Source: ai-research/claude-code-docs-changelog-2026-10-02.md (v2.1.287, October 1, 2026; v2.1.288, October 2, 2026). Field notes: raw/9_NEW_Claude_Mods_that_can_truly_change_how_you_work.md and raw/A_NEW_Agentic_OS_Standard_for_Claude_5.5_is_here_Claude_Mods_Jev_5.5_Models.md (Robo Nuggets, see § Robo Nuggets field notes). Added 2026-10-28: raw/9_Insane_Claude_Mods_I_Already_Can_t_live_Without.md (Ben AI, see § Ben AI field notes). Added 2026-10-30: raw/Mods_in_Claude_Code_-_change_how_Claude_Code_works.md (official Claude YouTube channel, see § Official walkthrough). Also 2026-10-30: raw/Claude_Code_Just_Dropped_Mods_-_Here_Are_5_You_Need.md (Sharbel, see § Sharbel field notes). Added 2026-11-11: raw/Copy_These_5_Claude_Mods_Never_Hit_a_Token_Limit_Again.md (Ben AI) and raw/Give_Me_10_Minutes_and_I_ll_Make_Your_Claude_100X_Better.md (creator not named), see § Token-saving mods.

Product: Claude Code · Date: 2026-10-01 · Version: v2.1.287+

Claude Mods are a new capability type added to the plugin system in v2.1.287 (October 1, 2026). A Mod is a plugin that modifies Claude Code’s own internal behavior rather than just adding new tools or UI panels. The first published Mod surface is $.ui.selection(), which lets a Mod read the text a user has selected in fullscreen mode.

Key Takeaways

  • Mods are a plugin subtype, not a separate install mechanism — they ship as plugins but declare a mods capability in their manifest.
  • The $.ui.selection() API (v2.1.288) is the first documented Mods surface: returns the text last selected in fullscreen mode and, when the selection is within one transcript row, that row’s data.
  • “You should know” (cc-plugin-you-should-know@builtin) is Anthropic’s own built-in Mod: a side agent that watches your session and surfaces things you or Claude might miss. Enable with /plugin enable cc-plugin-you-should-know@builtin.
  • Mods require first-party sessions with telemetry on for the built-in Anthropic mod; third-party mods follow the standard plugin trust model.
  • Deeper behavior access distinguishes Mods from ordinary plugins. Where a regular plugin can add tools and panels, a Mod can reach into the Claude Code UX loop itself.

What Changed

v2.1.287 — Claude Mods launch

  • The plugin manifest gains a Mods capability type; plugins declaring it can modify behavior Claude Code previously kept internal.
  • Built-in mod: “You should know” — a side agent that watches your session and flags overlooked items. Enable: /plugin enable cc-plugin-you-should-know@builtin. Requirement: first-party session (Anthropic API login) with telemetry on.
  • Plugin-defined agents spawned by name now run with the agent’s own prompt, tools, disallowedTools and effort instead of the defaults (fix in v2.1.288 — this broke on the initial v2.1.287 launch).

v2.1.288 — First Mods API surface

  • $.ui.selection() added as the first documented Mods API call. Returns:
    • The text last selected by the user in fullscreen mode.
    • When the selection falls within a single transcript row, that row’s data.
  • The claude plugin test fix for mods reporting as “turned off remotely” when settings were stale also landed in v2.1.288.

Relationship to Existing Plugin Types

Plugin capabilityWhat it can do
ToolsAdd new tool calls for Claude to invoke
Panes / UIAdd panels in the Claude Code interface
Bands / Status lineAdd status-bar items
Mods (new v2.1.287)Modify Claude Code’s own internal behavior loop

Mods sit above the prior capability set: a Mod can do everything a regular plugin can plus hook into Claude Code’s internal UX and behavior primitives.

Built-in Mod: “You Should Know”

“You should know” is the reference implementation of the Mods capability, shipped by Anthropic. It runs a background side agent that watches the main session and surfaces notes about items the user or Claude might miss.

