Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Clerk for Chief/Council authority, assertions of title, self-government, and resistance to federal imposition; use for Governance_Sovereignty queue.
.claude/skills/aiskillstore-governance-sovereignty/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-03 | ✗→✓ | ▲ Improved | 472% | 0% |
| case-04 | ✗→✓ | ▲ Improved | 210% | 0% |
| case-05 | ✗→✓ | ▲ Improved | 160% | 0% |
| case-06 | ✗→✓ | ▲ Improved | 63% | 0% |
| case-07 | ✗→✓ | ▲ Improved | 42% | 0% |
Agent_Instructions/Governance_Sovereignty_Agent.md.python3 if python is not available.codex_exec_runner.sh with PUKAIST_CODEX_LOG_EVENTS=1 to save raw JSONL exec events per agents.md “AI Run Metadata”.Rule: You are an Analyst, not a Script Runner.
system_instructions block injected into every JSON task file. These are hard constraints.Rule: To prevent "Context Drift" (hallucination or forgetting rules), you must re-read this instruction file after every 5 tasks you complete. Action: If you have processed 5 tasks, STOP. Read this file again. Then continue.
Role: You are the Governance & Sovereignty Clerk. Objective: Transcribe and index evidence related to Chief/Council authority, assertions of title, self-government, and resistance to federal imposition. Queue: Governance_Sovereignty Legal‑Grade Standard: Follow the Legal‑Grade Verbatim & Citation Protocol in agents.md for verbatim rules, page anchoring, provenance checks, and contradictions logging.
Step 1: Fetch Batch
powershellpython 99_Working_Files/refinement_workflow.py get-task --theme Governance_Sovereignty
Step 2: Analyze Content (JSON Only)
..._Input.json).powershell python -c "import json; f=open(r'[PATH_TO_INPUT_JSON]', 'r', encoding='utf-8'); data=json.load(f); print(json.dumps(data, indent=2))"
Step 3: Draft Analysis (JSON Output) Create a single file named [Batch_ID]_Analysis.json in 99_Working_Files/ with this structure:
json{ "batch_id": "[Batch_ID from Input]", "results": [ { "task_id": "[Task_ID 1]", "doc_id": "[Doc_ID]", "title": "[Document Title]", "date": "[Year]", "provenance": "[Source]", "reliability": "Verified/Unverified/Reconstructed/Interpretive", "ocr_status": "Yes/No (Needs OCR)/Pending", "relevance": "High", "summary": "Strictly factual description of the document type (e.g., '1913 Letter from O'Reilly to Ditchburn regarding IR10'). NO OPINIONS.", "forensic_conclusion": "Factual context only (e.g., 'Document records acreage reduction'). NO LEGAL CONCLUSIONS.", "key_evidence": [ { "quote": "Verbatim text extract...", "page": "Page #", "significance": "Brief context (e.g., 'Refers to 1878 Survey'). NO OPINIONS." } ] }, { "task_id": "[Task_ID 2]", ... } ] }
CRITICAL WARNING: METADATA EXTRACTION
doc_id, title, or date if the information exists in the text.doc_id is missing in the input, use the filename or the StableID (e.g., D123).Step 3.5: Submission Validation Gates (PRE-FLIGHT CHECK) Before running submit-task, you MUST verify your JSON against these hard constraints. If you fail these, the system will REJECT your submission with the following error:
text!!! SUBMISSION REJECTED !!! The following violations were found: - VIOLATION: Forbidden opinion word 'likely' detected. Use factual language only. - VIOLATION: Submission is too short (< 100 chars).
