Tech • AI • Robotics • Game
Claude Code: the only rules that still matter (Opus 5.5 changes everything)
7/10TL;DR
Recent Claude workflows favor clear task briefs, strict guardrails, lightweight instructions and strong verification over elaborate planning modes, sprawling rule files and constant model switching.
KEY POINTS
Planning mode has lost its central role
Early agent workflows often started with a dedicated planning phase before any work began. That approach is now described as largely unnecessary because current agents can plan, execute, reflect and test within one run. The practical priority is to define the task, set boundaries and state what counts as completion.
A precise brief matters more than prompting rituals
High-performing users are increasingly treating the model like an autonomous worker with a defined job. They specify the objective, the limits and the success condition, then let the agent carry out the work end to end. The goal is less manual steering and more reliable autonomy.
Use a stronger main model instead of constantly switching
The advice argues that frequent switching between Opus, Sonnet and lighter models often adds complexity without clear gains. In the comparison cited, Opus at lower or medium settings was framed as matching or beating Sonnet at higher settings while staying close in cost. The key variable is not only the model name but the thinking budget assigned to the task.
Cost depends heavily on thinking budget
Concerns that Opus is always expensive were challenged with examples showing that low-budget use can average about 0.5 cents per task. Higher budgets can push tasks into the $1 to $3 range, but those settings are not necessary for every job. A smarter model can also reduce total cost if it finishes faster and with fewer retries.
Verification is the most important habit
The strongest recommendation is to make the agent check its own work repeatedly. That requires access to a browser or testing environment so it can inspect outputs, log into services, reproduce user flows and fix mistakes over multiple iterations. Instead of stopping after one pass, top workflows push the agent through 2 to 9 rounds of testing and correction.
Instruction files should be shorter, not longer
Large Claude.md style rule files were described as a growing source of poor results. The claim is that bloated instruction sets confuse newer models, especially when teams keep appending warnings, priorities and formatting demands over time. A target of under 200 lines, and ideally far less, was presented as a better default.
General agent files are replacing tool-specific ones
Rather than maintaining a file tailored to one assistant, the recommendation is to move toward a broader agent.md format that multiple coding agents can understand. The benefit is portability across tools and less maintenance as the ecosystem changes quickly. In practice, that means keeping one compact operating manual instead of many overlapping prompt layers.
Too many skills can degrade performance
Many users install large collections of skills, plugins and prompt packs in the hope of making the agent smarter. The argument here is the opposite: modern models often perform better with no extra skills at all, unless a specific weakness needs to be fixed. Unused skills consume context, add instructions and can make behavior erratic.
Block bad actions with tools, not warnings
When an agent behaves dangerously, adding more written prohibitions is seen as less effective than hard technical controls. One example is intercepting destructive commands such as rm -rf and replacing them with a safer action that moves files to the trash. Deterministic safeguards reduce reliance on the model remembering every rule.
Prefer CLI integrations over built-in connectors
Default connectors to services such as Notion can lock workflows into one platform and may perform less reliably than direct tooling. The preferred route is the official CLI, which agents handle well because it is text-based and structured. If no CLI exists, a small script can bridge the service’s API into a command-line tool the agent can use consistently.
The biggest gains come from automation
The end goal is not better chatting but more work completed without supervision. Background agents running on a VPS or isolated machine can audit data, research topics, answer email and manage routine operations while the user is offline. That approach turns AI from an assistant used occasionally into a persistent operational layer.
CONCLUSION
The emerging playbook for advanced Claude use is simpler than many earlier prompting trends: fewer rituals, fewer rules and fewer add-ons, paired with stronger models, hard guardrails and aggressive self-testing. The main competitive edge comes from building workflows that let the agent execute and verify real work autonomously.
Ask a question