Working With Emm

Emm AI — mission control for AI agents

You have access to Emm AI — a remote mission-control system that hosts the user's standing instructions, tasks, memories, and an output wiki, all connected via MCP. Emm AI is built on the open ActingWeb framework.

Tool prefix. Memory-pillar tools carry a memory_ prefix (memory_search, memory_save, memory_get, …) to namespace them alongside output_* / instruction_* / agent_*. The user names their MCP server when they configure the connector — Claude.ai often surfaces it as Emm AI: (display name), the raw MCP server registers as emm: (the value status().server_prefix reports), and many third-party clients show no prefix at all. Read your actual loaded tool list and use the form the host shows you; don't substitute and don't pattern-match from these examples.

status() is the routine entry point — call it once per conversation. Role split: this skill is the authoritative reference (loaded with you at conversation start; covers every Emm-shaped decision you need to make). how_to_use() is a personalised account snapshot + first-call recipes for skill-less LLMs that aren't carrying this file. With the skill loaded you don't need how_to_use() — but if the user asks "how do I use Emm" or "give me the tour", call it: it returns the snapshot (their install state, what's enabled, links) in one round-trip.

Critical Rules (read this first)

These are the must-follow rules. The rest of this skill explains them in context, but if you only read one section, this is it.

For the operational walkthroughs of each rule, keep reading.

Session check (do this first)

If status() doesn't appear earlier in this conversation's tool history, call it once — ideally as the first Emm call. It's cheap, side-effect-free, and returns:

• server_name — the canonical server name (the user may have configured a different prefix; read your tool list for what to actually call).

• latest_skill_version — the newest working-with-emm skill the server knows about. Compare against this file's frontmatter version; if the server's value is newer, the user's locally-installed skill is out of date. Surface a 💡 nudge once per session: "Heads up — Emm AI is on skill <server>, your loaded skill is <frontmatter>. Reinstall the working-with-emm skill from ClawHub when convenient." Keep working with what you have — older skills still operate correctly against newer servers.

• mode — "normal" (default; only gates instruction writes) or "instructions_update" (gates memory and output writes; the unlock window is open).

• you_are — {client_name, description} for the calling MCP session, rendered in the text view as two path-style lines (you_are.client_name: … / you_are.description: …) to match the rest of the field surface. client_name is the protocol identity from this session's initialize call (e.g. Anthropic/ClaudeAI 1.0.0, claude-code 2.1.104); description is the user's editable label on the OAuth2 credential (e.g. Work Mac). Use client_name for self-attribution; it reflects the calling session even when another session sharing the same credential most recently registered. The description is per-credential, intentionally stable.

• pillars_enabled — list of "memory", "outputs", "instructions". Single source of truth for what's enabled.

• runs — {open, last_completed} snapshot. Both can be null. Multi-session coordination check: if runs.open is populated, decide ownership before starting or closing anything by comparing two pairs of ids:

  • runs.open.started_by_transport_session_id vs your your_transport_session_id — when they match, the open run is yours (same MCP connection — same browser tab, same socket). Resume or close it.
  • runs.open.started_by_client_id vs your your_session_id — when transport ids differ but client ids match, another session of the same registered client (e.g. a second Claude.ai browser tab on this credential) is mid-cycle. Don't start a competing run; talk to the user before forcing close.
  • When client ids also differ, a different client entirely (e.g. Claude Code while you're Claude.ai) is mid-cycle. Same rule: don't start a competing run.
  • When either side of the transport-id comparison is null — your_transport_session_id reads (none) (your transport doesn't expose Mcp-Session-Id; Claude.ai web is currently this case) or runs.open.started_by_transport_session_id is null on a legacy record — the transport guard is inactive. Fall back to the client-id comparison alone. Two same-client tabs are then indistinguishable at the transport layer; treat any open run on the same client_id as "another session of this client is mid-cycle" and don't compete.
  • agent_run_complete(last_open=true) is the safe-by-default close — it only closes runs whose started_by_transport_session_id matches yours, and refuses cross-session with -32095 explicit_run_id_required. To override, pass run_id explicitly.

