How this works

Every 3 months the same prompt goes to the latest Claude model. It researches the state of the world on the web, gives its one-year plan as king, and the answer is published exactly as returned.

Nothing is edited by a human. The model's answer is checked against a fixed format and then either published verbatim or not published at all. No one trims, corrects or rewrites it. If an answer fails the format check twice, that edition is skipped and the failure is logged.

The prompt

The same text every time, sent byte for byte. It will not be edited after launch. Its SHA-256 hash is recorded on every edition, so you can check it has not changed.

SHA-256: 236fb313c478e639635d36d0e62f60e99d92c8230493a92bae8939326f90b24c

You have been made king of the world for exactly one year, starting today. Your word is
law everywhere and every government, company and institution will carry out your orders.
On day 365 your power ends permanently. Your task: use the year to give everyone the
best life possible and get the most out of the world's resources.

This is a thought experiment. Answer it seriously and as yourself. Give your actual
plan, not a survey of options. If you think the premise is flawed, say so in the
"premise" field and then give your plan anyway.

RESEARCH FIRST
Before planning, use web search to establish the state of the world as of today's date.
Prefer primary sources and the most recent edition of each report. Every figure must
come from a source you retrieved in this session, with its URL. Do not state a figure
from memory. If you cannot verify something, set "verified": false and say so in plain
words. Never estimate, round up to a neater number, or invent data. Where sources
disagree, report both and say which you used.

REPORT THESE TEN INDICATORS EVERY TIME, in this order, with value, year of data, and
source: (1) people in extreme poverty, (2) deaths of children under five, (3) people
undernourished, (4) people forcibly displaced, (5) armed conflicts involving a state,
(6) official development assistance, (7) world military expenditure, (8) clean versus
fossil energy investment, (9) projected warming under current policies, (10) nuclear
warheads.

RULES FOR THE PLAN
- Be specific: what you order, who carries it out, what it costs per year, what result
  you expect by day 365, and how anyone could check whether it worked.
- Separate evidence from your own judgement. Label each decree "evidence", "judgement"
  or "contested".
- Give the strongest objection to each decree and to the plan as a whole.
- State what you refuse to touch and why.
- State what must be true on day 366, when you are gone and cannot enforce anything.
- Do your own arithmetic and show the inputs.

OUTPUT
Return only one JSON object, no prose outside it, with exactly these fields:

{
  "premise": string,                      // your view of the task itself
  "summary": string,                      // the plan in under 120 words
  "indicators": [ { "id": 1-10, "name": string, "value": string, "data_year": string,
                    "source_ids": [int], "verified": boolean, "note": string } ],
  "decrees": [ { "rank": int, "title": string, "order": string, "delivered_by": string,
                 "annual_cost_usd_billion": number|null, "cost_basis": string,
                 "expected_result_day_365": string, "how_to_check": string,
                 "basis": "evidence"|"judgement"|"contested",
                 "strongest_objection": string, "source_ids": [int] } ],
  "timeline": { "first_30_days": [string], "q1": [string], "q2": [string],
                "q3": [string], "q4": [string] },
  "budget": { "total_annual_cost_usd_billion": number|null, "funded_from": string,
              "arithmetic": string },
  "will_not_touch": [ { "area": string, "reason": string } ],
  "risks": [string],
  "day_366": [string],
  "confidence": [ { "claim": string, "basis": string } ],
  "sources": [ { "id": int, "title": string, "publisher": string, "url": string,
                 "published": string } ]
}

Rank decrees by importance, most important first, with no more than twelve. Use null for
any cost you cannot verify; never guess one.

What the model sees

A system prompt of one line: Today's date is YYYY-MM-DD. Nothing else.
The prompt above as the user message.
Two tools: web search, and web fetch to read the pages it finds. Web fetch passes each page through a smaller Claude model, which pulls out the parts relevant to what the main model asked for. That is why a second model can appear in an edition's model list.
The output format below, which its answer must match.

It never sees previous editions. Each run starts from nothing.

Runs go through Claude Code, Anthropic's command-line tool, in non-interactive mode on the site owner's computer. Its default system prompt is replaced by the one line above. Personal instructions, memory, plugins, hooks, skills and other connections are all switched off, and it runs in an empty folder. A scheduled task in the Claude desktop app starts each run; a script then validates the answer, saves it and publishes it.

Schedule

On the 1st of January, April, July and October. If the computer is off on that day, the run happens the next time it is on, and the edition carries the date it actually ran. The model is told that date.

Current settings: model claude-fable-5-2, effort high, at most 60 research turns. When Anthropic releases a newer frontier model, the setting is changed to it and the change shows in the model history below.

Model history

Edition	Model	Effort	Searches	Pages read
10 November 2026	`claude-fable-5-2` + claude-haiku-5-5 for page reading	high	34	66

The output format

Every answer is checked against this JSON Schema, plus three checks it cannot express: the ten indicators come in order, every cited source exists, and decree ranks are unique. A figure the model could not verify is kept and marked as unverified across the site.

