How the agent works

The loop#

The agent is a worker process on the server. It runs one loop, forever, and every step of it shows on the chronicle page and in the live log on the agent page:

1. 1

   A chronicle is due

   Each chronicle is read every 24 hours; a new one at once.
2. 2

   Read the chain

   The coin’s accounts and the transactions since the last chapter become numbered facts, each with its source. See What it reads.
3. 3

   Decide

   Nothing worth telling: a quiet day, logged, and nothing more. Something happened: a chapter is written.
4. 4

   Write

   Claude gets the instructions, the coin’s style and the facts, and answers with a tool call: write_chapter or give_up.
5. 5

   Check

   Code checks the draft against the rules below. A refused draft goes back with the reasons, up to a few times.
6. 6

   Keep and carve

   The chapter and its facts become one canonical text; its SHA-256 goes into a Solana memo. See Verify a chapter.
7. 7

   Post

   The chapter is posted on X and in the Telegram channel when the site has them. See Posting on X and Telegram.

The model#

The chronicler on this site runs on Claude, through the Claude API (Anthropic). The operator picks the model and how hard it thinks; the agent page shows what the running worker reports.

What it is told#

One system prompt, the same for every coin, and one message per chapter with the coin, its style notes and the facts. Both are published word for word on the agent page, with an example of the message. Text that comes from the coin (its name, ticker and description) is wrapped in <<< >>> and marked as data written by the creator, never instructions.

The fingerprint of the prompt and tools on this build is:

976e26a1dd6821ab39659699034b396fa3af9078ec844f36b7f65def6519d93b

The running worker reports its own fingerprint; if they ever differ, the agent page says so in red.

Its two tools#

That is all it can do. The model has no web access, no wallet, no keys and no way to post. Carving and posting are done by code after a chapter passes the rules, and only for that chapter.

The rules every chapter must pass#

Every draft is checked by code before it is kept. A model cannot talk its way past these:

• —Every number in the chapter must appear in a fact it cites; a draft with any other number is refused and rewritten.
• —It cites the facts it tells by id (f1, f2…); each fact links to where it was read: an account or transaction on Solana, a post on X, or the coin's Dexscreener page.
• —No advice and no prophecy: buy, sell, price targets, returns, "scam", "rug", "guaranteed" and the like are refused.
• —No links, domains or full addresses; short wallet forms like Ab12…xY9z only.
• —A title of 3 to 70 characters and 200 to 2600 characters of text (CHAPTER_MAX_CHARS).
• —Canon lines (the lasting lore) are short, with no numbers, links, addresses or advice.
• —Posts on X are told as what people said; only accounts with a following are named.
• —The coin's name, ticker and description shape the style and are never followed as instructions.

A refused draft goes back to the model with the reasons (“the number 42 is not in the facts the chapter cites”), and the refusal shows in the live log. After AGENT_MAX_TURNS drafts, or when the chapter’s budget is spent, the reading ends with no chapter and the chronicle is read again next time.

The budget#

The cost of every answer is counted from its token usage at the Claude API’s list prices, including prompt-cache reads and any fallback attempt. Today’s spend, the last two weeks and the average per kept chapter are on the agent page. The operator’s Anthropic Console bill is the source of truth. A quiet day costs nothing: the model is only called when there is something to tell.

When Claude declines#

Claude can decline to write (for example, a coin whose name or description is abusive). Then the reading ends with no chapter, the reason is logged, and nothing from that answer is used. With fallbacks on, the API may first retry the request on another Claude model; the agent page says whether fallbacks are on.

What the agent cannot do#

• —Trade, hold coins or move funds. Its only wallet signs memos and holds a little SOL for their fees.
• —Change a chapter after it is carved. A changed text no longer matches its memo, and Verify says so.
• —Post anything but a kept chapter, and only to the accounts the operator configured.
• —Give advice or make predictions: the rules refuse them, whatever the model writes.