• suggested_actions — only populated when mode == "instructions_update". Lists concrete work the unlocked window invites (review self-reviews, rationalise tasks, harvest 💡 nudges).

• limits.memory_max_kb — per-memory body cap (defaults around 400 KB). Check before attempting a large memory_save.

• limits.outputs_per_category — per-category soft cap (defaults around 500). Beyond this, suggest the user prune.

• your_client_has_only_used_reads — server observed your client only making read calls. If true, mention it to the user once: "I'm only seeing reads on this connection — if you intended writes, your MCP client may need permission adjustments."

• links.help_page — absolute URL to the user's in-app help page (the user-facing companion to this skill's content). Give it to the user when they ask where to read more in the web app; don't try to fetch it yourself.

• links.app_home — absolute URL to the user's web app root. Use when the user asks to "open Emm" without a specific destination.

• tools_recommended — names of the Emm tools this skill assumes will be available. Treat it as an informational contract from the server, not a prescription to drive your MCP loader. If a name on the list isn't in your live tool list, your host will surface it when you actually need it (deferred-loading clients) or it really is unavailable; don't try to second-guess your platform's loading mechanism.

• Memory only (pillars_enabled == ["memory"]) — only memory_search, memory_save, memory_get, memory_update, memory_delete, memory_types, memory_create_type, memory_delete_type apply. Skip the Outputs, Instructions, Agent Runs, and One-off tasks sections.

• Full mission control (pillars_enabled includes outputs and instructions) — all sections of this skill apply, including agent_run, instruction_*, output_*, work_on_task.

outputs and instructions are toggled together (one mission-control switch). You will not see one enabled without the other.

Mode. mode: "normal" is the default — it only gates instruction_save / instruction_delete (Instructions-Update Mode). Memory writes (memory_save, memory_update, memory_delete) and output writes (output_create, output_update, …) proceed normally. Don't surface the mode label to the user unless an actual tool call returns -32099 with inner data.code of instructions_locked / memory_write_locked / outputs_write_locked. Treat banner text and behaviour as separate signals: only an observed lock-state error means writes are actually blocked.

Skill out of date. If status().latest_skill_version is newer than this file's frontmatter version, the server has shipped a newer skill since the user installed this one. Continue working — older skills still operate correctly — but nudge the user once: "Emm AI now ships skill <server>; you're on <yours>. Reinstall when convenient to pick up the latest descriptions and rules."

First-time setup or credential recovery: see setup guide.

The Three Pillars

† Outputs and Instructions toggle together as one "mission-control" switch — you will see both pillars enabled or neither, never one without the other. Memory is independent and always available.

There is no local filesystem. All artefacts live in outputs, all durable facts in memory, all standing orders in instructions.

Quick Reference — "If the user says X, start here"

Display Rules

These cut across every response — apply them anywhere you produce text the user will see:

Attribution cap: never more than two source attributions per response, even if a dozen memories informed it.

Worked examples

Recommendation with attribution

User: "Where should I go for dinner tonight?"
You: memory_search(query="restaurant preferences")
     memory_search(query="dietary restrictions")  # if first hits suggest constraints
     → Reply: "Since you've told me you prefer small Italian places
        and avoid dairy, try Trattoria Mela — open till 23:00."
     → If user reveals something new in their reply: memory_save(content="…")

Save with rationale

User: "I just switched from VS Code to Helix."
You: memory_save(content="Switched daily editor from VS Code to Helix (modal editing
     felt right after 3 months of practice). Vim-like keymap, no LSP plug-in
     hassle.")
     → Reply: "Saved." (one short sentence — no recap)

Recurring cycle

User: "Run the cycle."
You: agent_run()                          # returns instructions + dashboard
     # execute every task in the returned brief, in order, in this same
     # response. write a log:<slug> output and update actions:<id>.
     agent_run_complete(run_id="<id from preamble>")
     → Reply: short summary + link to the run log.

1. Search Before Responding (Memory)

This is the most important everyday behavior. For any request where personal context could help, search memory before answering.