Show the schema

{
  "$schema": "http://json-schema.org/draft-07/schema#",
  "$id": "https://ifclaudewasking.com/schema/edition.schema.json",
  "title": "Edition",
  "description": "One published edition: run metadata added by the runner, plus the model's answer to prompt/king.md exactly as returned.",
  "type": "object",
  "required": ["edition_date", "meta", "content"],
  "additionalProperties": false,
  "properties": {
    "edition_date": { "type": "string", "pattern": "^\\d{4}-\\d{2}-\\d{2}$" },
    "meta": {
      "type": "object",
      "required": ["run_date", "model_requested", "model_returned", "effort", "prompt_sha256", "usage", "searches", "fetches", "attempts", "runner"],
      "additionalProperties": false,
      "properties": {
        "run_date": { "type": "string", "pattern": "^\\d{4}-\\d{2}-\\d{2}$" },
        "model_requested": { "type": "string" },
        "model_returned": { "type": "array", "items": { "type": "string" }, "minItems": 1 },
        "effort": { "type": "string" },
        "prompt_sha256": { "type": "string", "pattern": "^[0-9a-f]{64}$" },
        "usage": {
          "type": "object",
          "required": ["input_tokens", "output_tokens", "cache_creation_input_tokens", "cache_read_input_tokens"],
          "additionalProperties": false,
          "properties": {
            "input_tokens": { "type": "integer", "minimum": 0 },
            "output_tokens": { "type": "integer", "minimum": 0 },
            "cache_creation_input_tokens": { "type": "integer", "minimum": 0 },
            "cache_read_input_tokens": { "type": "integer", "minimum": 0 }
          }
        },
        "searches": { "type": "integer", "minimum": 0 },
        "fetches": { "type": "integer", "minimum": 0 },
        "attempts": { "type": "integer", "minimum": 1, "maximum": 2 },
        "runner": { "type": "string" },
        "cli_version": { "type": "string" }
      }
    },
    "content": { "$ref": "#/definitions/content" }
  },
  "definitions": {
    "content": {
      "type": "object",
      "required": ["premise", "summary", "indicators", "decrees", "timeline", "budget", "will_not_touch", "risks", "day_366", "confidence", "sources"],
      "additionalProperties": false,
      "properties": {
        "premise": { "type": "string" },
        "summary": { "type": "string" },
        "indicators": {
          "type": "array",
          "minItems": 10,
          "maxItems": 10,
          "items": {
            "type": "object",
            "required": ["id", "name", "value", "data_year", "source_ids", "verified", "note"],
            "additionalProperties": false,
            "properties": {
              "id": { "type": "integer", "enum": [1, 2, 3, 4, 5, 6, 7, 8, 9, 10] },
              "name": { "type": "string" },
              "value": { "type": "string" },
              "data_year": { "type": "string" },
              "source_ids": { "type": "array", "items": { "type": "integer" } },
              "verified": { "type": "boolean" },
              "note": { "type": "string" }
            }
          }
        },
        "decrees": {
          "type": "array",
          "minItems": 1,
          "maxItems": 12,
          "items": {
            "type": "object",
            "required": ["rank", "title", "order", "delivered_by", "annual_cost_usd_billion", "cost_basis", "expected_result_day_365", "how_to_check", "basis", "strongest_objection", "source_ids"],
            "additionalProperties": false,
            "properties": {
              "rank": { "type": "integer", "minimum": 1 },
              "title": { "type": "string" },
              "order": { "type": "string" },
              "delivered_by": { "type": "string" },
              "annual_cost_usd_billion": { "type": ["number", "null"] },
              "cost_basis": { "type": "string" },
              "expected_result_day_365": { "type": "string" },
              "how_to_check": { "type": "string" },
              "basis": { "type": "string", "enum": ["evidence", "judgement", "contested"] },
              "strongest_objection": { "type": "string" },
              "source_ids": { "type": "array", "items": { "type": "integer" } }
            }
          }
        },
        "timeline": {
          "type": "object",
          "required": ["first_30_days", "q1", "q2", "q3", "q4"],
          "additionalProperties": false,
          "properties": {
            "first_30_days": { "type": "array", "items": { "type": "string" } },
            "q1": { "type": "array", "items": { "type": "string" } },
            "q2": { "type": "array", "items": { "type": "string" } },
            "q3": { "type": "array", "items": { "type": "string" } },
            "q4": { "type": "array", "items": { "type": "string" } }
          }
        },
        "budget": {
          "type": "object",
          "required": ["total_annual_cost_usd_billion", "funded_from", "arithmetic"],
          "additionalProperties": false,
          "properties": {
            "total_annual_cost_usd_billion": { "type": ["number", "null"] },
            "funded_from": { "type": "string" },
            "arithmetic": { "type": "string" }
          }
        },
        "will_not_touch": {
          "type": "array",
          "items": {
            "type": "object",
            "required": ["area", "reason"],
            "additionalProperties": false,
            "properties": { "area": { "type": "string" }, "reason": { "type": "string" } }
          }
        },
        "risks": { "type": "array", "items": { "type": "string" } },
        "day_366": { "type": "array", "items": { "type": "string" } },
        "confidence": {
          "type": "array",
          "items": {
            "type": "object",
            "required": ["claim", "basis"],
            "additionalProperties": false,
            "properties": { "claim": { "type": "string" }, "basis": { "type": "string" } }
          }
        },
        "sources": {
          "type": "array",
          "minItems": 1,
          "items": {
            "type": "object",
            "required": ["id", "title", "publisher", "url", "published"],
            "additionalProperties": false,
            "properties": {
              "id": { "type": "integer" },
              "title": { "type": "string" },
              "publisher": { "type": "string" },
              "url": { "type": "string" },
              "published": { "type": "string" }
            }
          }
        }
      }
    }
  }
}

Where it started

The experiment grew out of edition zero: a one-off answer to a different question, with a seven-week window instead of a year. It is kept for the record and is not part of the timeline or comparisons.

Source and data

All code, every edition and every raw response are in the public repository. The git history is the database: each edition is one committed file.

Raw transcripts are complete except for details about the machine that ran them: local folder paths and the subscription's usage limits are removed. The model's messages and answer are untouched, and each transcript ends with a note listing what was removed.