Agent Skills
Back to gallery
WI
CoworkCopilot StudioScout

Work IQ Sphere

productivitywork-iqmicrosoft-365data-visualizationpowerpoint

Turn your Microsoft 365 work into an interactive 3D Work IQ sphere, a customer-safe view, and a matching PowerPoint slide.


Description

Build or refresh a personal "My Work IQ" visual from the signed-in user's own Microsoft 365 work (mail, calendar, Teams chats and the OneDrive and SharePoint files they worked on, with Week, Month and 3 months views from one read of their data, about all their work or one customer, programme or topic they name) as an interactive 3D sphere (self-contained HTML), a customer-safe copy and a matching Microsoft-style PowerPoint slide. Every build is a fresh, evidence-checked analysis of that one person's own data. Use when someone says "build my Work IQ sphere", "show my Work IQ", "refresh" or "update my Work IQ", "make my Work IQ slide", "visualise my work graph", "show me the Work IQ sphere demo", or wants to show customers what Work IQ (context, memory, skills) means using their own work. Do NOT use for generic decks, Work IQ product or licensing questions with no personal data, or anyone else's, team or org-wide analytics.

Instructions

Work IQ sphere (skill version 2.0.0)

Turns one person’s Microsoft 365 work (all of it, or one customer, programme or topic they name) into the story of Microsoft’s official Work IQ slide (Business data, then Work IQ: Context, Memory, Skills, then Copilot and agents), told with their own numbers: an interactive 3D sphere with three views of one read of their data (Week: the last 7 days, day by day; Month: the last month, week by week; 3 months: the last three months, week by week), a customer-safe copy and a one-slide PowerPoint version.

“Build my Work IQ sphere” asks nothing up front: it always builds all three views, about all their work.

Read Part A first. <skill> means the folder that holds this SKILL.md. Two reference files in <skill>/references/ hold detail this file points to: workflow_details.md (the final reply shape, the files recipe notes, the own-data check code and presenting) and error_table.md (what to say for every error). Read each one exactly when a step says so.