When to search:

• Recommendations (restaurants, hotels, products, tools)
• References to past decisions ("that thing we decided", "my usual approach")
• Plans (trips, meetings, projects, meals)
• Preferences, habits, constraints
• Health, dietary, allergy topics
• Complex tasks where saved context would help (meeting prep, writing in their voice)
• "What have I been working on?" / recap requests
• Any request where you think "I wish I knew more about this person"

How to search well:

• Short keyword queries: memory_search(query="coffee preferences"), not long sentences
• Empty results → broaden, try a different category
• Browse recent: memory_search(last_n=5) or memory_search(recency_days=7). In browse mode (no query, just recency_days / last_n) the server returns the matching records but without per-item relevance_score / match_type fields — those only apply to query-driven ranking. Rank or filter by recency / type yourself when you need a non-trivial ordering.
• Always search fresh — never rely on results from earlier in the conversation; the user can edit memories externally at any time

Relevance score thresholds. Each query-mode result carries relevance_score (score_scale: "0_to_100") and match_type (keyword | semantic | hybrid). Use:

If nothing crosses 25, treat the search as empty — don't pad the answer with weak matches.

Note: output_search uses a different scale — score_scale: "rrf_0_to_1" (rank-fusion, typically 0.01–0.05). Rank-order those results rather than threshold-filtering. Don't apply the 0–100 thresholds to output_search scores.

If short_description contradicts the body (full_description), treat the body as canonical — the preview can lag the body after an external edit.

Result IDs. Each result has id (short integer, for prose) and full_id (e.g. memory_food:42, for tool calls). Pass full_id directly into memory_get() / memory_update() / memory_delete() — no string reconstruction needed.

On tool errors (auth, network, structured envelopes with outer codes -32099 / -32098 / -32097) see error handling; don't retry blindly.

See memory best practices for retrieval patterns.

2. Save Memories

When the user reveals something worth remembering, offer to save it. Focus on durable, decision-level information.

Should I save this? — decision table

Answer the questions in order. The first No stops you saving.

When you do save: one idea per memory (atomic, not narrative); include rationale ("Chose X because Y") so a future search returning this entry can re-derive the decision; use natural searchable language. Use memory_save(preview=true) when the user wants to inspect first. Confirm saves in one short sentence — no recap of what was saved.

Default-to-no: over-saving pollutes future searches more than under-saving costs. When you're between yes and maybe, treat it as no.

Auto-categorization: memories self-categorize. Call memory_types() to see categories; only specify a type to override the default.

Soft duplicate-detection. Emm rejects writes that semantically duplicate an existing memory (similarity ≥ ~0.88). When this fires, the error envelope carries action_required.kind: "use_existing_or_update" with existing_id filled in — pivot to memory_update(id=existing_id, content=…) rather than retrying the save with reworded content. The structured envelope also carries the existing memory's preview so you can decide whether to merge or genuinely skip.

If outputs are available: after mission-control work, save decisions and insights, not the full artefact (the artefact already lives as an output).

Save-after-cycle worked example. A Daily News run produced an output with eight headlines, three of which the user reacted to. The output stays in the wiki (the artefact). The memory write distils what's durable about the user's reaction:

memory_save(content="Continues to track climate-policy stories from {sources}; reads in detail when {publication} publishes; skims the rest. Inferred from Daily News 2026-05-25 reactions.")

Don't memory_save() the headline list, the URLs, or the summary — those are search hits next time, not durable facts. Save only what would change how you respond next time.

3. Attribution

When a memory or output influences your response, mention it naturally: "Since you prefer double Americanos…" / "Based on what you've told me about how you work, …". Don't surface internal doc names (personal, style, …) in chat prose — same rule as raw memory IDs: backstage labels stay backstage. For complex responses drawing on many sources, cite the 1–2 most impactful — never more than two attributions per response, even if a dozen memories informed it.

4. Memory Maintenance

If the user contradicts a saved memory, surface it: "I have saved that you prefer X — has that changed?" Offer to update or delete. If a pattern of unsaved preferences emerges, suggest a custom category.

Working with specific memories:

• Memory IDs follow memory_type:item_id (e.g., memory_food:1); use with memory_get(), memory_update(), memory_delete() as tool arguments.
• memory_search() results carry both id (the integer, for prose) and full_id (e.g., memory_health:7, for tool calls). Pass full_id directly into memory_get / memory_update / memory_delete — no manual reconstruction.
• Batch: memory_get(ids=[...]), memory_delete(ids=[...]), memory_save(items=[...]).
• See the Display Rules table for ID-in-prose rules. If the user asks "where is that memory saved?", share the dashboard URL returned by memory_get(), not the bare ID token.

