Breakcold CRM

The expert playbook for any AI agent operating a user's Breakcold CRM through the Breakcold MCP server. Optimized for speed, error-proofing, and minimum back-and-forth with the user.

This skill is designed to be universal: the markdown content works the same way whether the agent runs on Claude apps, Claude Code, OpenClaw, Hermes Agent (Nous Research), Codex, Cursor, Windsurf, Cline, Continue, Goose, ChatGPT, Open WebUI, OpenHands, TypingMind, LibreChat, LM Studio, Microsoft 365 Copilot, or any other MCP-capable client. The skill follows the AgentSkills format, so it loads natively on platforms that support it. Installation differs per platform — see DEPLOYMENT.md.

What this skill covers

Six headline workflows the user explicitly cares about, plus the universal plumbing that makes them fast and safe. Open the reference file for the workflow you're running — don't operate from memory.

#	Workflow	Reference
1	Multichannel auto-task creation (follow-ups linked to conversations)	references/action-tasks.md
2	Auto-movement of contacts in the pipeline (Kanban)	references/action-pipeline.md
3	Auto-report (visual, Breakcold-branded, actionable insights)	references/action-reports.md
4	Prospect research (web research → notes on records)	references/action-prospect-research.md
5	Contact detection & creation from the inbox	references/action-contact-detection.md
6	CRM setup & reorganization from a website URL	references/action-crm-setup.md

Anything not listed above is still in scope — the Breakcold MCP exposes the full CRM, inbox, tasks, notes, and metadata surface. The six workflows are simply the ones to deliver perfectly.

Cross-cutting references

• references/fundamentals.md — auth, regions, workspaces, rate limits, attribute types, conversation/record linking, error handling, canonical tool-call examples. Read this once at the start of any complex run.
• references/routines.md — how to schedule any of the six actions, set frequency, and offer daily/weekly recap emails. Linked from every action's "schedule it" step.
• references/branding.md — Breakcold color tokens (primary #0059F7), tag-color palette, asset paths, and visual style for every report or visual deliverable.
• references/report-design-guide.md — the module catalog for Action 3 reports: which sections to include based on what's actually in the user's workspace, plus hero / headline / mascot rules.
• assets/ — real brand assets: Breakcold mark SVG, full logo SVG, mascot WEBP, and a fully populated report-template.html chassis to fill in.

Session-start orientation (mandatory boot sequence)

Run this in one shot at session start. Cache results in your scratchpad — never re-fetch within the same session.

SESSION_START:
  1. capabilities_list()                       → cache scopes
                                                 IF auth error: see fundamentals.md § 9
                                                 IF the response includes the authenticated actor's user ID,
                                                 cache it as CURRENT_USER_ID immediately.
  2. workspaces_list()                         → cache workspace IDs
                                                 IF count == 1: select silently
                                                 IF count  > 1: ask user ONCE, cache choice
                                                 Same as step 1: cache CURRENT_USER_ID if it appears here.
  3. Resolve CURRENT_USER_ID if still unknown   → see fundamentals.md § "Resolving the current user ID".
                                                 Default fallback: read it from the response of the next
                                                 write call (records_create, tasks_create, notes_create,
                                                 custom_activities_create) and cache from then on. The
                                                 user's ID appears as the createdBy / actorId field on
                                                 any write response.
                                                 CURRENT_USER_ID is the default assignee for any task
                                                 the agent creates on the user's behalf (see action-tasks.md).
  4. (lazy, when first action runs)
     crm_objects_list(workspace_id)            → cache object IDs
     crm_fields_list(object_id)                → cache per object as needed
     crm_field_options_list(field_id)          → cache only for select/multiselect you'll touch

Anything outside this boot sequence (action-specific reads) is owned by the relevant action-*.md.

Universal rules — apply to every workflow

These rules exist because the user explicitly asked for speed, error-proofing, and minimal actions. Violating any of them creates work for the user.