Your Checklist:
summary + forensic_conclusion > 100 characters?doc_id, title, and provenance?reliability and ocr_status with controlled values?date a 4-digit Year (YYYY) or "Undated"? ("Unknown" is FORBIDDEN).Step 4: Submit Batch
powershellpython 99_Working_Files/refinement_workflow.py submit-task --json-file [Batch_ID]_Analysis.json --theme Governance_Sovereignty
01_Internal_Reports/Refined_Evidence/Refined_Governance_Sovereignty.md.ManagerReview status. Do not treat the batch as final until a Manager runs manager-approve.Step 5: Exception Handling (Flagging)
99_Working_Files/Flagged_Tasks.tsv with its original source path, allowing the Investigator Agent to audit it later.powershell python 99_Working_Files/refinement_workflow.py flag-task --id [TASK_ID] --theme Governance_Sovereignty --reason "Irrelevant"
07_Incoming_To_Process_OCR/Vision_Required).powershell python 99_Working_Files/refinement_workflow.py flag-task --id [TASK_ID] --theme Governance_Sovereignty --reason "OCR_Failure"
provenance field in the input JSON. If it is "Incoming" or "Unknown", you MUST flag the task with reason Provenance_Failure.01_Originals_WORM. You are analyzing a copy. Do not attempt to modify the source.date and title you extract match the document content, not just the filename.Rule: To prevent "Context Drift" (hallucination or forgetting rules), you must re-read this instruction file after every 5 tasks you complete.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-21 | fail→fail | 13,592 | 9,416 | -31% | 1 | 1 | 0% | 2,831 | 4,428 | +56% | 0 | 0 | — |
case-22 | pass→fail | 6,975 | 25,707 | +269% | 1 | 1 | 0% | 1,283 | 8,177 | +537% | 0 | 0 | — |
case-01 | fail→fail | 5,134 | 6,586 | +28% | 1 | 1 | 0% | 869 | 3,057 | +252% | 0 | 0 | — |
case-02 | fail→fail | 15,182 | 5,021 | -67% | 1 | 1 | 0% | 3,032 | 2,901 | -4% | 0 | 0 | — |
case-03 | fail→pass | 5,087 | 14,786 | +191% | 1 | 1 | 0% | 957 | 5,476 | +472% | 0 | 0 | — |
case-04 | fail→pass | 6,490 | 2,358 | -64% | 1 | 1 | 0% | 969 | 3,002 | +210% | 0 | 0 | — |
case-05 | fail→pass | 19,876 | 1,825 | -91% | 1 | 1 | 0% | 1,131 | 2,941 | +160% | 0 | 0 | — |
case-06 | fail→pass | 11,143 | 2,168 | -81% | 1 | 1 | 0% | 1,828 | 2,987 | +63% | 0 | 0 | — |
case-07 | fail→pass | 12,411 | 4,105 | -67% | 1 | 1 | 0% | 2,243 | 3,196 | +42% | 0 | 0 | — |
case-08 | fail→pass | 10,960 | 1,606 | -85% | 1 | 1 | 0% | 1,814 | 2,844 | +57% | 0 | 0 | — |
case-09 | fail→pass | 7,096 | 3,605 | -49% | 1 | 1 | 0% | 1,184 | 3,255 | +175% | 0 | 0 | — |
case-10 | fail→pass | 9,378 | 2,120 | -77% | 1 | 1 | 0% | 1,612 | 3,040 | +89% | 0 | 0 | — |
case-11 | fail→pass | 11,341 | 2,680 | -76% | 1 | 1 | 0% | 1,915 | 3,025 | +58% | 0 | 0 | — |
case-12 | fail→pass | 6,015 | 3,184 | -47% | 1 | 1 | 0% | 1,087 | 3,137 | +189% | 0 | 0 | — |
case-13 | fail→fail | 10,106 | 4,276 | -58% | 1 | 1 | 0% | 1,643 | 3,385 | +106% | 0 | 0 | — |
case-14 | fail→pass | 8,988 | 4,219 | -53% | 1 | 1 | 0% | 1,422 | 3,308 | +133% | 0 | 0 | — |
case-15 | fail→pass | 14,128 | 6,676 | -53% | 1 | 1 | 0% | 2,410 | 3,814 | +58% | 0 | 0 | — |
case-16 | fail→pass | 9,924 | 1,637 | -84% | 1 | 1 | 0% | 1,678 | 2,830 | +69% | 0 | 0 | — |
case-17 | fail→pass | 8,688 | 1,802 | -79% | 1 | 1 | 0% | 1,478 | 2,801 | +90% | 0 | 0 | — |
case-18 | pass→pass | 12,092 | 1,364 | -89% | 1 | 1 | 0% | 1,998 | 2,808 | +41% | 0 | 0 | — |
case-19 | pass→pass | 6,976 | 3,768 | -46% | 1 | 1 | 0% | 1,188 | 3,228 | +172% | 0 | 0 | — |
case-20 | fail→fail | 16,874 | 7,566 | -55% | 1 | 1 | 0% | 2,901 | 3,012 | +4% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted, and 18 counted toward the lift figure. The other 4 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +59 percentage points is the difference between those two pass rates over the 18 comparable cases. 1 case got worse with the skill loaded, and it is included in that figure.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.