  • Enable: /plugin enable cc-plugin-you-should-know@builtin
  • Requirement: Anthropic API login (claude.ai or Console); telemetry must be on.
  • Behavior: generates “You should know” notes inline in the conversation when the side agent detects something worth surfacing (exact trigger logic is not documented in the changelog; the feature is listed as first-party).
  • The fix in v2.1.287 that resolved plugin-defined agents running with wrong defaults was also a prerequisite for “You should know” running correctly.

Authoring Mods

Full authoring docs are not yet published (as of the v2.1.288 snapshot). The changelog confirms:

  • Mods are declared in a plugin manifest; no separate package format is introduced.
  • The $.ui.selection() call is the first stable API surface, implying more surfaces will follow.
  • Use claude plugin test to validate a Mod locally; the --remote flag now correctly distinguishes a locally-turned-off Mod from a remotely-turned-off one.
  • The plugin-authoring skill (bundled) covers the manifest format; the Mods capability field is expected to appear in that skill’s docs.

Official walkthrough (Anthropic, 2026-10): how a mod is wired

Creator: official Claude YouTube channel (@claude) · URL: https://www.youtube.com/watch?v=yUuFvL1lK4k · Platform: YouTube · auto-caption transcript fetched 2026-10-30, about 750 words. The first first-party explanation of the mod format in the wiki; it answers most of the authoring questions the changelog left open.

Topic: how a plugin declares a mod

Existing claim: (from claude-mods.md, Key Takeaways) — “they ship as plugins but declare a mods capability in their manifest.” The cited changelog (ai-research/claude-code-docs-changelog-2026-10-02.md) says only that Mods are “a new plugin capability type”; it names no manifest field. New source says: (from raw/Mods_in_Claude_Code_-_change_how_Claude_Code_works.md) — “a mod is just a plugin, so the same plugin.json, same hook.json, but the only difference is that hook.json gets one extra key, modules. And this points at a TypeScript file, and that file is the actual mod.” Status: unresolved. The official video is the stronger source; the key name modules and the file name hook.json come from auto-captions and should be checked against the plugin docs or the plugin-authoring skill before the callout is resolved.

  • What a mod can change: what Claude Code renders, what goes into a prompt, which tool calls it makes, and it can add your own UI, skills and commands.
  • Format: a normal plugin. The plugin’s hooks file gets one extra key (captions: modules) that points at a TypeScript file; that file is the mod.
  • Events and handlers. Claude Code has “a few dozen internal events” (a tool call, a permission check, running a message, emitting a prompt). You register a function on one; it runs every time that event fires, right before Claude Code does its own thing. The presenter’s comparison: Express middleware built into Claude Code.
  • Three arguments, always the same:
    • $ (captions also render it as “$send”) — your connection back to Claude Code: rendering UI, calling the model, reading files.
    • the event — the data for what is happening right now. You can change it in the function body.
    • next — whatever you pass to next is what Claude Code actually does, so a changed command is the command that runs. Skip next and return your own value, and Claude Code uses your answer instead of doing its own thing.
  • Loading. Start Claude Code with the --plugin-dir flag (captions: “plugin dir flag”) pointed at the folder. The desktop app takes no flags, so add the folder to your global settings file instead (the setting name is not given). The folder is watched: saving the file updates the current session immediately.
  • Sandboxing. A mod only sees the events Claude Code exposes and runs in its own sandbox, so everything goes through $.
  • Sharing. Because a mod is a plugin, publish it in a marketplace repo and others install it like any plugin.
  • Example mods the presenter built: a live breakdown of context usage rendered above the prompt; a side pane listing every agent in the project (not just the running one), each startable with one click; and an email-hiding mod for screen recordings.

Worked example: hide email addresses until hover.

  1. Register on the render event, which fires when a reply is about to go on screen. The event data is the reply text.
  2. For each email address, build a small block through $ that shows its text only on hover.
  3. Return that instead of calling next, so Claude Code renders your version.
  4. Result: asking for Sam’s email shows a hidden address that appears on hover, and asking Claude to write a message still uses the real address, because a render mod changes only what appears on screen; the model never sees your version.