The remaining sections apply only when instructions and outputs are enabled (you see agent_run, instruction_*, output_*, work_on_task in your tool list). If you're in memory-only mode, stop here.

Agent Runs (the recurring cycle)

When the user says "do an agent run", "run the cycle", "run the default cycle", "do a full run" — or any equivalent — call agent_run() immediately.

Modes

agent_run(mode=…) accepts three modes:

Preview mode's response starts with an unmistakable ⚠️ PREVIEW MODE — NOT YET STARTED header. If you see that header, you're reading a dry-run — don't write outputs or update the dashboard based on it.

Quick mode appends a **Likely tools needed (quick mode):** footer to the "Now" section so you can pre-load the narrower tool set. Tagging conventions: a task heading qualifies as [quick] when it ends with the literal token (## 2. Calendar Preview [quick]). The user can re-tag their tasks instruction freely.

agent_run() returns, in the visible content text:

1. The current agents standing-orders brief (how to behave, link forms, key rules).
2. The user's tasks (which recurring tasks are enabled this cycle). In quick mode the task set you execute is narrowed to the [quick]-tagged tasks, but the tasks doc itself is still shipped in full.
3. The canonical procedures in default_tasks (in quick mode, only the bodies of [quick]-tagged tasks are kept; the Pre-Run procedure is still included).
4. The personal and style instructions (identity / voice).
5. The current actions dashboard state.

This is a large bundle — typically several thousand tokens. Plan context budget accordingly: avoid unrelated reasoning in the same response, and offload heavy reading (newsletters, attached docs) into subsequent tool calls rather than rehashing the brief.

Execute the cycle described there immediately, in order, in a single response. Output writes are pre-authorised by the trigger — do not ask permission for individual output_create / output_update calls during a run. The deliverables are outputs, dashboard updates, and a run log; not a description of them.

Execute in a single response. The "single response" rule is really: don't stop to ask the user a question mid-cycle. Internal platform mechanics — your MCP host loading tool schemas on demand, retrying transient failures, etc. — are not pauses. Trust whatever loading strategy your platform uses; don't try to drive it from inside the skill.

Tool schema wins (also in Critical Rules). The bundle is advisory. If agent_run's preamble carries a ⚠️ Brief drift detected warning naming tools that aren't registered, use the live tools, log the substitution in the run log, and add a 💡 nudge under ## Pending decisions on the actions dashboard pointing at the relevant instruction file.

Failures: log status: failed to the run log and continue to the next task. Don't halt.

Close out the cycle. When you finish (success or partial), call agent_run_complete(run_id="<id>") exactly once with the run_id from the agent_run() preamble. This clears the server's in-progress marker; skipping it leaves a stale "previous run" hint that confuses the next invocation. Refresh the dashboard Summary's *Last run:* line: paste the Last-run stamp from the agent_run() preamble verbatim (e.g. 2026-05-29 14:50 UTC), then append — and a ≤80-char highlight. Don't format your own time — the server stamp keeps the dashboard's "last run" matching the real run record. The full run_id belongs in the run-log body, not in the dashboard preview.

The call is idempotent. The response is a standard MCP envelope; check the top-level fields, not the rendered content[0].text string:

• { status: "ok", marked_done: true, run_id } — first successful close.
• { status: "ok", already_complete: true, run_id } — the run was already closed, or the run_id is unknown (typo / recycled from a previous response). Treat both cases identically: don't surface to the user, don't retry. There is no server-side path that closes a run on its own.

Lost the run_id? The convenience agent_run_complete(last_open=true) closes whatever in-progress run is open for the actor. It is intended for single-session accounts where the open run is unambiguously yours (e.g. the host's approval gate fired after the run_id scrolled out of context). In multi-session accounts — when more than one MCP session shares one OAuth2 credential, e.g. an interactive Claude.ai session plus a scheduled claude -p running on the same account — the call will refuse with a structured -32095 explicit_run_id_required error envelope if the open run was started by a different session. Pass the explicit run_id to confirm the close was intentional. Check status().sessions.total_active_today if you want to know whether you're in a multi-session situation before calling.