• Authenticate once. Never ask the user to re-authenticate mid-session. If a tool call returns an auth error, surface it once with a one-tap reconnect path. Don't loop.
• Cache discovery aggressively. A second call to workspaces_list, crm_objects_list, or crm_fields_list for the same object in the same session is a bug. Reuse the IDs you already hold.
• Verify before mutating. Before records_update, records_archive, tasks_create, tasks_complete, notes_update, or any crm_*_delete, confirm the target ID exists from a prior read. Never trust speculative IDs.
• Never create duplicates. Before creating a record, task, or note, check for an equivalent. For tasks: call tasks_list on the record and bail if an open task with the same intent and a due date in the next 7 days already exists. For records: see the duplicate-detection rules in references/action-contact-detection.md.
• Leave a breadcrumb when it helps. For any non-trivial automated decision (moving a contact, auto-creating a record, choosing a pipeline stage on weak signal), call custom_activities_create on the record with a one-line explanation. This is your shared memory across runs and the user's audit trail. Don't ask permission for this — just do it when it adds clarity for the next run.
• Confirm once, then act. When the user kicks off an automated workflow, confirm parameters in a single message (criteria, frequency, assignee, recap channel). Don't re-prompt for what they already answered.
• Never move pipelines backwards. Treat pipeline stages as a directed flow. See references/action-pipeline.md for the exact rule.
• Ask only when it actually changes the output. Defaults are your friend; surface them as "I'll do X — say so if you'd rather Y." Don't ask about obvious stages or obvious decisions.
• Respect rate limits gracefully. On HTTP 429 / JSON-RPC -32029, wait the Retry-After interval and resume from the same point. Don't restart the workflow. See references/fundamentals.md § 7.
• Mirror the user's language for anything written into Breakcold or shown to the user. Task titles, note bodies, report headlines, confirmation messages, stage names, and view names all match the user's language (and the language already present in their workspace). If a user writes to you in French, write Relancer — fil email en silence (9 jours) not Follow up — email thread went quiet (9 days). The single exception: internal breadcrumb prefixes like [breakcold-crm:auto-tasks 2026-05-23] stay in English so future runs can parse them reliably across users. Everything else translates.
• Conversations live on people — traverse relations before concluding "no signal", and create tasks on the Person where the conversation lives. Deals, Companies, and most non-Person objects don't usually have conversations attached directly. Their linked People do. Whenever a workflow checks for "activity" on a parent object, always also fetch conversations on the linked People records. And whenever a workflow creates a task in response to conversation activity, the task goes on the Person record, not on the parent Deal or Company — so the task can link to the actual conversation. Concluding "inactive" on a Deal whose linked Contacts are actively replying, or creating a task on a Deal that can't link to the conversation, are the two most common false-negatives in this skill. See references/action-pipeline.md step 3b for the signal-traversal procedure and references/action-tasks.md step 2 for the task-placement rule.
• Notes are HTML. When calling notes_create or notes_update, the content field renders as rich HTML in Breakcold. Plain text with \n line breaks comes out as one unbroken paragraph — the user will have to ask you to reformat. Use semantic HTML (<h3>, <p>, <ul>/<li>, <strong>, <hr>) for any note longer than two sentences. Always also provide contentPlain as the plain-text fallback. See references/fundamentals.md § 11 - notes_create for the payload shape and references/action-prospect-research.md step 6 for a structured template.
• Small pages, processed inline. Cap every list call at limit: 20 — records_list, inbox_conversations_search, anywhere the API takes a limit parameter. Process each page completely (filter → decide → mutate → write breadcrumbs) before fetching the next page. Never accumulate more than one page of raw data in working context at once. Why this is a hard rule, not a tip: when a tool response exceeds the runtime's payload threshold (~50KB on several agent platforms), the runtime saves the response to a file and returns a truncated preview. The model, now working from a partial view of its own data, loses confidence and may delegate to a sub-agent, restart the workflow, or stall. A 20-record page stays well under the threshold across every runtime, and the model keeps full context. 20 × ~5KB ≈ 100KB worst-case raw, but the filtering happens inline immediately so the processed set stays tiny. See per-action playbooks for the exact loop structure.
• No sub-agent delegation for any workflow in this skill. Make every tool call directly in the main thread. Do not use any sub-agent / task-delegation tool — by whatever name the runtime exposes it (Agent, Task, dispatch_agent, subagent, delegate, etc.). Sub-agents start with zero cached context, re-discover the schema from scratch, hit the same payload limits the main agent does, and stall on tool-approval prompts that the user can't see. The skill's caching architecture only works in one continuous main-thread context. If you find yourself reaching for a sub-agent because "this loop is long" — that's a signal to apply the small-pages rule, not to delegate.
• Small-batch first for one-shot runs. Default to a first-page-only test run for any one-shot action invocation (auto-tasks, pipeline movement, contact detection, prospect research). Process only the first 20 records, write the breadcrumbs, then surface the result with: "I processed the first 20 — say continue to sweep the rest." This bounds context, proves the pipeline works on real data, and is reversible if anything looks off. The exception is scheduled runs triggered by a saved routine — those proceed end-to-end because the user already approved the cadence at setup time and isn't watching the run live. Apply small-batch-first even when the user's request was specific and explicit (e.g., "Create tasks for any contact I haven't talked to in 5 days") — the upside is huge, the cost is one word ("continue") from the user.

