meta-competitive-intel

# meta-competitive-intel

Connector / competitive-intel meta-skill. Monitors one to a handful of
target companies across pricing, product, leadership, hiring, and news
dimensions; extracts grounded signals; classifies a verdict; produces
a brief with concrete actions. Always read-only — never sends emails,
posts, or modifies tracker data.

## Composition philosophy — multi-skill bundled orchestration

This meta-skill uses **only OpenSquilla-bundled atomic skills** plus
the five built-in step kinds — no external dependencies. The DAG calls
into **5 distinct bundled atomic skills**:

| Skill | Step(s) | Role in the DAG |
|---|---|---|
| `multi-search-engine` | `web_research`, `web_research_target_1..3`, `web_research_retry` | Primary, target-level, and fallback research source, fed by LLM-generated short search queries rather than raw target grids |
| `deep-research` | `deep_dive` | Extra rounds for `SINGLE_DEEP` mode only |
| `memory` | `recall_baseline`, `store_brief` | Cross-session continuity. `recall_baseline` pulls the last brief from durable memory automatically so `baseline_diff` works even when the user didn't paste a baseline; `store_brief` writes this run for next time. **This pair effectively delivers what the proposed `state:` primitive would give us.** |
| `xlsx` | `signals_xlsx` | When the user explicitly asks for a spreadsheet / `xlsx` / download / export, export the signal table + baseline diff as a workbook |
| `docx` | `export_docx` | Optional final DOCX export |

Step kinds used: `llm_chat`, `llm_classify`, `user_input`, `skill_exec`,
`agent`.

Before search, `search_strategy` uses the LLM to translate structured intel
context into compact `SEARCH_QUERY` lines with target aliases and product
names. The flow then adds focused target-specific searches for the first
three monitored targets, so a broad competitive-intel request does not depend
on one generic multi-company result. This keeps generalized inputs such as
Chinese company names, baseline-diff requests, and "all dimensions" replies
from becoming brittle YAML-like search strings.

Direct URL fetching of competitor pages is intentionally out of scope
for this skill — the search results from `web_research` are normally
enough, and adding URL-scrape would require a compliance-aware fetcher
not currently bundled. If the user wants page-level detail, they can
paste the page content into the next turn.

Persistence to Notion / external knowledge bases is also out of scope:
the deliverable is the markdown emitted by `deliver_intel_brief`; the
user copies it wherever they want.

## Mode design

Four depth labels via `llm_classify: depth`:

- `SINGLE_DEEP` — one target, multi-round deep-research, full
  signal table, full actions. Best for "tell me everything about
  $competitor's last quarter."
- `MULTI_QUICK` — 2-5 accounts, scan-level coverage, top-10 signal
  table. Best for "what did these 5 do this month?"
- `DIFF_VS_BASELINE` — baseline text was pasted, the run leads with
  the diff section. Best for "what's new since I last looked?"
- `EXEC_BRIEF` — 5-bullet executive summary + verdict + top 3
  actions, nothing else. Best for "give me something to forward to
  the CEO."

## Honest limitations (first-wave)

- **No `state:` primitive.** Baselines are pasted each turn; the
  proposed `state:` would persist last run's signals automatically.
- **No `foreach`.** The first three targets get focused search lanes; beyond
  that, remaining targets share the broad query context. With `foreach`, each
  target would get its own isolated step + audit trail per target.
- **No alerting.** The skill produces a brief on demand; it doesn't
  push notifications. A future combination with `cron` (bundled) +
  the proposed `event_trigger` primitive would give a "monitor this
  every Monday" mode.