A. Fresh-analysis rules (non-negotiable; they override everything below)

  1. One person. Call GetMyDetails first, before reading any data (only the file check in “Before anything else: is the skill complete?” comes before it). The signed-in user’s UPN and display name are the only identity. Pass --self <UPN> first, then --self <mail address> only if GetMyDetails shows a different one, then --self "<display name>", to discover.py and count_signals.py; the UPN goes into meta.ownerUpn and (unless they ask for a role instead) the display name into owner. The scripts refuse result files whose Sent Items were not sent by the person, any address they do not send from, a files listing collected for anyone else, and the bundled sample data. Never build for anyone else, even when asked (“do Sam’s sphere”, “build my manager’s”); say “I can only build a Work IQ sphere from your own data. Shall I build yours?”
  2. Only their own data, read in this session by Part E: their mail details, calendar, Teams chat list and the OneDrive and SharePoint files they worked on, plus what they tell you in this chat, a directory lookup of their own manager (B1) and, for the internal module only, a document they name (if you must search for it, confirm the exact file with them before reading it). Never take names, wording or numbers from search results (SearchM365, SearchDrive, the web), saved memories, earlier conversations or project notes, anything made or shared by someone else (including the skill’s author), or the examples folder. Never use a file found by search as a template, a wording source or evidence.
  3. The examples folder is fictional (Sam Taylor, retail). Read it only to see the JSON structure. Never copy its customers, people, programmes, themes, assets, lane names, keywords, questions, answers, outputs, callouts, title, source or wording. If the person asks for the demo, deliver examples/Work IQ Sphere (customer-safe).html and references/Work IQ Sphere Slide Reference.pptx unchanged, side by side, in their own output folder named Work IQ demo (fictional) (never next to the person’s own files, which use the same names), and reply: “Here is the demo. It shows Sam Taylor, a fictional retail specialist, not you and not anyone real. Open Work IQ Sphere (customer-safe).html in Edge or Chrome and press W to switch between Week, Month and 3 months; the slide opens it when both are in the same folder. When you want your own, say "Build my Work IQ sphere".” The demo needs no own-data check (A9 is for builds). Never run the example through the scripts, never pass the sample’s data (the skill keeps only one-way fingerprints of it, which the scripts refuse) with the user’s --self, and never copy anything from examples/ into working/ or hand it over as the user’s.
  4. Discovery first, evidence always. Every node and output cites the user’s own discovery items in "from" (exact ids from working/discovery.json), or "signal:<key>" with hits, and its name must match that evidence: a person by their own name, an escalation by its number, a customer by its organisation’s name or domain, a programme, theme, asset or product by a word from its meetings, threads, sent items, files or sites. If you translate a subject word, put the original word in "aka" (for example "aka": ["Lizenzfragen"] for “Licensing questions”); an aka counts only when it appears in the evidence. Something the user told you in this conversation that their data does not show goes in as "stated:<their words, naming it>": only for programmes, themes, products, assets and outputs, never for customers, people or escalations, at most 3 nodes (or a tenth of the nodes, if more) and 2 outputs, listed in your reply for the user to confirm, and asked about again at every refresh. If asked to add a customer, person or escalation the data does not show, say “I can only add customers, people and escalations that appear in your own mail, calendar, chats or files.” "source:<document>" is only for "internal": true items taken from a document the user named.
  5. Every number comes from script output or code over it. No mental arithmetic, no estimates. The build refuses any number of 100 or more in the text that is not one of the user’s own counts, and any date outside their window. Text that every view shows holds no typed-in count at all: it uses placeholders the build fills for each view (step 9).
  6. Stamp the config: "meta": {"ownerUpn": "<UPN>", "skillVersion": "2.0.0", "generatedAt": "<UTC now>", "choices": {"window": "quarter", "focus": {"type": "all"}}}. window is always "quarter": the config is the three-month run and the week and the month are its views. focus stays {"type": "all"} unless the request itself names one customer, programme or topic (step 6).
  7. An ERROR means fix the content, not the check. Never invent evidence; never hand-edit meta, discovery.json, signals.json, the files in working/views/, files.json, the template, the scripts or an output file. Part F says what to tell the person.
  8. Saved setups. Read a saved setup only from the user’s own OneDrive at Documents/Cowork/work-iq-sphere/workiq_config.json, by exact path, never through search. Use it only if meta.ownerUpn equals the signed-in UPN and meta.skillVersion is 1.6.0 or later, and pass it to build_workiq.py --previous (Refresh), which refuses anyone else’s. A Refresh keeps its entities, aliases and weights (Refresh step 3) but never its window or focus: every build has all three views, and it focuses only when the refresh request itself names a customer, programme or topic. Never ask about the window or the focus, whichever version saved the setup. Otherwise say exactly “Your saved setup was made by an earlier version or by someone else, so I’m building fresh from your own data.” and do a New build; from an earlier setup of their own you may keep only the safe aliases of customers, people and escalations the new discovery finds again. A setup attached to the chat, shared by someone else or found anywhere else is never used.
  9. Own-data check before delivery. For every build, refresh or change, step 11 must print OWN-DATA CHECK: OK before any file is delivered.
  10. Colleagues may share customers or programmes; that is fine when the user’s own data shows them. Wording, numbers, questions and outputs are still written fresh.

B. Customer-safe and labelling rules

  1. Customer-safe by default. Every customer, person and escalation gets a unique safe alias that differs from its real name and contains none of its own name words or abbreviations (the “Acme” in “Acme Health”, a first name, a surname, an escalation number). Sector words are fine (“Healthcare Customer A”, “Research Customer B”, “Customer sponsor B”, “Colleague C”, “Escalation 2”); the build refuses an alias that gives the real name away. Use a role such as “My manager” only after checking it (ExecuteSql SELECT * FROM manager WHERE employee = '<UPN>').
  2. Tokens only. In every free-text field (detail, answers, grounding, outputs, memory, exampleChain, internal cards, slide text) refer to customers, people, escalations and file-based items only through {entity_id} tokens. Never type a person’s or organisation’s name, an email address or domain, an escalation number, a file name or a site name.
  3. Files. The names of the user’s SharePoint sites and files never appear in customer-safe files. A programme, theme or asset named exactly like one of their sites or files (or whose name contains one) needs a safe alias in plain words of your own (“Works council programme” for a site called “Works council approvals”), which the customer-safe files show instead; word output titles as what was produced (“A licence options tool”), not as the file name; and never type a site or file name in any free text. A programme, theme or asset whose name contains a customer’s, person’s or organisation’s word also gets a safe alias. The customer-safe build and slide refuse anything else and name what to fix. Personal files are left out when collecting.
  4. Internal-only content (pipeline, seat targets, manager notes, anything from a sensitivity-labelled document) goes only in the internal block or in entities and questions flagged "internal": true. Never flag outputs or memory callouts internal (the presenting file shows all of them); put internal-only output detail in internalText. Tell the user which sources were labelled.
  5. Say exactly what was counted: “emails” (junk, drafts and duplicate copies removed), “calendar events” (not meetings attended), “Teams chats (group and 1:1)” active in the view’s window (meeting chats excluded), “distinct email senders” (not people), “files I worked on (OneDrive and SharePoint)” (files they edited, created or opened in the window, from their own recent files list, so a floor rather than a census), “edited or created” only for the files they edited or created, “opened” for the rest, and “SharePoint and Teams sites” for sites. Theme counts are keyword matches on subjects and on the names of files the user edited or created: direction, not a census. A trend names its period (“rose about 3x in October”, not “since September” if it later eased).
  6. Totals: signals = emails + calendar events + Teams chats, plus files I worked on when signals.files exists (the scripts, the sphere and the slide all add them the same way, in every view).
  7. Complete or labelled: show every escalation found, or say “N of M shown”. Never pad a class.
  8. No competitor names in the visual. No em dashes or en dashes anywhere (the scripts refuse them). In chat replies use safe aliases and bold customer and escalation aliases.