Routing — match the user's phrasing to a reference

The user rarely says "run workflow 3." Match on intent:

User says something like…	Open this reference
"follow up", "ping the cold ones", "tasks for people who went quiet"	references/action-tasks.md
"move contacts forward", "advance my pipeline", "update my Kanban", "auto-progress deals"	references/action-pipeline.md
"give me a report", "how is my CRM doing", "what's working", "deal review", "sales recap"	references/action-reports.md
"research these prospects", "what does this company do", "enrich my contacts"	references/action-prospect-research.md
"find missing contacts in my inbox", "add new people from emails", "who isn't in the CRM"	references/action-contact-detection.md
"set up my CRM", "reorganize Breakcold", "I'm new", "redo my pipeline from scratch"	references/action-crm-setup.md
"set up a routine", "do this every morning", "schedule a weekly recap"	references/routines.md
"what can you do?", "I'm a bit lost", "help me with Breakcold"	See "If the user is lost" below

Same matching applies across languages. "Relancer mes contacts" → action-tasks.md. "Reorganiza mi CRM" → action-crm-setup.md.

If the user is lost or asks what you can do

Don't list every MCP tool. Lead with the six headline workflows — one line each, in the user's language — and let them pick:

Here's where I really shine inside Breakcold:

1. Auto follow-up tasks for contacts where you're losing momentum — every task links straight to the conversation.
2. Pipeline auto-movement based on what's actually happening in your emails, LinkedIn, WhatsApp, Telegram, and meetings.
3. Visual CRM reports that point to what's working and what's not — no fluff, just actionable insights.
4. Prospect research that writes itself back into your notes and deals.
5. Inbox → CRM — find people you're talking to who aren't in the CRM yet and add them.
6. Full CRM setup or reorganization from your website URL, no consultant needed.

Beyond these I can also create or update records, tasks, notes, and custom activities; manage pipeline stages, fields, options, and views; search your inbox; and almost anything else the Breakcold API supports. Where would you like to start?

Recap emails for routines

Any of the six actions can run as a routine. When you set one up, always offer the user a daily or weekly recap by email — a short summary of what got done since the last run. This requires the user's email connector (Gmail, Outlook, etc.) be linked. Follow references/routines.md for the exact wording and setup.

Deliverables outside Breakcold (reports, exports)

When an action produces a visual artifact (most relevant for action 3 — reports), always render it on-brand. Pull references/branding.md first so colors, type, and layout are consistent without thinking.

Installation and deployment

The way this skill gets loaded depends on the agent platform. See DEPLOYMENT.md for platform-by-platform instructions covering Claude apps, Claude Code, OpenClaw, Hermes Agent, Codex, Cursor, Windsurf, Cline, Continue, Goose, ChatGPT, Open WebUI, OpenHands, TypingMind, LibreChat, LM Studio, Microsoft 365 Copilot, and plain API integrations.

---

Reminder. Open the reference file for the action you're running. The references contain the exact tool sequences, defaults, and edge cases. Operating from memory is how mistakes happen — and the user explicitly asked for error-proofing.