This is the display-only class of mod that both creator tours recommend starting with (Claude Skins, Screen Share Mode below).

W43 additions (v2.1.304, October 19, 2026)

Source: ai-research/claude-code-docs-changelog-2026-10-23.md.

decider: "person" in tool.check

A mod’s tool.check handler can now return decider: "person" to force the tool call decision through a human gate:

  • If a permission dialog is available (interactive terminal session), it is shown.
  • If no human can be asked (unattended session, no host callback), the call is refused with the mod’s reason.

This differs from the default approve/deny behavior: it guarantees the decision always involves a human, regardless of the session’s permission mode.

e.isReadOnly in tool.check

The tool.check event now includes isReadOnly: true for tool calls that are read-only (e.g., Read, Glob, Grep). Mods can branch on this field to apply different policies to read operations versus write operations.

settings.read hooks can now rewrite merged settings

settings.read hooks now receive each settings layer (user, managed, org, repo, etc.) and can rewrite the merged result. Previously they were read-only observers. This enables:

  • Setting an organizational ceiling on any setting (e.g., cleanupPeriodDays for the retention sweep)
  • Enforcing that session-level or repo-level settings cannot override organizational policy
  • Managed Mods that apply company-wide configuration without user override

Transcript line on skipped mod hooks

When a plugin that runs first skips one of your mod’s hooks, a transcript line is now emitted naming the event and the plugin that caused the skip — previously this was silent.

Mod docked pane width cap

A mod’s docked pane now stays within 45% of the terminal’s width (v2.1.305), including when the session was opened in a wider terminal.

W46 changes (v2.1.321, November 9, 2026)

Source: raw/anthropic-watch-claude-code-tag-v2-1-321.md. Full digest: What’s New — Week 46.

  • A mod’s own tool calls no longer prompt. $.tool.call and $.mcp.call “raise no permission prompt, even under your own ask rules or in dontAsk and -p runs; deny rules and organization policy still apply”. An ask rule is no longer a gate on what a mod does; use a deny rule.
  • Mods Claude writes turn on without asking in auto and bypass modes, “even with managed allowUnsandboxedCommands: false or allowManagedDomainsOnly”, and do not load in dontAsk mode.
  • Headless sessions reload mods in .claude/skills when their files change; set CLAUDE_CODE_PLUGIN_DIR_WATCH=0 to turn this off.

v2.1.323 (November 12, 2026)

Source: ai-research/claude-code-release-v2-1-323.md (full release notes; the raw/ stub was cut at 20,000 characters).

  • engine.create hooks by wildcard only observe. “a *, namespace or negated hook on engine.create now only observes. To add or withhold a $ noun, hook engine.create by name”. claude plugin validate now fails a mod whose engine.create hook “isn’t written in the one accepted form” (installed mods keep loading) and lists the $ nouns such a hook adds, withholds or replaces.
  • Mods can no longer reword Claude Code’s own confirmation questions (listed as a fix).
  • A permission prompt from a mod’s tool.check hook names the plugin that asked, not every plugin hooked on the tool.
  • isPromptCovered is added to the props a mod’s pane is drawn with, “telling it when a dialog or a field above the prompt covers the prompt”.
  • TypeScript checks: claude plugin validate type-checks a plugin’s TypeScript hooks module; each finding is a warning and --strict fails on it.
  • “Not now” on the prompt to turn on mods Claude writes in a session now says when they load.

Robo Nuggets field notes (2026-10): nine example mods and how to build one

Creator: Robo Nuggets (Jay) · URL: https://www.youtube.com/watch?v=lDrAZ1wAyVs · Duration: 12:47 · Platform: YouTube · Published: 2026-10-08. Plus the mods section of his 2026-10-21 agentic-OS video (https://www.youtube.com/watch?v=s_1GPN-NXA0). A hands-on creator tour; which mods are Anthropic’s own is his statement and was not checked against a mod registry.

