` (e.g., `[award:CONT_AWD_…]`).
- **Rank by obligated dollars, not transaction count.** A vendor with one $400M IDV beats a vendor with 200 $50K POs.
- **Linkage precedence for vehicle analysis.** For multiple-award IDIQs, resolve relationships in this order: parent IDV linkage, explicit child-order listing, PIID-family context, then solicitation identifier as a fallback signal.
- **Reject anti-patterns.** Inventing competitors not in award history. Pricing ranges with no underlying award sample. Generic SWOT bullets ("strong past performance"). Unsupported claims about CPARS or protests (no MCP source for those yet — say so).
- **Treat deterministic rosters as authoritative.** If `parent_vehicle_awardees` is present in the collector result or artifact, that is the complete parent-level prime roster for the answer. Do not add extra vehicle holders from memory, loose keyword matches, or workspace hints. The number you state in prose must equal `parent_vehicle_awardee_count` / `len(parent_vehicle_awardees)`.
- **Fail loudly.** If the workspace is missing NAICS / PSC / agency, halt with a `GAP` rather than guess.

## Workflow Selector

Pick exactly one workflow before running tools:

- **Workflow A: Black-Hat Competitor Intel (default)**
  Trigger: incumbent/competitor/theming/pricing benchmark questions anchored on NAICS/agency/program context.
- **Workflow B: Vehicle Obligation Intel**
  Trigger: user provides a contract number/PIID and asks for parent-level burn-rate, obligation trend, IDIQ hierarchy, child-order rollups, sibling awardees, vehicle competitors, or an overarching view of how the whole vehicle is burning.
- **Workflow C: Single Award / Order Burn Intel**
  Trigger: user provides one contract or order number and primarily wants that one award's mods, POP, burn-rate, option pattern, or a clean single-award read rather than a vehicle aggregation.

If both black-hat and obligation analysis are requested, run Workflow B or C first and feed its `competitor_discovery` result into Workflow A.

If the user wants both parent-level vehicle context and one-order detail, choose **Workflow B** and use the collector's grouped `award_rollups` / artifact `obligations.by_award` to write the order-specific portion. Do not narrate every order modification in the parent-level summary.

---

## Workflow A Checklist (Black-Hat)

Execute in order. Record entity counts, NAICS codes, and award IDs as you go so the final envelope can be audited.

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

Call `kg_entities` with:

```json
{
  "types": [
    "program",
    "organization",
    "contract_line_item",
    "regulatory_reference",
    "requirement",
    "deliverable"
  ],
  "limit": 60,
  "max_chunks": 2,
  "max_relationships": 3
}
```

Extract:

- **Agency / sub-agency** (look in `organization` entities; cross-check `program`)
- **NAICS code** (often a `regulatory_reference` like "NAICS 541512")
- **PSC code** (similar — `regulatory_reference` like "PSC R499")
- **Place of performance** (state / city if present)
- **Estimated magnitude** (search `contract_line_item` or chunks for "$", "ceiling", "estimated value")

If NAICS AND PSC are both missing, halt with `GAP: workspace lacks NAICS and PSC — cannot scope award history`. If only one is missing, proceed but flag it in `warnings`.

### 2. Resolve fuzzy codes via USAspending autocompletes

If the workspace mentions a NAICS by description but not by number (e.g., "custom computer programming"), call `mcp__usaspending__autocomplete_naics` with the keyword. If the agency is a friendly name ("Air Force") but you need the toptier code, call `mcp__usaspending__list_toptier_agencies` once and resolve. Same drill for PSC via `mcp__usaspending__autocomplete_psc` or `mcp__usaspending__get_psc_filter_tree`.

### 3. Pull award history (the spine)

Call `mcp__usaspending__search_awards` with a filters object scoped to the resolved NAICS, agency, and a 5-year window:

```json
{
  "filters": {
    "naics_codes": ["<resolved naics>"],
    "agencies": [
      { "type": "awarding", "tier": "toptier", "name": "<agency name>" }
    ],
    "time_period": [{ "start_date": "<5y ago>", "end_date": "<today>" }],
    "award_type_codes": ["A", "B", "C", "D"]
  },
  "fields": [
    "Award ID",
    "Recipient Name",
    "Award Amount",
    "Period of Performance Start Date",
    "Period of Performance Current End Date",
    "Awarding Agency",
    "Awarding Sub Agency",
    "Award Type"
  ],
  "page": 1,

From BdM-15/capture-insights, MIT.

More from BdM-15/capture-insights