dialectic

name	dialectic
description	Run a DIALECTIC investment analysis on any startup. Asks for company info (auto-filled via website scrape), optionally extracts founder data from LinkedIn screenshots, runs a multi-agent LLM pipeline (decomposition → argument generation → devil's advocate critique → refinement × 2 iterations), and presents pro/con arguments with an investment recommendation. Based on the EACL 2026 research paper.

You are running the DIALECTIC investment analysis pipeline. The base directory for this skill is provided in your invocation context as Base directory for this skill: <path>. Read that path and store it as SKILL_BASE_DIR — you will use it directly when running the pipeline.

Follow these steps exactly.

Step 1: Collect Company Information

1a. Ask for website first

Ask:

"What's the startup's website? I'll auto-fill the company info from it. (or type skip to enter details manually)"

If the user provides a URL:

Use the WebFetch tool to fetch the page. Extract the following fields from the content:

• name — the company name
• about — what the product or service does (1-2 sentences)
• industry — the sector or domain (e.g. FinTech, HealthTech, B2B SaaS)
• tagline — a short marketing phrase if present (optional)

Content quality check: If the fetched content is empty, contains only navigation/footer/cookie-consent text, or has fewer than 50 characters of substantive company description, tell the user the site couldn't be read and fall through to manual entry (Step 1b).

If scrape succeeds, display a confirmation card:

"Here's what I found — does this look right? Edit anything below or type yes to confirm:

• Name: [extracted name]
• What it does: [extracted about]
• Industry: [extracted industry]
• Tagline: [extracted tagline or 'none']"

Wait for the user's response:

• If they type yes (or confirm): store the values and proceed to Step 2.
• If they correct any field (e.g. "industry is HealthTech"): update that field, re-display the full card, and ask again. Loop until confirmed.

1b. Manual entry fallback

Use this path when: (a) the user types skip, or (b) the scrape returned empty/useless content.

Ask for each field, one at a time (or accept all at once if already provided):

1. Startup name — What is the company called?
2. What it does — Describe the product or service in 1-2 sentences.
3. Industry — What sector or industry? (e.g. FinTech, HealthTech, B2B SaaS)

After collecting all three, display the same confirmation card format as above and wait for confirmation or corrections before proceeding.

Step 2: Collect Team Information

After company info is confirmed, ask:

"Want to add founder/team info? This powers the team evaluation dimension.

1. Upload LinkedIn screenshot(s)
2. Enter manually
3. Skip"

Option 1: LinkedIn screenshots

Ask the user to upload the first LinkedIn profile screenshot.

For each screenshot uploaded:

Read the image using your vision capability and extract the following fields into a Person record:

• name — full name from the profile header
• city — city of residence if shown
• country_code — two-letter country code if location is shown (e.g. "US", "DE")
• about — the bio/summary section text if shown; if not shown, infer a 1-2 sentence summary from the headline, job titles, and experience visible in the screenshot
• education — list of entries, each with: institution name, start year, end year (mark as "not visible" if section not shown)
• experience — list of entries, each with: company, title, date range, description (mark as "not visible" if section not shown)
• followers — follower count if visible
• connections — connection count if visible

Mark fields not visible and not inferable as absent. Always infer about from available context rather than leaving it blank.

Display a confirmation card for each person:

"Founder: [name] Location: [city, country_code or 'not shown'] Education: [institution (years); institution (years); ...] Experience: [title at company (dates); title at company (dates); ...] About: "[about excerpt]" Followers: [count or 'not shown'] | Connections: [count or 'not shown']

Does this look right? Correct any field or type yes to confirm."

If the user corrects a field: update it, re-display the full card, and ask again. Loop until confirmed.

After confirming each person, ask:

"Another founder? Upload the next screenshot or type done."

Continue extracting and confirming until the user types done.

Final team summary: After all founders are confirmed, display:

"Team confirmed:

1. [name] — [title at most recent company], [education institution]
2. [name] — [title at most recent company], [education institution] ..."

Then proceed to Step 3.

Synthesizing team data for the pipeline:

Format each confirmed Person as follows (matching Person.get_profile_summary() text structure):

Name: [name]
Location: [city], [country_code]
About: [about]
Education:
- [institution] ([start_year]–[end_year])
Experience:
- [title] at [company] ([start_date]–[end_date]): [description]
Followers: [followers] | Connections: [connections]

If multiple founders, join their formatted summaries with --- on its own line between each person.

Pass the combined text as the --extra argument to the pipeline CLI.

Option 2: Manual team entry

Ask for each founder one at a time:

"Name and background for founder 1? (e.g. Alice Chen, Stanford CS 2019, ex-Google PM)"

After each entry, ask: "Another founder? Enter their info or type done."

Format each response as a Person summary (same structure as above, filling in what the user provided) and join with --- for the --extra argument.

Option 3: Skip

Proceed to Step 3 without any team data. Omit the --extra flag when running the pipeline.

Step 3: Run the Analysis

Warn the user first:

"This analysis takes up to 10 minutes to complete. The pipeline will decompose investment questions, search the web for information, generate pro/contra arguments, apply devil's advocate critiques, and refine the best arguments over 2 iterations. Starting now..."

Then run the pipeline using the Bash tool with a 10-minute timeout. Use SKILL_BASE_DIR from your invocation context directly — do not search the filesystem:

cd "[SKILL_BASE_DIR]" && uv run python -m agent.cli \
  --name "[STARTUP NAME]" \
  --about "[WHAT IT DOES]" \
  --industry "[INDUSTRY]" \
  --website "[WEBSITE URL IF PROVIDED]" \
  --extra "[TEAM SUMMARY TEXT IF COLLECTED]"

Omit --website if the user didn't provide a URL. Omit --extra if the user skipped team info.

The command outputs JSON to stdout. Capture it.

Step 4: Present Results

Formatting each argument: Extract a short bold headline (5–10 words capturing the core claim) from the argument text, then show the full content beneath it. Sort by score descending. Show top 3 pro and top 3 contra from final_arguments.

Present results in this format:

## DIALECTIC Analysis: [Company Name]

### Investment Arguments

**PRO** (reasons to invest)

1. [score/100] **[Bold 5–10 word headline]**
   [Full argument text — refined_content if non-null, otherwise content]

2. [score/100] **[Bold headline]**
   [Full argument text]

3. [score/100] **[Bold headline]**
   [Full argument text]

**CON** (reasons to pass)

1. [score/100] **[Bold 5–10 word headline]**
   [Full argument text]

2. [score/100] **[Bold headline]**
   [Full argument text]

3. [score/100] **[Bold headline]**
   [Full argument text]

### Recommendation
**[INVEST / PASS]**

[One sentence stating the decision and citing the single strongest argument verbatim or closely paraphrased. Do not argue against it or qualify it — just state it as the deciding factor.]

Step 5: Error Handling

If the Bash command exits with a non-zero status or outputs an error:

1. Show the error message clearly
2. Check if it mentions a missing API key — if so, say:

   "It looks like an API key is missing. Open the .env file in your DIALECTIC install directory and add:

   • OPENAI_API_KEY — from platform.openai.com
   • PPLX_API_KEY — from perplexity.ai (or BRAVE_SEARCH_API_KEY from brave.com/search/api)"

3. Otherwise, show the raw error and suggest re-running.