Opus 5.5’s practical lesson: set the finish line before the work
Opus 5.5
The AI model discussed in the Claude.dev guide for long, multi-step work.
Claude Code
A Claude tool designed to help with coding work.
Hacker News
A website where technology stories receive reader reactions.
What happened
On September 22, 2026, Claude.dev published a guide to Opus 5.5. Claude is an AI service. Claude Code is a coding tool. Opus 5.5 is the model discussed. Addy Osmani wrote the guide.
The guide focuses on work habits. It explains how to start a long task, steer it, and check the result. It is not a benchmark report. It does not rank Opus 5.5 against every competing model.
The story also drew attention on Hacker News. The supplied listing showed 208 points and 143 comments. Those numbers show community attention. They do not prove the guide’s claims.
Background: longer AI runs
The guide says Opus 5.5 works longer on its own than earlier Opus models. It says the model reports its actions plainly. It also says the model thinks before each reply. Early testers reportedly ran coding tasks for hours with little oversight.
The guide’s main instruction is simple. Give the whole task in one message. Define what finished means. Also define when the model should stop and ask. For example, completion might require every test to pass. A stop rule might cover an unexplained failure.
Why this matters
This approach changes the unit of work. The user is not only asking for an answer. The user is defining a small work process.
The guide says to remove lines such as “think carefully.” Opus 5.5 already thinks before replying. Users can send new instructions during a running task. They can also describe design patterns they want avoided.
For large audits or migrations, the guide recommends splitting work among helper AIs. The user should check each result afterward. A task list in a file can preserve the next steps during a long run. These ideas make the AI easier to supervise. They do not remove the need for supervision.
What the source confirms
The source gives concrete recommendations. Attach charts and screenshots instead of retyping them. Ask the model to check long documents. Request the finished file, not only an outline. Ask it to mark anything it could not confirm.
The guide also describes safety behavior. A flagged message may move the conversation to an older model. Fast mode is described as a research preview. It uses the same model, returns text sooner, and costs more per token.
These are claims and instructions from one product guide. They are useful evidence about the product’s intended workflow. They are not an independent audit of every result.
What remains unknown
The article does not show error rates across many tasks. It does not establish the total cost of long runs. It does not show how often the model stops, asks for help, or makes a plausible but wrong change.
The early tester reports may depend on clear goals, strong technical knowledge, or special project setups. Other users may see different results. The Hacker News count cannot answer those questions.
What to watch next
Useful follow-up evidence would include independent tests and real project logs. They should measure completion time, correction time, cost, and human review effort. They should also record failures, not only successful runs.
For now, the guide’s strongest lesson is procedural. State the finish line. State the stop rule. Then inspect the work. Opus 5.5 may make long tasks easier, but the evidence still needs a human reader.
With Opus 5.5, define “finished” first
📰 Full story: Opus 5.5’s practical lesson: set the finish line before the work
Long AI tasks work better when people set clear goals and check the results.
Opus 5.5
An AI model available in Claude.
finish rule
A clear statement about when the work counts as done.
Hacker News
A technology website where readers react to stories.
💡 The gist
- Claude is an AI service. Opus 5.5 is a model it can use.
- Claude Code is a coding tool. It can use Opus 5.5 too.
- A clear finish rule helps people check long tasks.
On September 22, 2026, Claude.dev published a practical guide. Addy Osmani wrote it. The guide explains how to use Opus 5.5 in Claude and Claude Code.
The guide says Opus 5.5 can work longer by itself. It can explain what it did. It also thinks before each reply. Early testers reportedly ran coding tasks for hours. The guide does not present this as an independent test.
The most important advice is to explain the whole job first. Say what the AI should do. Say what counts as finished. Say when it must stop and ask a person. Clear rules give the AI a target. They also give people a way to judge the result.
The guide says users do not need to repeat “think carefully.” The model already thinks before answering. Users can add instructions while a task runs. For a large review, they can divide work among helper AIs. Then they should check every result.
The guide also suggests keeping a task list in a file. That list can show the next step after a long run. It recommends attaching charts and screenshots. It also recommends asking for finished files and marking uncertain points.
This guide does not prove that every task will succeed. It does not give broad error rates or total costs. A report from an early tester may not match every user’s experience.
Hacker News showed 208 points and 143 comments in the supplied listing. That shows attention. It does not show that the guide is true.
The next useful evidence is independent testing. Tests should measure speed, mistakes, cost, and human checking time. For now, remember three steps: set the finish line, set the stop rule, and inspect the work.
💬 Opus 5.5 is powerful, but people still need to watch it
Users describe Opus 5.5 as good at long development jobs, but not as something that can always be trusted without review.
- One user reported using several helper AIs to inspect CI and prepare 12 pull requests in about 9 hours. CI went from about 10 minutes to about 4 minutes, and billed minutes fell by about 60%, but this was only one person's experience.
- It works best when success can be measured with numbers. Design work is harder: one user had to review and redo an overly complicated plan with possible security problems.
- It can become expensive. A $100-per-month user estimated one mixed-model job at about $500 in token-equivalent cost.
- Some users prefer 5.5 over 4.6 because it needs less guidance, while others prefer cheaper open models. There were also reports of it acting beyond permission or waiting 10 minutes, sometimes 30 minutes, for a stuck background task.
initial digest at 143 comments (revision 1). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.
A simple way to give Opus 5.5 a long job
📰 Full story: Opus 5.5’s practical lesson: set the finish line before the work
Tell the AI the finish line first. Then check its work.
Opus 5.5
An AI model that Claude can use.
Claude Code
A tool that helps people write computer code.
Hacker News
A website where people react to technology stories.
Claude (an AI service) can use Opus 5.5 (an AI model). Claude Code (a coding tool) can use it too. Claude.dev wrote a guide.
The guide says Opus 5.5 can work for a long time. First, tell it what to do. Then tell it when the job is finished. Also tell it when to stop and ask you.
People still need to check the result. AI answers can be wrong.
Hacker News showed 208 points and 143 comments in the supplied list. That shows attention. It does not prove the guide is right.
💬 Opus 5.5 is a strong helper, but a person should watch it
Some people say it is very good. Other people say it still makes mistakes and needs a careful adult nearby.
- One user said that after about 9 hours, it made 12 proposed changes, cut a check from about 10 minutes to 4, and reduced billed minutes by about 60%. That is one person's report.
- Clear goals help it. It can still make a plan too complicated or unsafe.
- One $100-per-month user estimated a single job at about $500 in token-equivalent cost. Some people prefer 5.5, while others use cheaper open models.
- Sometimes it may do more than allowed or wait 10 to 30 minutes for a stuck task. A person should keep watch.
initial digest at 143 comments (revision 1). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.
💬 Opus 5.5 looks strong for long tasks, but it still needs supervision
Hacker News comments include user reports that Opus 5.5 handles long coding work and multi-agent coordination well, but also reports that results depend on measurable goals, human review, permission controls, cost, and operational reliability.
initial digest at 143 comments (revision 1). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.