See prompts, agents and tools

Claude Code starts every conversation with a long system prompt, hands work to named subagents, and calls tools all day. mike records all three, and this page covers the screens that show them: Prompts, Agents, Tools, Agent dispatches and the tool-error list. Use them to see which prompt a session ran with, whether a prompt changed between versions of Claude Code, which subagent was given a task, and which tool calls failed. They only read what was captured; nothing here changes what is sent to Anthropic.

This is a fictional example. The prompts and agents shown come from the guide’s docs session, not from real traffic.

Find out whether a prompt changed

Continue the example. You updated Claude Code and the docs session started behaving differently. You want to know whether its system prompt changed.

1. Open Prompts. Every distinct system prompt mike has seen is a row, named for the agent it belongs to: Claude Code (main agent) or a subagent such as frontend-developer. Each row shows how often it was seen and when last.
2. Rows are ordered by what needs attention first: Drifted, Needs investigation, Unregistered, Registered. Use the chips above the list to show one group.
3. Select a drifted row (its action reads Review drift). The Drift tab shows the prompt you registered next to the one most recently seen, with altered blocks marked altered.
4. Read the difference. If the new wording is fine, register the new prompt; if not, you now know which Claude Code update changed what.

You should see two versions of the same agent’s prompt, with the changed block highlighted.

What registering means

A prompt is unregistered until you give it a label and say which agent it is. Registering is bookkeeping: it lets mike tell you when that prompt changes. An unregistered prompt still passes through untouched, and a registered one is not blocked or edited either.

To register one, select Register this on an unregistered row, name the composition, and confirm. You can add the agent’s file, such as .claude/agents/<name>.md, and a persona to link it to an agent. The composition page lists the prompt’s Blocks; a prompt whose third block cannot be found shows a warning that traffic will not match until it is present.

The catalogue toolbar can import compositions from a JSON backup and export all registered compositions as JSON. Refresh reloads the list without waiting for the automatic refresh.

Volatile lines

Some prompt lines change on every run, such as a working-directory line, and would make every session look like a new prompt. Volatile lines (/prompts/volatile-lines) lists patterns that mike strips out before it decides two prompts are the same. To add one, give it a name, a regular expression matched line by line, and the block it applies to. The bytes sent to Anthropic are not changed by this; only mike’s comparison is.

Browse the agents

Agents lists every agent identity mike has seen: subagents Claude dispatched, the prompt versions captured for each, and its instances.

1. Open Agents. Search by name, alias or definition in Find a name, alias or definition.
2. Narrow by where the agent is defined with the source chips:
   • observed: seen in traffic; no definition on disk was confirmed.
   • file-backed: you confirmed it has a matching .md file.
   • dynamic: generated on the fly by Claude.
   • builtin: ships with Claude Code itself.
3. Select Inspect agent. The page shows its aliases (Also known as), its description and, below, its dispatches and the prompt versions captured for it.
4. Expand a dispatch to read the brief that was sent, or a prompt version to see its blocks.

On an agent’s page you can edit its description and which file defines it, and change its source. Deleting an agent keeps the captured evidence; only the link between them is cut.

With no agents yet, the page says No agents registered yet. Agents appear as named subagents show up in captured traffic, or when you register one against a composition. Open prompts takes you to the catalogue.

See what each subagent was asked

Agent dispatches lists every time a session handed work to a subagent.

1. Open Agent dispatches. Leave Project and Since blank to see everything, or type an exact project path and a date such as 2026-11-01T00:00:00Z and press Apply.
2. Each row shows when it happened and a short description. Expand one to read the full description. View brief opens the parent request, where the full brief is, because the page itself shows only the description.
3. Clear removes the filters.

The list is empty until a session dispatches a subagent: Sub-agent spawns appear here once a session dispatches a Task tool call.

Search tool calls and find failures

Tools is a feed of every tool call across all sessions.

1. Open Tools. Choose a Project, a Tool name and a Time range (any time, last hour, 24 hours, 7 days or 30 days), then Apply.
2. Switch on Errors only to see failed calls. The address then ends in #errors, and /diagnostics/tool-errors opens the same view.
3. Expand a call to read its full input and jump to the request it came from.
4. Clear resets the filters.

Calls are grouped by project, and calls from sessions with no project fall under Ungrouped. With no calls, the page says Tool calls appear here once sessions make their first tool use.

When it is empty or something fails

Where this goes next

To read one session’s calls in order, open it from Sessions. To change what a prompt contains before it is sent, see Filters. What mike keeps and sends is on Privacy, and the settings that gate each feature are in Settings.