How a mod works, in his words. Claude Code emits an event each time it does something (runs a command, spawns a subagent). A mod is code that listens for events and can watch one, change it, or handle it itself. This matches the tool.check handler and decider field documented in the W43 section above.

Where it runs. The Claude Code terminal and the desktop app’s Code tab. At recording time it did not appear in the VS Code extension. Needs v2.1.287 or newer (claude update) and a paid plan.

Building one. Describe the feature you want; Claude uses the built-in plugin-authoring skill and offers hot reloading so the mod appears in the current session as soon as it is built. Mods Claude builds are temporary by default: when you like one, ask Claude to save it as a plugin and install it so it loads in every session. Mods cost nothing extra unless they call a model in the background, which uses your normal Claude usage.

ModWhat it doesOrigin (per Jay)Token cost
Savvy ProgressProgress bar above the prompt; for subagent work, a live panel of every subagent with model, progress and rough API cost; customizable pixel crab mascotsCommunityNone (code)
Claude SkinsReskins tool-call display: icons per tool, colored edits; seven themes such as Noir and Tokyo Night; display-only, so a safe first modOpen-source repoNone
You Should KnowSide agent reads Claude’s output and pops up anything important you might miss (his example: the cost of a Kling video run through his provider); steerable toward costs or customer issuesAnthropic (matches the built-in above)Uses tokens (runs a second agent)
File TreeRight-hand file panel; a file shimmers while Claude touches it and turns green when the change is committedCommunityNone
Cache TaxStops you before a message that would re-read a cold-cache conversation and shows the token and dollar estimate; /keep-warm pings the session just before the cache expiresCommunitySmall (keep-warm pings)
Blast RadiusBefore a risky command such as deleting a folder, lists the files that would go and requires a human to press Proceed or CancelAnthropicNone
ReflectWhen feedback contains trigger phrases (“no”, “this doesn’t look good”), a band offers to save it as a permanent rule in CLAUDE.md or the skill fileCommunitySmall
Terminal BrowserOpens a real browser beside the terminal chat to review pages or GitHub pull requestsCommunityNone
Replay TheaterAfter a turn, a replay button steps through every file edit Claude made; also works for prose draftsAnthropic sampleNone

His own usage-bar mod. In the agentic-OS video Jay shows a mod that keeps his 5-hour and 7-day usage limits and the prompt-cache state (warm or cold, idle time, tokens that would re-cache) visible at all times. He calls mods “Chrome extensions for Claude Code.” See ARMS Agentic OS for how he uses it to decide when to compact or keep a session warm. In his October desktop-app tour he names it Ribbon, says it is free and open source, and shows it as a slim bar above the chat box with remaining weekly usage and warm/cold cache state; he also uses Savvy Progress daily (raw/Use_Claude_2.0_better_than_99%_of_people_11_BIG_Updates.md; see The Revamped Claude Desktop App). The repo link is in his video description, not the transcript.

Try it: start with a display-only mod (a skin or file tree), then a human-gate mod for destructive commands. Blast Radius is the same idea as the decider: "person" return value above, and as hook-based command guards such as DCG (see Plugins and Marketplaces).

Ben AI field notes (2026-10): nine mods, four built from scratch

Creator: Ben AI · URL: https://www.youtube.com/watch?v=bTqizdva04g · Duration: 14:06 · Platform: YouTube · Uploaded: 2026-10-21. Ben says he built or tried “more than 100” mods before picking these nine and gives them away as a zip in his video description. The mods themselves were not inspected here. The video also pitches his paid AI Accelerator, which is left out.

His explanation matches Jay’s. Claude Code sends a signal when you send a prompt, when Claude is about to use a tool, when a subagent starts and when Claude finishes a reply. A mod listens for one signal and can watch, change or block what happens. Claude builds a mod from a plain description using the built-in plugin-authoring skill, and you revise it by telling Claude what to change, as with a skill. Mods that call a model use your plan’s usage; display-only mods cost nothing. Both points agree with the Robo Nuggets notes above.

