git:20260907.65e4374 to git:20261020.2bc152e

46 added, 18 removed. Audit A to A.

---

name: startup_profile

- description: Generate a neutral five-part startup diagnostic from dataset evidence or supplied files. Use to explain the business, technology, market and current challenges.

+ description: Generate a startup diagnostic with sector from a named dataset, including a prepared temporary dossier. Use to explain the business, technology, market and current challenges.

---

# Startup profile

Describe what the startup does and the structural questions requiring review.

## Inputs and outputs

- The async `startup_profile(startup, files=None)` returns one

+ The async `startup_profile(startup)` returns one

`InsightFile` in a list, named `startup-profile-<startup>-<model>.md`.

- The configured framework covers oneliner, industry, technology, business model

- and current challenges; keep detailed assessment instructions in configuration.

+ The fifth bullet, **Classification**, is appended directly from the Dealum result

+ in the startup_sector JSON insight, using resolved display names. Its sub-bullets

+ are **Primary-industry**, **Secondary-industries**, **Primary-technology** and

+ **Secondary-technologies**, preserving the selection order from startup_sector.

+ The current Dealum reference has no subindustry hierarchy.

+ Unresolved primary selections are shown as insufficient evidence; empty secondary

+ lists are shown as "None selected".

+ Legacy manual SICTIC classifications retain

+ their explicit SICTIC label; they are not translated into Dealum industries.

+ Product type is owned by dd_checks and is not added to this profile.

+ The configured LLM framework covers oneliner, technology, business model and

+ current challenges; it does not classify the startup.

## Workflow and dependencies

- Prepare the startup dataset and synchronize it, even when files are supplied.

- With no files, select a manual override or reusable generated report using indexed

- startup revisions and the configured query/instructions.

+ Prepare and synchronize the named startup dataset, then select a manual override

+ through InsightFile. If no manual profile exists, call startup_sector's existing

+ `startup_sector_as_dict` adapter so edited manual category IDs receive current

+ display names and derived fields without rewriting the source file.

+ Include the Dealum record, rendered classification bullet and profile

+ configuration in config_key before checking generated-profile reuse; InsightFile

+ owns hashing and freshness. Classification freshness belongs to startup_sector.

+ Changes to the other providers alone do not invalidate this profile. For a

+ legacy manual SICTIC classification, include that legacy classification instead.

+ Retired product-type fields in older manual classification files are ignored

+ for profile rendering and freshness, without modifying those files.

+ Existing generated datasets bypass startup preparation and external imports;

+ their profiles are saved within that generated dossier.

- With files, prepare the existing ephemeral dataset and analyze those documents.

- This path bypasses output reuse and still saves under the startup's normal

- profile identity. Its cache metadata does not include the supplied file content;

- do not interpret that metadata as proof of file-input freshness.

+ For supplied documents, follow the shared

+ [temporary-dataset lifecycle](../standards_and_architecture/SKILL.md#temporary-datasets)

+ and pass the prepared dataset name. Add further documents for the same startup

+ to its existing `datasets/` folder without resetting it. The skill accepts no

+ file-input argument.

- Use `dataset_chat` with the configured questions and prompt. The registry has

- no profile prerequisite; direct preparation may import Dealum source data.

+ Use `dataset_chat` with the four configured questions and prompt, then append

+ the fifth classification bullet in Python (or the legacy SICTIC sector, with

+ technologies explicitly unavailable). This adds no LLM call and no separate header.

+ The registry declares

+ startup-sector as a prerequisite; direct calls invoke the same public skill.

+ Direct preparation may import Dealum source data.

## Side effects and failure behavior

Preparation and retrieval may convert/index documents; generation calls a model

- and saves a report. Supplied files can replace the current model's startup report.

- Technical failures and empty generated output raise before saving.

+ and saves a report within the named dossier.

+ Technical failures (including classification failures) and empty generated output

+ raise before saving. A direct call returns an existing manual profile without

+ calling the classification skill.

- A nonempty insufficient-context sentinel can be saved as the profile; it is not

+ The insufficient-context sentinel is saved unchanged, without classification

+ labels, to preserve sentinel detection. It is not

a successful substantive diagnosis. Consumers can inspect

`InsightFile.has_insufficient_context()`.

## Usage

```bash

conda run -n sictic-env python -m skills.harness '/startup_profile "<STARTUP>"'

```

- The direct CLI accepts comma-separated `--startups` names and repeated `--files`

- options. It continues after individual startup failures, prints successful

+ The direct CLI accepts comma-separated `--startups` dataset names. It continues

+ after individual startup failures, prints successful

reports, then exits with code 1 if any failed. The harness handles one startup.

`--startup` remains a direct CLI alias for `--startups`.

## References

- [Implementation](startup_profile.py), [direct CLI](__main__.py)

- [Configuration](../../config/startup_profile/)

- [Dataset chat](../dataset_chat/SKILL.md)