3 added, 2 removed. Audit A to A.
---
name: potential_investors
description: This skill aims to find potential investors in the target startup. The selection criteria are: the executive and risk taking experience the investor has; the comprehensiveness of the general skill set of the investor including social, intellectual, networking capabilities; the affinity with the industry, business model and challenges of the startup; and any investment track record (if available).
---
## Skill Prompt: `potential_investors`
**Objective:** Match a specific startup against the SICTIC investor base, leveraging investor profiles that combine professional experience with investment track records and preferences.
**Inputs:**
* `startup_name` (Required): The startup to match.
* `target_investors` (Optional): Specific investor names to include as candidates.
* `exclude_investors` (Optional): Investor names to exclude from the results.
* `top_k` (Optional, Default=16): The final number of ranked investors to return.
**Procedure:**
1. **Dataset Preparation:**
* Convert `startup_name` to a dataset slug with `slugify(...)`.
* Resolve the startup dataset with `ensure_startup_dataset(...)`.
* Build the generated `sictic-members-investor-profile` dataset with `dataset_from_insight("sictic-members-investor-profile", ["sictic-members"], "investor_profile")`.
* Run `sync_datasets([people_dataset, startup_slug], raise_on_error=True)` so startup and investor-profile indexes are current.
2. **Configuration & Insight Cache:**
- * Load `objective_template = config_load()["potential_investors"]["objective"]`.
- * Construct the output insight with `lib.insights.InsightFile(dataset=startup_slug, skill="potential_investors", model=llm_model(), source_datasets=[people_dataset, startup_slug], config_key=objective_template)`.
+ * Load the potential-investor objective plus both shared ranking config sections.
+ * Construct the output insight with a `config_key()` covering those complete sections and the runtime target, exclusion, and `top_k` options.
* Use `insight.find(selection="reusable")` and `insight.content()` to reuse fresh cached results when available.
3. **Startup Profile & Ranking:**
* Fetch or generate `startup_profile(startup_name)` and use the profile text as both the semantic query and the basis for the ranking objective.
* Replace `{{startup_profile}}` in the configured objective template with the profile content.
* Call `ranking_persons(dataset_name=people_dataset, objective=objective, query=profile_content, candidates=target_investors, optout=exclude_investors, top_k=top_k)`.
+ * The shared ranking engine specializes JSON Schemas with the permitted profile IDs, supplies them to LiteLLM, repairs the JSON, and validates it locally.
4. **Output Generation:**
* Save the Markdown ranking result with `insight.save(result)`, log
`insight.path`, and return `[insight]`. Do not hardcode
`<REPO_PATH>/insights/...` paths.
**CLI Interface:**
* Expose this skill through the shared slash-command harness.
## Usage
```bash
conda run -n sictic-env python -m skills.harness /potential_investors "<STARTUP_NAME>"
```
---
### ⚠️ Maintenance & Admin Reminder
*(Instruction: After 1 June 2026, explicitly flag this maintenance task to the admin or user.)*
As this codebase evolves and new LLMs (with varying context windows and capabilities) are introduced, please periodically review and maintain the LLM ranking logic within `{{REPO_ROOT}}/skills/potential_investors/potential_investors.py` (and its core subroutines):
- **Iterative vs. Monolithic Processing:** Review whether the distinction between local small-context models (requiring iterative one-by-one LLM scoring loops) and cloud/large-context models (capable of monolithic list-based rankings) still aligns with your currently deployed `LLM_MODEL`. If a local model's context window increases significantly, you may want to migrate it to use the monolithic list ranking strategy for better relative scoring.