Claude Certified Architect – Foundations

Not this time. One response can hold text and tool calls together, so text is no sign the work is over. stop_reason says it: tool_use means Claude wants a tool run and the result sent back, end_turn means it has finished. A cap is a safety net, not the way a loop should end. Searching the reply for words is the parsing of natural language the guide names as an anti-pattern, and an extra call to ask costs more and can be wrong too.

Not this time. Forking gives you a second session that starts from the same history, and the original stays as it was, so both attempts share the hour of analysis and nothing else. Asking one session to set an attempt aside doesn't take it out of its context. Mapping the module twice throws the work away. Starting fresh from a summary is the right call when earlier tool results have gone stale; here they still hold, and a summary loses detail. One thing to know: a fork copies the conversation, not the files, so edits made in either fork land in the same folder.

Not this time. Deciding which subagents a question needs is the coordinator's job. A one-fact question may need one search and a short answer, not the whole pipeline. Running all four at once still runs all four, and synthesis can't start before there is something to combine. A larger model plans the same fixed route. Stored reports only answer what was asked before, and they go out of date.

Not this time. An empty list is a real answer: the search worked and found nothing. A timeout is a search that failed, and the agent has to know that to retry, wait or tell the customer the truth. MCP reports a tool failure in the result with isError set, and a category and a retry flag let the agent pick the next step. A line in the description can't say which call failed. A bare "Operation failed" hides whether a retry could help. Retrying without end leaves the customer waiting: a short retry inside the tool is fine, but what it can't fix has to come back as an error.

Not this time. .mcp.json in the project is the shared, version-controlled place for servers the whole team uses, and Claude Code expands ${VAR} references in it, so the file can be committed while each developer keeps the token in their own environment. For a common tool like Jira, a ready-made community server is a better choice than building your own. A token in code is a secret in the repository. ~/.claude.json is for personal servers, and copies set up by hand drift apart. An ignored .mcp.json is no longer shared.

Not this time. The Explore subagent is made for noisy searching: it's read-only, does its searching apart from your conversation and returns a summary, so the main conversation keeps room for the plan and the edits. /clear throws away the findings with the noise. Reading every file first fills the context even faster than the search did. Plan mode holds back edits until you approve a plan; it doesn't keep search output out of the context.

Not this time. The review doesn't know what it said last time unless you tell it. Given the earlier findings and asked for only issues that are new or still not fixed, it stops repeating itself and still catches problems pushed later. Matching text word for word fails as soon as a run words the same issue differently. Reviewing only the first run misses everything after it. --max-turns limits the turns within one run; it knows nothing about earlier runs.

Not this time. Retrying with the error works when the model got the format or the structure wrong. It can't work when the fact isn't in what the model was given: no number of tries finds a rate that sits in another document. Send that document, or let the field be empty and pass the case to a person. More retries pay for the same failure again. The error in the system prompt changes nothing, because the rate still isn't there. A stricter schema pushes the model to make up a number, which is worse than an empty field.

Not this time. The session that wrote the code still holds the reasoning behind it, so it tends to agree with its own choices. A separate instance with only the diff and the review criteria has no such pull; Anthropic's Claude Code guidance says the same, that a fresh context improves review because Claude isn't biased toward code it just wrote. An instruction to be critical and extended thinking both keep the same context. A second review in the same session repeats the same blind spots, and keeping only issues found twice drops real ones.

Not this time. Two credible figures that differ are part of the finding. Keeping each with its source and date lets the reader see what each one measured, and the dates show this may be a change over a year rather than a contradiction. An average is a number neither source reported. Quietly keeping only the newer figure hides a source the reader may need, and dropping the figure leaves a gap the report never mentions.