Primr Research Skill

Conceptual Framework

This skill is a thin orchestrator over Primr MCP.

Use MCP resources to discover current behavior instead of assuming fixed costs, providers, or defaults from the skill text. full is the standard end-to-end workflow, premium is the higher-depth option, and scrape and deep are narrower research modes.

Research runs are async jobs. Start the run, then monitor status until terminal. Standard runs often take 35-50 minutes; premium multi-vendor runs can take 75-120 minutes.

Operational Capabilities

1. Select a mode

• Read primr://research/modes before advising on mode selection.

• Use scrape for first-party reconnaissance.

• Use deep when the site is blocked or low-signal.

• Use full for the default strategic analysis workflow.

• Use premium when the user explicitly wants maximum depth.

2. Estimate before execution

Always call estimate_run before research_company, then state that the run incurs real API cost and wait for explicit user approval.

estimate_run(company_url="https://example.com", mode="full")

3. Start and monitor the job

After user approval, call research_company and pass the approved max_estimated_cost_usd when available, then monitor with wait_for_status_change, check_jobs, or primr://research/status. Do not assume the client session will stay attached for the entire run.

research_company(company_name="ExampleCo", company_url="https://example.com", mode="full")

4. Retrieve results

When the run completes, read primr://output/artifacts/by_job/{job_id} first to inventory artifacts without report body content. Read primr://output/qa_summary/by_job/{job_id} when QA artifacts are attached. Read primr://output/usage_summary/by_job/{job_id} when cost, timing, approval, or artifact-count metadata is needed. Read primr://output/source_summary/by_job/{job_id} when citation/source appendix metadata is needed. Read primr://output/trace_summary/by_job/{job_id} when scrape trace health metadata is needed. Read primr://output/verification_summary/by_job/{job_id} when claim verification trust score, claim counts, status counts, first-party downgrade counts, or source-reference counts are needed without raw claims, source URLs, search queries, explanations, or report body content. Read primr://output/calibration_summary/by_job/{job_id} when label-calibration counts, inference source-copy counts, evidence-review count buckets, judge provenance, or judge-agreement metadata is needed without raw claims, source URLs, evidence reviews, rationales, or report body content. MCP resource reads are audit-logged with hashed URI/result values and normalized resource kind, without raw URI query values or resource bodies. A2A skill calls are audit-logged with hashed message/result payloads, hashed caller ids, granted scopes, duration, outcome, and job id when present, without raw message text, task ids, URLs, report paths, raw results, or caller ids. Read primr://output/latest only when the next step needs a report preview.

Error Handling

• job_in_progress: report the existing run and monitor it instead of starting another.

• invalid_url or ssrf_blocked: require a valid public URL, or consider deep mode.

• rate_limit_exceeded: wait for the retry window.

• api_error: surface the provider error clearly and recommend retry or diagnostics.