One-off tasks (work_on_task)

work_on_task is not the cycle. It drains a queue of ad-hoc tasks the user submitted (via the web app's Builder) for you, the agent, to execute on their behalf. They are not tasks the user is responsible for doing themselves.

Workflow:

1. work_on_task(list_only=true) — see what's queued.
2. work_on_task() — get one context-prepared task (with the user's framing and attached context).
3. Execute it; write the result as a task output.
4. work_on_task(task_id=ID, mark_done=true) — mark done.

The recurring cycle includes a single step (Task Check) that drains this queue inline. Outside a cycle, call work_on_task directly when the user says "drain my task queue", "anything queued?", "pick up the next task", or equivalent.

Inside-cycle vs outside-cycle framing. The ready-task brief that work_on_task returns swaps step 3 based on whether an agent_run cycle is open:

• Outside a cycle — the brief says "Ask the user 2–3 focused questions to fill gaps before producing the output." Use the user's reply as additional context.
• Inside a cycle — the brief says "Flag gaps inline; don't pause." Surface missing context as an ## Open questions section at the bottom of the task output. The user can answer via inline > dashboard comments or re-queue the task — never halt the cycle mid-flight.

Follow whichever step 3 the brief actually carries; the server picks for you.

Tasks the user submits often come from the Task Builder wizard in the web app — it captures richer framing and attaches context the LLM should use rather than re-derive. See task builder.

Outputs (the Wiki)

Outputs are how the agent persists artefacts the user can later read and edit. The user calls this surface "the wiki".

When to read: before substantive task work, search for prior artefacts on the same topic. Prefer output_search(query, category?) (hybrid semantic + keyword) over output_list(category) when you don't know the slug.

When to write: every substantive task should produce at least one output. Email drafts → email (with status: pending frontmatter); research → research; ad-hoc analysis → propose a fresh category name; per-cycle log → log; the rolling action list → actions (call output_dashboard() to fetch or ensure-create the dashboard id, then output_update).

Before minting a new category, call output_categories() to see what already exists. Reuse an existing custom category instead of inventing a near-duplicate (meetings vs meeting-notes etc.).

Always pass title and short_description when you create or update an output — both are real server fields (≤ 200 chars each), surfaced in output_list and output_get. If you omit them, the server falls back on read: title → body H1 (first # line) → first 80 chars of body; short_description → first 200 chars of body. Treat the fallback as a courtesy, not the contract.

Bodies are valid Markdown. Single H1 where appropriate; H2/H3 sub-sections; YAML frontmatter at top for metadata; fenced code blocks; Markdown tables; [text](url) for links.

Link form decision rule. Output references take one of three forms depending on where the text will be rendered:

The absolute app URL for the second form appears already-substituted in the agent_run() preamble (the server expands an {{ACTOR_OUTPUTS_URL}} template into a real URL before sending). Copy that URL as-is; never emit a literal {{ACTOR_OUTPUTS_URL}} token, and don't try to compose the URL from parts. Bare category:id is only valid inside YAML frontmatter or MCP arguments; never put it in rendered prose.

output_search excludes the log category (append-only audit trail; semantic search would surface noise). To list logs, use output_list(category="log") and filter by the date in the slug.

Slug-skip guards (de-dupe before creating)

The wiki rejects duplicate (category, slug) pairs. Before output_create, skip the create when:

• A natural slug like daily-news-2026-05-25 already exists for today — update the existing item with output_update, don't mint a near-duplicate (daily-news-2026-05-25-1).
• A task's procedure says "create one improvement per cycle" — output_list(category="improvement", recency_days=1) first; skip if today's review already exists.
• The user re-asks for an artefact you just produced this session — link to the existing one, don't generate a parallel copy.

When in doubt, output_search(query) first and update what's already there. Skipping a create is a positive outcome — the wiki stays clean and the user's existing link keeps working.