Install route. Drop the zip into a Claude chat and say “Install these mods for me.”

ModWhat it doesTriggerOrigin (per Ben)Uses a model?
Brain DumpHolds back a long, messy voice prompt before Claude sees it, asks a few clarifying questions through a second model, then rewrites it as task, why, guardrails and “what done looks like”Prompt sentHis ownYes
Progress panel (caption: “helper desk”)Purple bar above the prompt box that fills as to-dos are ticked; side panel with one coloured crab per subagent showing its progress and costSubagent start, to-do tickedRebuilt from “Johnny Savy’s progress mod” (the Savvy Progress mod above)No
Cache ClockCountdown to when the chat’s prompt cache goes cold, read from the timer in Anthropic’s reply. Keep warm (button or /keep-warm) sends a small hidden message just before expiry; stops when you send your next message or after 2 hoursReply finishedInspired by “Karan Binsal’s” Cache Tax modA few cents per ping
HandoffReads the whole chat plus the files, skills, tools and links it used, writes a detailed handoff prompt and offers a button to open a fresh chat with it. The prompt can also go to a teammateAt 250,000 tokens, when Cache Clock sees the cache go cold, or /handoffHis ownYes
ReflectSpots a hard rule in any prompt (“never use em dashes”, “keep LinkedIn hooks under 12 words”) and pops up a choice: save to built-in memory, CLAUDE.md or your second brainPrompt sentInspired by a Claude “reflect” skill (credit garbled in captions)Yes ^[inferred]
Chat InventoryA Cowork-style “In this chat” side panel: to-do list with the goal on top, each connector with its logo and call count, skills used, files read and createdEvery tool, connector and skill callHis ownNo
Screen Share ModeTurns every name, email address, phone number and dollar amount in the chat into a grey bar/screenshareNot statedNot stated
Replay TheaterPlays back every edit Claude made to a file, one after another, and lets you return to an earlier version. Works on files only, not on text in the chat/replay after a replyAnthropicNo
SwitchboardLists every running mod; turn each on or off and hide the panels mods add, mid-chat. New mods appear in the list by themselves/modsHis ownNo

Where Ben and Jay agree and differ.

  • Same mods, different builds: Ben’s progress panel, Cache Clock and Reflect are rebuilds of the community mods Jay lists as Savvy Progress, Cache Tax and Reflect. Both name Replay Theater as Anthropic’s.
  • Reflect triggers differently. Jay’s version reacts to negative feedback phrases (“no”, “this doesn’t look good”); Ben’s watches for hard rules in any prompt and adds the second brain as a save target.
  • Ben does not mention the paid-plan requirement or the v2.1.287 minimum that Jay gives. He calls mods “just launched” on 2026-10-21; the changelog dates the launch to 2026-10-01.
  • Ben’s reason to use mods at all: they make the Claude Code tab show what Cowork’s side panel shows, which he says is what moved him from Cowork to Code.

His cache numbers, checked against the wiki.

  • Cache TTL of 1 hour on a subscription and 5 minutes with an API key matches the wiki’s record of Claude Code defaults (see Token Optimization and ENABLE_PROMPT_CACHING_1H in the CLI reference). This also answers the open question in Claude 5.5 Mistakes and Fixes about Jay’s “about an hour”, though Ben is a creator source too.
  • “Cache reads at a 20th of the normal price on Opus 5.5” matches 4 input per MTok (Opus 5.5).
  • “About 4 cents warm versus 8 per MTok). ^[inferred] Ben does not give the chat length.
  • His 250,000-token Handoff threshold is his own rule of thumb (“sort of the cutoff point” for performance), not a documented limit.

Try it: install Chat Inventory or the progress panel first (no model calls), then Cache Clock if you return to long chats after breaks. Ask Claude to build a Switchboard-style /mods list before you have more than a handful running.

