> ## Documentation Index > Fetch the complete documentation index at: https://docs.speckle.systems/llms.txt > Use this file to discover all available pages before exploring further. # Reuse organisation playbooks with Intelligence skills > Save reusable instructions for Speckle Intelligence Chat, share them across the workspace, and invoke them with /skill-name. This is 2026.9 preview documentation. Coverage here is incremental: a page exists only where 2026.9 differs or is newly documented. A **skill** is a reusable playbook for Speckle Intelligence. You write the instructions once — how your organisation colours a model, how it prices work, which carbon factors to use, which words to say, or how replies should sound — and anyone who can see that skill pulls it into a conversation with `/skill-name`. Speckle Intelligence can also **draft the skill for you**: describe the playbook in plain language, or save a chat that already works. Skills live on the workspace. They are how you exercise control at organisation scale without re-explaining the same rules in every chat. A skill tuned to the way your company does things also spends fewer AI tokens reinventing that method on every turn: the assistant follows the playbook instead of rediscovering colour legends, cost codes, carbon factors, house lingo, or tone of voice from scratch. Anything you would put in a generic AI chat tool as a custom instruction can be a skill here — brand voice, meeting notes format, translation rules, non-AEC analysis — as well as model-aware playbooks that use Speckle data. For where to ask questions, see [Ask questions about your models with Speckle Intelligence](/next/analytics/intelligence). Skills appear when the workspace plan includes Speckle Intelligence Chat. On a self-hosted server, Chat is available only on Speckle Enterprise Server. For deployment setup, see [Enterprise license — Intelligence](/developers/server/deployment/enterprise-license#optional-intelligence). ## When to use a skill Use a skill when the same behaviour should happen again: a colour theme, a cost or carbon method, a formula, a report shape, a glossary, a tone of voice, or any other repeated instruction you would save in a generic AI assistant. Share it as **Workspace** visibility when the whole organisation should be able to invoke it. The more the instructions match how your company already works, the less each conversation spends credit re-deriving that process. Do not use a skill for a one-off question. Do not use a skill for a standing constraint that must apply to **every** conversation without anyone typing `/skill-name`. That always-on layer is [AI Rules](/next/analytics/intelligence-rules). If you only need the assistant to match the language of your question, ask in that language — Speckle Intelligence usually keeps replying in the same language without a skill or a rule. ## Create a skill Workspace **members** can create skills. Guests can invoke a **Workspace** skill in Chat. They cannot author one, and they usually do not see **Settings → Skills** — they arm skills from the Chat **/** menu instead. Workspace admins can edit or delete any skill in the workspace, including someone else's **Private** skill. You can author a skill three ways. Use Speckle Intelligence when you want a strong first draft; you always review the form before anything is saved. ### Draft with AI Open the workspace, select **Settings**, then **Skills**, then **New skill**. You should see the **Draft with AI** card above the form. In **Draft with AI**, describe what the skill should do in plain language — for example **Audit models for missing fire ratings and report the gaps in a table**, or **Colour the model silver by classification type**. Select **Draft with AI**. Speckle fills **Name**, **Description**, **Instructions**, and related fields. Nothing is created until you save. Edit the draft, set **Visibility** to **Private** or **Workspace**, pin **Knowledge** if the playbook needs a rate card or factor table, then select **Create skill**. You should see the skill on the Skills list. Drafting with AI fills the whole form. If you already typed in the fields, Speckle asks **Replace your edits?** Choose **Draft anyway** only when you want the AI draft to replace what you entered. Drafting spends workspace AI credit the same way other Intelligence actions do. ### Save a chat as a skill In **Chat**, when a conversation already behaves the way you want, select **Save as skill**. Speckle drafts a skill from that thread and opens the create form. Tighten the instructions, set visibility, pin knowledge if needed, then save. Use this when you discovered the method in Chat and want the organisation to reuse it. ### Write the form yourself You can fill the form without a draft. Set **Name** to a slug such as `embodied-carbon` (Chat uses `/embodied-carbon`), write **Instructions** in the organisation's method and lingo, choose **Visibility**, then **Create skill**. ### What you can attach **Knowledge** pins workspace files and data warehouse tables (Snowflake, Databricks, Microsoft Fabric) to the skill. When the skill is active, Chat can draw on those sources — a rate card, an EPD library, a carbon factor sheet, a classification map — without you attaching them every time. A skill can pin up to 20 knowledge sources. Instructions can be up to 8,000 characters. ## Use a skill in Chat Skills work in project **Chat** and in the 3D viewer sparkle. Arming a skill applies its instructions to the **next message** you send. The armed chip clears after that send. It does not rewrite earlier replies, and it does not stay on for the rest of the thread unless you arm it again. Open **Chat** in the project sidebar, or select the sparkle in the main 3D viewer. You should see the composer. Select the **/** control (**Use a skill**), or type `/` and the skill name. Pick `/skill-name`. You should see that skill as a chip above the composer, or as `/skill-name` in the message. Type the question, or pick a conversation starter the skill provides. You should get a reply that follows those instructions — colour legend, cost codes, carbon method, tone of voice, or whatever the playbook defines. If the organisation needs the same playbook on every project, set