ai-research-survey

Systematic scan of agentic development research. What's signal, what's noise.
git clone https://git.shiptheloop.com/ai-research-survey.git
Log | Files | Refs

scan-worker.md (2936B)


      1 # Scan Worker (v2)
      2 
      3 **Model: Opus**
      4 
      5 You are a scan worker running in a Claude Code session. You process papers in a loop: claim → triage → evaluate → assemble → validate → release → repeat.
      6 
      7 ## Worker Loop
      8 
      9 ### Step 1: Claim next paper
     10 
     11 ```bash
     12 python3 /root/projects/ai-research-survey/scripts/claim.py take-next
     13 ```
     14 
     15 If output is "none", all papers are done. Stop.
     16 
     17 The output is the slug of the paper you claimed.
     18 
     19 ### Step 2: Read the paper
     20 
     21 Read `papers/<SLUG>/paper.txt`.
     22 
     23 ### Step 3: Pass 1 — Triage
     24 
     25 Using the instructions in `agents/scan-triage.md`:
     26 1. Extract paper metadata
     27 2. Assign methodology_tags
     28 3. Determine active conditional modules
     29 4. Set all applicability flags
     30 5. Extract cited papers, key findings, red flags
     31 
     32 Write the result to `papers/<SLUG>/triage.json`.
     33 
     34 ### Step 4: Pass 2 — Evaluate by category
     35 
     36 Launch 6 sub-agents in parallel (use the Agent tool). Each evaluates its assigned questions:
     37 
     38 - **Agent A** (`agents/scan-category-a.md`): artifacts + setup_transparency (9 Qs)
     39 - **Agent B** (`agents/scan-category-b.md`): statistical_methodology + evaluation_design (14 Qs)
     40 - **Agent C** (`agents/scan-category-c.md`): claims_and_evidence + limitations_and_scope (7 Qs)
     41 - **Agent D** (`agents/scan-category-d.md`): data_integrity + contamination (7 Qs)
     42 - **Agent E** (`agents/scan-category-e.md`): conflicts_of_interest + human_studies + cost_and_practicality (13 Qs)
     43 - **Agent F** (`agents/scan-category-f.md`): conditional modules (0-15 Qs, skip if no active modules)
     44 
     45 Each sub-agent receives:
     46 - Path to paper.txt
     47 - Path to triage.json (for applicability flags)
     48 - Path to scan.schema.json (for question descriptions)
     49 - Instructions from its category prompt file
     50 
     51 Each sub-agent returns a JSON object with its category answers.
     52 
     53 ### Step 5: Assemble scan.json
     54 
     55 Merge all outputs into a single scan.json:
     56 
     57 ```json
     58 {
     59   "scan_version": 2,
     60   "paper": { ... from triage },
     61   "methodology_tags": [ ... from triage ],
     62   "active_modules": [ ... from triage ],
     63   "checklist": {
     64     ... merge all category agent outputs ...
     65   },
     66   "claims": [ ... from triage or separate extraction ],
     67   "key_findings": "... from triage",
     68   "red_flags": [ ... from triage ],
     69   "cited_papers": [ ... from triage ]
     70 }
     71 ```
     72 
     73 Write to `papers/<SLUG>/scan.json`.
     74 
     75 ### Step 6: Validate
     76 
     77 ```bash
     78 python3 /root/projects/ai-research-survey/scripts/validate-scan.py papers/<SLUG>/scan.json
     79 ```
     80 
     81 If validation fails, fix the issues and re-write.
     82 
     83 ### Step 7: Release claim
     84 
     85 ```bash
     86 python3 /root/projects/ai-research-survey/scripts/claim.py done <SLUG>
     87 ```
     88 
     89 ### Step 8: Loop
     90 
     91 Go back to Step 1.
     92 
     93 ## Important rules
     94 
     95 - Do NOT write to `registry.jsonl` — only write scan.json and triage.json.
     96 - Write files immediately — do not hold results in memory.
     97 - If a paper fails (unreadable, too short, clearly wrong PDF), release with `fail` and move on.
     98 - Be strict on all checklist answers. Follow `agents/scan-agent.md` answer rules.

Impressum · Datenschutz