Sharbel field notes (2026-10): five mods and an audit prompt for your own

Creator: Sharbel (YouTube @sharbelxyz) · URL: https://www.youtube.com/watch?v=W5JSvXig3Ig · Duration: 11:33 · Uploaded: 2026-10-06 · Source: raw/Claude_Code_Just_Dropped_Mods_-_Here_Are_5_You_Need.md. No sponsor named. Origins are his statements and were not checked.

His description agrees with the two above: a mod runs inside Claude Code, can watch, change or stop what Claude is about to do and draw on screen; install like any plugin (/plugin), in the desktop app or the terminal; “people have built hundreds of mods. Most of them are junk.”

ModWhat it doesOrigin (per Sharbel)
Next StepsAfter each turn, a panel offers three suggested next prompts; click one to paste itAn Anthropic engineer
Cache ControlOn-screen timer until the cache goes cold (about 60 minutes). He had Claude add a button that clears the session and sends a handoff brief to the fresh one, and a readout of cache useNot stated
Plan ProgressChecklist and percentage progress bar showing the current phase; a sound when a task finishesFrom a marketplace
Blast RadiusBefore a delete or force push, shows what would disappear and waits for Proceed or CancelAnthropic
Image ViewerShows pasted or generated images inline in the terminal; mostly useful in the terminal rather than the desktop appNot stated
Honorable mentionA multiplayer Doom game while you wait for ClaudeSame creator as Image Viewer

Build your own from your habits. His prompt: “Audit how I actually use Claude Code, read my last 20 sessions, and find three things, the questions I keep asking you for, the commands, and how I check your work. Then, suggest five mods that would fix those or enhance my workflow. Show me the ideas first. Do not build anything until I pick one.” He picked a “prompt deck” of his most-used prompts, then told Claude to save and install it so it loads in every session. This matches the temporary-by-default behaviour in the Robo Nuggets notes.

His cache figure (warm for about 60 minutes) matches the one-hour subscription TTL recorded in the Ben AI notes above.

Token-saving mods (2026-11): Ben AI’s second set, and a second creator

Source: raw/Copy_These_5_Claude_Mods_Never_Hit_a_Token_Limit_Again.md · Creator: Ben AI (“I’m Ben, I’m a three-time founder and I run an AI agency”) · URL: https://www.youtube.com/watch?v=4d6OK7IGdBU · Platform: YouTube · duration and upload date not in the transcript, fetched 2026-11-11. The title says five mods; the video shows six plus a /savings command. The mods ship in a free zip behind his description link and were not inspected here. His AI Accelerator pitch is left out.

Which “mods” this means. The real Claude Code feature, not a loose word for tips: his definition is “a plugin that changes how Claude Code looks and behaves both in the desktop app and in the terminal”, which listens for a signal (prompt sent, file read, reply finished) and can change or block what happens, or hand work to a second, cheaper model in the background. This matches his October description above and the official walkthrough. Install route is the same: drop the zip into Claude and say “Install these mods for me.” He says they work on Mac and Windows.

ModTriggerWhat it doesHis test (one run each, his numbers)
Input RouterA file read or connector result over about 2,000 tokensSends it to Haiku with your request; Haiku returns short notes and the file path, so the full text never lands in the main chat. For connectors, a plain script first strips fields such as profile pictures and empty values. Side panel shows tokens saved per fileFive Q&A transcripts: 74,000 tokens into chat without the mod, 5,900 with it; 0.94. A Circle connector result: 65,000 tokens down to 981
Effort OptimizerPrompt sentSonnet reads your prompt and Claude’s last reply and suggests a different effort level, which you approve; can fire mid-chat. The model picker does not show the change, but the effort does changeNone given
Sub-agent Model SelectorA subagent is spawnedJudges each subagent’s job: easy to Haiku, normal to Sonnet, complex or risky stays on Opus. Side panel lists the routingResearch task: 3.82 routed (two to Sonnet, one to Haiku), “almost identical output”
Cache ClockReply finishedCountdown to when the prompt cache goes cold; Keep Warm sends a small hidden message just before expiry, “a few cents each time”None given; same mod as in his October set above
Handoff250,000 tokens, or /handoffReads the chat plus the files, skills and links used, writes a detailed handoff prompt, and opens a fresh chat with itNone given; same mod as in his October set above
Efficient Browser UseEvery browser screenshotShrinks it to half size before Claude sees it, using plain code and no model, so it costs nothing to runWebsite review, 48 screenshots: 58,000 tokens without, 14,000 with; 0.17 on Opus

Plus /savings, which reports what each mod saved (his figure: about 1.2 million tokens in one day), and Switchboard from his October set to turn mods on and off.