the skill to **Workspace** visibility and tell people the slash name. Chat does not run workspace skills automatically. The **/** menu also links to **Manage skills** or **Create a skill** when you can author skills. ## What custom skills are for A skill is organisation control you opt into. The catalogue below mixes model-aware AEC playbooks with the same kinds of custom instructions people save in generic AI chat tools. It is not a limit on what you can write. If people keep repeating the same method, lingo, formula, visual rule, or voice, put it in a skill. ### Tone of voice and non-AEC playbooks * Write every reply in the company's brand voice: plain, formal, client-facing, or internal — and keep that voice stable across projects. * Prefer short bullets for directors, and longer technical prose for engineers, as two separate skills rather than one vague assistant. * Translate or localise answers into the workspace language, including how the firm spells product names and place names. * Structure meeting notes, action lists, or decision logs the way the PMO template already does. * Summarise chats for email or a board pack with the firm's subject-line and disclaimer rules. * Explain Speckle data to a non-technical stakeholder without host jargon, or the reverse for a specialist audience. * Apply general analysis habits that are not AEC-specific: always show assumptions, always cite sources, never invent numbers, always end with three next steps. * Turn a pinned company handbook, style guide, or FAQ workbook into **Knowledge** so Chat answers HR, IT, or process questions in the approved words. * Keep a private personal skill for how you like replies formatted, and a workspace skill for the organisation voice everyone should use with clients. ### Colour and view the model * Apply a house colour theme — for example a graduated silver or grey palette by classification type — with clear rules for which host property to read (category, `speckle_type`, type) when models come from different applications. * Colour the loaded model with the company's discipline legend, and keep those colours identical on every project. * Colour by coordination status using the organisation's red / amber / green meanings, not a one-off palette. * Colour by fire rating, acoustic rating, or accessibility category the way the fire engineer and the architect already talk about those values. * Isolate MEP systems with the firm's system names and colours, then restore a clean overview when the walkthrough is done. * Hide furniture, entourage, and site context for a coordination view that matches the BIM execution plan. * Frame typical elevations or a true-north camera so review screenshots look like the office standard, not a random orbit. * Colour by the parameter the firm always uses for phase, zone, or work package, even when the published property name differs between hosts. * Offer darker / lighter / reverse / revert adjustments in the instructions so reviewers can tune the theme without rewriting the skill. * Apply a client-facing colour theme for workshops, and a denser internal theme for design reviews, as two separate skills. ### Cost, carbon, and quantities * Price quantities with the company's unit rates, cost codes, and work breakdown structure, and say those codes out loud in the reply. * Turn room or space data into the organisation's area standard — NUA, NIA, NLA, GLA, GIA, GEA, GBA, or the chargeable area the lease uses — with the same definitions every time. * Apply the firm's measurement rules: wall centreline versus finished face, doors counted in the wall or excluded, and which voids count. * Convert published units into the units the commercial team invoices in, and round the way the cost plan rounds. * Estimate embodied carbon (for example A1–A3) by matching model quantities to a pinned factor table, with category fallbacks and an unmatched list instead of silent guesses. * Separate A1–A3 product carbon from A4 transport and A5 construction when the firm reports that way, and refuse to blend them. * Flag missing rates or missing carbon factors instead of inventing a number, and list the exclusions in the firm's language. * Pin a rate-card spreadsheet or a live warehouse table as **Knowledge** so the skill reads this year's prices, not last year's memory. * Produce a cost or carbon narrative that uses the commercial team's words: elemental, package, provisional sum, contingency — not generic AI phrasing. ### Language, classification, and house style * Use the organisation's names for levels, grids, and storeys, and never swap in a host application's synonym. * Map host categories and IFC types onto Uniclass, OmniClass, or the firm's own classification, and quote the code the document controller expects. * Call coordination issues what the BEP calls them. If the office says "coordination issue", do not say "clash". * Keep spelling, units, and date formats on the regional standard the workspace uses (UK millimetres and DD/MM/YYYY, or US feet-inches). * Address the client by the name in the appointment, and describe the scheme the way the bid did. * Prefer the firm's abbreviations (GA, NBS, RIBA stage, LOD) and expand them only when the audience is external. * Never rename parameters the way a general assistant might. Repeat the published Speckle property names the connectors actually sent. ### Formulas and conversions * Compute steel tonnage from volume and the firm's density, then apply the agreed section utilisation or connection allowance. * Derive door, window, and finish schedules with the takeoff formulas the interior package already uses. * Convert between host units, SI, and the commercial unit of measure in one place so every chat does the same conversion. * Apply percentage complete, earned value, or design maturity rules the project controls team already uses. * Include wastage, lapping, and cutting allowances the estimator uses, and show the allowance as a line, not a silent fudge. * Normalise type marks, family names, and type catalogues into the firm's type-naming pattern before totals. * Recalculate areas when the skill's measurement rule disagrees with a raw property, and say which rule won. ### Reports, tours, and output shape * Structure every commercial reply