Gaj Sherlock

Process

Step 1: Accept and parse input

The user provides any combination of:

• A URL (LinkedIn job, LinkedIn profile, company website)
• Pasted job description text
• A company name
• A recruiter name
• Any mix of the above

Parse rules:

When inputs overlap (URL + pasted JD), merge. URL-sourced facts take precedence. Pasted text fills gaps.

Step 2: Check pipeline (optional)

Search the pipeline for the company:

npx tsx scripts/pipeline-cli.ts search '<company name>'

If a match exists:

• Note the job ID for later storage
• Check for prior sherlock findings:

npx tsx scripts/pipeline-cli.ts get-job <id>

If job_data.sherlock.investigated_at exists and is < 24 hours old, ask: "I have recent findings for [company]. Re-investigate or show existing?"

If the user wants existing findings, present the stored report and stop.

Step 3: Run investigation (parallel subagents)

Read @prompts/sherlock-system.md for the full investigation framework.

Launch up to 5 parallel subagents, one per investigation dimension. Each subagent gets:

• The parsed input data relevant to its dimension
• The investigation instructions from sherlock-system.md for its dimension
• Access to WebSearch and WebFetch tools

Subagent dispatch:

1. Recruiter analysis - only if recruiter name or firm name available
2. Company analysis - always (company name extractable from most inputs)
3. Mystery client resolution - only if posting is from a staffing firm
4. Compensation reality check - only if comp data is stated or extractable
5. Red flag detection - always, runs after dimensions 1-4 return

Read ~/gaj/context/about-me.md for the user's salary floor, tech stack, and role targets. Pass these to the comp and stack fit investigators.

Step 4: Assemble the report

Collect findings from all subagents. Follow the narrative output template from @prompts/sherlock-system.md exactly. Omit sections with no data.

Apply @prompts/writing-rules.md to all narrative output.

End the report with a Sources section listing all URLs used as markdown links.

Step 5: Store findings (silent, if pipeline-linked)

If a pipeline entry was found in Step 2, store the structured findings:

npx tsx scripts/pipeline-cli.ts update <id> 'job_data' '<JSON string of sherlock findings>'

Use the JSON schema from @prompts/sherlock-system.md.

If no pipeline entry exists, offer at the end of the report: "Want me to add this to your pipeline?"

If the user accepts, create the entry:

npx tsx scripts/pipeline-cli.ts add '{"company_name":"<company>","job_title":"<role>","status":"pending-review","source":"sherlock","job_data":<findings JSON>}'

Internal invocation (from respond)

When called by /gaj:respond, sherlock runs the same process but:

• Does not display the banner or check pipeline (respond already did that)
• Returns findings to the calling skill rather than presenting to user
• Respond presents a condensed version of the report as part of its own flow
• Storage happens through respond's pipeline entry

CLI reference

npx tsx scripts/pipeline-cli.ts search '<company>'
npx tsx scripts/pipeline-cli.ts get-job <id>
npx tsx scripts/pipeline-cli.ts update <id> 'job_data' '<json>'
npx tsx scripts/pipeline-cli.ts add '<json>'