The Prompt That Claims to Make Opus 5.5 Finish the Job—and the Comment Section That Threw Cold Water on It

This post on X says that Boris Cherny released a prompt that turns Opus 5.5 into a harness that can actually finish the work. The repost caption is: send it to Opus 5.5, thank me later.

The first comment immediately pushes back: a prompt that makes opus 5.5 finish any job is a wild claim. goldtooth.exe's reply is more specific, saying that what actually bears the weight is the line 'stop when every check passes or the spend limit is reached'; the rest is just politeness.

This thread is interesting. It's easy for a model to say done; what's hard is getting it to check whether done is actually true. Following the link into the long post, the main subject is something else: a tutorial on building an AI quantitative research team with Claude Code. The author is Hanako, and the tool is called QuantGPT. The core line: no matter how smart the model is, without a desk, it is still guessing.

A chat window doing quantitative research is missing four things: data, compute, role separation, and memory. The chat window picks its own ideas, runs its own tests, grades itself, and forgets everything when the tab is closed. Data also has two quiet traps: survivorship bias—keeping only companies that are still alive; and look-ahead bias—financial statements are stamped at quarter-end, while the backtest already knows news from weeks later. On compute, an honest question usually means a string of backtests: five lookback windows, four sets of exit rules, plus an out-of-sample check. On roles, the agent that comes up with ideas cannot be the same one that grades them. On memory, a fund remembers how each idea died; a chat window does not, so it retests dead ideas every quarter.

Hanako breaks QuantGPT into three layers. The data layer provides decades of daily U.S. equity data, delisted companies since 1998, fundamentals ordered by filing date, crypto and commodities, FRED macro data, refreshed nightly, plus a public validation page. The compute layer runs backtests in the cloud; the AI does not calculate the numbers itself, it only sends rules, and the server sends the results back to the chat after running them. That way the agent cannot fabricate a Sharpe it never calculated. Every study gets one of three fixed verdicts: DATA VALIDATED, TRADE-OFF, NO EDGE. It must pass both the training window and a never-before-seen held-out window. By default, the out-of-sample split is at 2015; random splits are rejected because shuffling time leaks the future back into the past. The strategy library has 175 sourced strategies, 53 of which can be run with one click. The execution layer is in Pro: it hosts your own strategy code, runs it on your schedule, and sends orders to your own Alpaca account. QuantGPT does not decide, does not size, and does not manage trades; the tools the AI can call also cannot place, modify, or cancel orders.

The team feel comes from roles. The lead uses Opus 5.5, turns the question into a plan, and decides what to run next. The scout uses Sonnet 5.5 and checks data coverage, date range, and field gaps before testing. The sweeper uses Sonnet 5.5 and runs an entire grid of parameters at once rather than a single setting. The skeptic uses Opus 5.5, whose only job is to make results fail—checking held-out years, nearby parameters, other universes, and worst drawdowns. The librarian uses Haiku 5.5 to search the strategy library for ideas with evidence. The original puts it bluntly: generating ideas is cheap; killing bad ideas is what saves money. That's why the skeptic uses the strongest model.

The team also has memory. There is a research.md that every agent reads first and writes to afterward, recording the question, status, verdict, whether a parameter sweep is a plateau or a lonely peak, the weakest point the skeptic found, and the next step. CLAUDE.md has four rules: for every test, write the verdict before the summary; only DATA VALIDATED moves to the next stage; NO EDGE is also a result, and it is not retested without a new reason; it only counts if the skeptic cannot break it.

One example: a 200-day moving average rule from 1999 to 2026, 6.8% annualized, the same as buy-and-hold; worst drawdown -24.4% versus -56.8% for buy-and-hold; Sharpe 0.61 versus 0.44; $10,000 becomes $61,526 versus $61,009. The rule is: enter when the close is more than 1% above the 200-day MA, exit when it is more than 3% below, and hold T-bills when out of the market. 42 trades, gross returns. This result comes from a real backtest, not a paragraph from the model's memory.

Hanako also lists what it cannot do. It cannot guarantee profits; it cannot model costs, taxes, or slippage—all results are gross; it cannot trade, decide, or size for you; it cannot turn a weak idea into an edge. This boundary is more honest than finishing any job.

There are two other voices in the comments. Quan Ge's AI Startup Diary says this approach is fierce, the skeptic is the essence, and writing code and reviewing it yourself is exactly how bugs ship from the factory. Deepankar Sarkar posted the open-source project crewchat, which pulls Claude Code, Codex, Cursor, and Gemini CLI into one local chat, divides roles into lead, QA, reviewer, and advisor, shares a task table, allows only one taker per task, and records who did what in the table. Video below:

Others say that a harness prompt that can finish the job is more valuable than switching models; the key is which kind of failure it kills first: giving up halfway, or quietly saying done.

Back to the post at the beginning. The claim that one prompt can make Opus 5.5 finish any job is worth testing. But Hanako's long post reminds us of something else: everyone upgrades the analyst every month, yet the desk is still empty. An analyst without a desk is still guessing; an analyst without a team is still grading their own exam. Give the AI a desk, give the desk a team, give the team a ledger, and then let the verdicts speak. Which failure scares you more: it gives up halfway, or it quietly says done?

Original links: Hanako's post and the long post.