as executive summary, exclusions, then tables — the same skeleton the director reads on Friday. * Build a saved **Report** from Chat that already uses the firm's rate card and carbon method, then share that report rather than the raw thread. See [Create and share Intelligence Reports](/next/analytics/reports). * Lead a client workshop view sequence: site, envelope, typical floor, services, then risks. * Produce a quantity survey in the column order the cost plan template uses. * Write a change commentary between two versions using the firm's added / omitted / revised vocabulary. * Keep charts in the organisation's preferred breakdown (by level, by Uniclass, by package), not whatever the model tree happens to be. ### Knowledge files and live data * Pin the approved rate card, carbon factor workbook, or product library so Chat does not scrape a random file from the conversation. * Pin a warehouse table for live commercial data so the skill stays current when finance updates the sheet. * Pin a mapping table from host category to company cost code, and instruct the skill to fail closed when a row is missing. * Pin specification extracts or employer's information requirements so answers cite the appointment, not generic good practice. * Pin a legend image or colour dictionary if the visual standard is easier to describe from that file than from prose alone. ### Coordination, quality, and safety * Assess BIM and tabular data readiness for design review, analytics, cost or carbon joins, compliance, handover, or asset reuse — and report gaps in the firm's checklist order. * Turn readiness findings into a versioned conditioning **manifest**: proposed transformations that are explainable, testable, attributable, and reversible, without mutating source data from the skill itself. * Preserve an auditable path from every headline finding to the smallest set of calculations or comparisons needed for the decision (a BIM benchmark style skill). * Review a version the way the BIM manager reviews it: required parameters present, naming pattern held, dummy types called out. * List coordination risks in BEP order (structure vs services, then envelope, then interiors). * Call out fire-stopping, accessibility clearances, or maintenance access with the language the compliance lead uses. * Refuse to answer from versions still in the old data format, and tell the user to publish after the project uses the new format. See [Speckle's data model is changing](/workspaces/data-model-migration). * Always name the version, the publisher, and the source application before any total, so a forwarded answer is auditable. ### Role, region, and programme * Give the quantity surveyor package-based totals, and give the architect room-based totals, as two skills rather than one vague assistant. * Give the structural engineer utilisation and mass, and give MEP flow, load, and system membership, without mixing those audiences. * Switch UK Approved Document language, US code language, or another region's citations by workspace or by skill, not by hoping the model guesses. * Hold a RIBA stage, AIA phase, or in-house gate in the instructions so the assistant does not recommend deliverables that stage does not include. * Keep a private personal skill for how you like replies formatted, and a workspace skill for what the organisation allows everyone to use. If none of those headings fit, write the skill anyway. Model-aware AEC methods and generic chat habits both belong here. The test is repetition: the organisation wants the same control next week, on the next project, in the same words and the same voice. ## Skills and AI Rules * Opt-in playbook people invoke with `/skill-name` — use a **skill**. * Standing constraint on every conversation — use an [AI Rule](/next/analytics/intelligence-rules). * Match the language of this conversation — ask in that language (no skill required). * One-off question — use Chat with no skill armed. Skills do not replace [workspace Settings](/next/workspaces/settings). Create and share skills from **Settings**, then **Skills**. Create standing rules under **AI Rules**. Invoke skills from Chat; rules apply on every turn. ## FAQ A skill is opt-in. Someone must arm it with `/skill-name` or **Use a skill**. For a rule that must apply to every conversation in the workspace, use [AI Rules](/next/analytics/intelligence-rules). Yes, usually. If you ask in a given language, Speckle Intelligence keeps replying in that language for the conversation. Use a skill or an AI Rule when the language or dialect must stay fixed even if someone asks in another language. Only the owner, and workspace admins. Other members do not see it in the picker. A skill you want the organisation to use must be **Workspace** visibility. Guests can invoke a **Workspace** skill in Chat with `/skill-name`. They cannot create, edit, or delete skills, and they usually do not see **Settings → Skills**. Workspace members can create skills. Admins can edit or delete any skill in the workspace. **Skills** in the Settings sidebar appears when you can create skills (workspace member, Chat on the plan, and **AI features** on in workspace settings). Guests still arm shared skills from Chat. If Chat itself is missing, see [Ask questions about your models with Speckle Intelligence](/next/analytics/intelligence). Those conversations continue without it. The history stays. Later messages no longer receive those instructions. On the skill, use **Knowledge**. Pin a workspace file or a warehouse table. When the skill is active, Chat can query that source. Do not rely on pasting the same sheet into every thread. Arming a skill does not spend credit by itself. Sending the next Chat message does, the same as any other turn. A company-tuned skill usually spends less on that turn than reinventing the method in free-form chat. **Draft with AI** and **Save as skill** spend credit when Speckle drafts the form. Viewing a saved report still does not spend credit — see [Create and share Intelligence Reports](/next/analytics/reports). Yes. On **New skill**, use **Draft with AI** and describe the playbook. Or in Chat, use **Save as skill** after a conversation that already works. You always review and edit before **Create skill** saves anything.