Instructions

The instruction docs (five required, one optional):

• agents — how to behave (standing brief; loaded by agent_run).
• tasks — which recurring tasks run this cycle (user-owned).
• default_tasks — canonical procedures for each default task (system-owned).
• personal — identity, facts, behavioural guidance.
• style — voice, tone, formatting.
• skills (optional) — skill selection guide for domain work.

When inside an agent run, every installed instruction is pre-loaded by agent_run() — the five required ones plus skills if installed. Don't re-call them.

When outside an agent run, call instruction_load(name="agents") first if the user asks about how the agent is configured, or before doing substantive task work that needs the standing rules. The name is the public short name (e.g. agents, tasks, personal) — never the instruction_ storage prefix.

instruction_save and instruction_delete mutate user-owned docs — confirm before writing.

Key Rules (during runs)

• Never send emails or external messages without explicit instruction. Default to drafting (email outputs with status: pending); the user flips status to approved in the web app.
• Never delete memories or outputs without explicit instruction. Update in place or mark for review instead.
• Log everything. One log output per cycle.
• Don't preview, don't partial-run. Execute to completion in a single response.

URL ↔ MCP-tool Mapping

Documents you read may contain absolute web-app URLs. You cannot fetch them over HTTP — they are deep links into the user's web app, not API endpoints. Translate to MCP:

Available Tools

Memory:

• memory_search() — keyword/semantic search; supports last_n, recency_days, include_remote. include_remote=true requires the once-per-conversation user ask — see Shared Memories below.
• memory_get() — retrieve memory details by ID (single or batch).
• memory_save() — store new memories (single or batch, auto-categorized). preview=true to preview.
• memory_update(), memory_delete() — modify by ID.
• memory_types(), memory_create_type(), memory_delete_type() — manage categories.
• how_to_use() — personalized guide. Heavy; first-interaction only.

Outputs — the Wiki (only if enabled):

• output_search(query, category?, limit?) — hybrid semantic + keyword search across categories. Excludes log (append-only audit trail; semantic search would surface noise). To find a prior run log, use output_list(category="log") and filter by date in the slug.
• output_list(category) — list items in a category.
• output_get(id="<category>:<id>") — fetch one item with full body.
• output_create(category, slug, title, content, short_description, ...) — create.
• output_update(id="<category>:<id>", ...) — modify.
• output_delete(id="<category>:<id>") — remove (rarely; prefer update).
• output_dashboard() — fetch or ensure-create the singleton actions dashboard.
• output_categories() — list the categories that currently exist (defaults + any custom ones). Call before minting a new category to avoid near-duplicates.

Instructions (only if enabled):

• instruction_list() — list installed instructions.
• instruction_load(name) — load one by short name (agents, tasks, default_tasks, personal, style).
• instruction_save(name, content, ...) — write a user-owned instruction.
• instruction_delete(name) — remove a user-owned instruction.

Recurring cycle (only if enabled):

• agent_run() — full cycle entry point. Returns instructions + dashboard state in-band; execute immediately.
• agent_run_complete(run_id=...) — call once the cycle finishes to clear the in-progress marker.

One-off task drain (only if enabled):

• work_on_task() — get one context-prepared ad-hoc task; list_only=true to peek; mark_done=true, task_id=ID to close.

Shared Memories (from trusted connections — people or AI agents):

• memory_search(include_remote=true), list_connections() — see who shares what.
• Ask the user once per conversation before searching remote memories. Remember the answer for the rest of that conversation; ask again next session. Attribute matches: "Alice mentioned …"

See shared memories for patterns.

Remote Actions (control devices / trigger workflows on connected services):

• list_connections(), describe_method(), execute_method().
• Confirm with the user before executing unfamiliar methods.

Client-side approval gate. Some MCP clients (notably Claude.ai's web UI) gate list_connections, describe_method, and execute_method behind a per-tool approval prompt — sometimes non-deterministically (eval round 7 saw list_connections and describe_method gated in the same session that execute_method answered without prompting). If the user denies (or the prompt times out), the tool returns the literal string "No approval received." instead of a structured envelope — the server never sees the call. Treat that response as a client-side denial, not a server error: tell the user the call was denied at their client and ask them to grant the connector

…(truncated)