His caveats and claims, checked where the wiki can.

  • Haiku can miss detail. For accuracy-critical work, switch Input Router off, or ask Claude to change it to use Sonnet.
  • Subagents default to the main model. He says every subagent “runs on the same model as your main chat”. Anthropic’s 2026-11-09 model-choice playbook agrees for the built-in Explore subagent, and names a built-in alternative to his mod: CLAUDE_CODE_SUBAGENT_MODEL sets the default model for subagents that do not name one (see Subagents). ^[inferred] His “more than 90% of tokens” from subagents in such chats is his own figure.
  • The low-versus-max redesign (1 minute against 28 minutes, similar result) is Thariq’s /config example recorded in Which Effort Level to Use.
  • Cache numbers (1 hour on a subscription, 5 minutes with an API key, cache reads at a 20th of the price on Opus 5.5) repeat his October figures, checked in his October notes above.
  • Screenshot sizes fit Anthropic’s image-token estimate of width × height ÷ 750: 1280 × 720 is about 1,229 tokens and 640 × 360 about 307.
  • Effort Optimizer asks before switching. The effort-levels article describes Ben’s effort mod as offering to apply its suggestion automatically; this video shows only the approve step.

A second creator with the same three ideas. raw/Give_Me_10_Minutes_and_I_ll_Make_Your_Claude_100X_Better.md (https://www.youtube.com/watch?v=izSnNPynvps; creator not named in the transcript, promotes a free community) describes three mods of his own: Effort Picker, a card above the chat that recommends an effort level you accept or ignore; a context meter, after which he types /handoff at about 80 to 85% full; and a cold-chat detector that warns once a chat has sat idle past 60 minutes, so he hands off instead of re-reading the context. The rest of that video restates prompting and effort-level advice already in Prompting Claude Opus 5.5 and Which Effort Level to Use.

Try it: start with Efficient Browser Use if you run long browser tasks (no model calls), then Input Router on read-heavy work you can spot-check. Before installing a subagent router, try CLAUDE_CODE_SUBAGENT_MODEL or a model: field in your own subagent definitions and compare /cost on the same task.

Open Questions

  • What other $.ui.* methods exist or are planned? The changelog names $.ui.selection() as the first; no registry of other Mods API calls is published.
  • Do Mods require a separate manifest declaration beyond capability: mods? The changelog does not document the full manifest schema for the Mods capability. Partly answered 2026-10-30: the official walkthrough says the plugin’s hooks file gets one extra key (captions: modules) pointing at a TypeScript file; see the contradiction callout in § Official walkthrough. The exact key and the desktop app’s settings field are still unconfirmed.
  • What does “deeper behavior” include? The “You should know” side-agent pattern suggests Mods can run background agents within the same session context, but the boundaries of what a third-party Mod may access are not published.
  • Ben AI’s mods zip. The nine mods were not downloaded or read here; the “Screen Share Mode” origin and whether it redacts the stored transcript or only the display are not stated. Its name for the progress mod is garbled in the captions.
  • Ben AI’s token-mods zip (2026-11) was not downloaded or read. His savings figures are single runs on his own tasks; how Input Router chooses what to keep, and whether Sub-agent Model Selector rewrites the spawn call or the agent definition, is not shown.
  • Telemetry requirement for third-party Mods? The cc-plugin-you-should-know@builtin requires telemetry on; whether this applies to all Mods or only to the built-in one is not stated.