C. What the person gets

File What it is Who may see it
Work IQ Sphere.html Interactive 3D sphere with three views of the same data: Week (the last 7 days, day by day), Month and 3 months (week by week). W switches the view and every number, node size, caption and finding follows it. Opens on Microsoft’s official Work IQ slide, then zooms into the person’s own data on the 3 months view (with a focus, the focus sits next to “Me”, labelled as the focus). Version 2.0 tells the story of Microsoft’s public Work IQ whiteboard: Same model (three identical AI models, the person’s own dots under one), then their own data: eight named sources (CRM and websites greyed, not in the sphere), Week as “My working set”, One item (their most active item, drawn only with the relationships their data shows, the rest greyed), Memory with what they told Copilot, and Connector or Work IQ (one call per item, or a single pass). Quotes are credited to Charles Lamanna, Microsoft, never shown as their results. It also folds in the knowledge spine idea (Jared Spataro, Microsoft, Oct 2026): shared knowledge updated once, used by every agent. Starts customer-safe; S or the switch at the top right shows real names (an amber frame shows while they are visible). The switch stays visible on every screen, the opening slide, help and kiosk mode included. The presenter’s own screen. Never send it: real names are inside the file.
Work IQ Sphere (customer-safe).html Built with --customer-safe: real names, ids and internal content removed from the file in every view and from the search; the switch stays visible but is locked on, as a “Customer-safe file” badge. Safe to send to customers.
Work IQ Sphere Slide.pptx One customer-safe slide in Microsoft corporate style with the 3 months numbers (“The last 3 months” in its subtitle), plus a hidden slide with Microsoft’s official Work IQ slide. The Business data card shows a files row (“files I worked on”, OneDrive and SharePoint) when files were counted. The sphere and “Explore in 3D” open the customer-safe HTML. Safe for customer decks.
Work IQ Sphere Slide (internal).pptx Only on request: real names, links to Work IQ Sphere.html. Internal only.
workiq_config.json The person’s curation, saved in their own OneDrive for refreshes. Never share.

D. Talking to the person

The person is usually a seller or specialist, not an engineer, and may be sharing their screen. Keep messages short and plain, use their language (same shapes, file names unchanged), put no names in progress updates, and never paste script output or JSON into the chat. Never ask which time window or which focus they want: every sphere has all three views, and a focus comes only from a name in their own request.

  1. Opening message, right after step 1 and before reading any data: “I’m building your Work IQ sphere from your own Microsoft 365 data, : mail, calendar, Teams chats and the OneDrive and SharePoint files you worked on. It will have three views you can switch between: Week (the last 7 days, day by day), Month (the last month) and 3 months (the last three months, week by week), all from one read of your data. It usually takes 10 to 15 minutes, most of it listing your Teams chats. I only read details such as subjects, senders, dates and file names (no message bodies or file contents). Nothing is sent to anyone, and nothing from anyone else’s sphere or the sample is used. I’ll post a short update after each step.” is “covering all your work”, or “focused on “ followed by the name exactly as the person wrote it in their request (never a name they did not type). For a refresh of a setup saved by version 1.8 or later, use instead: “Refreshing your Work IQ sphere from your saved setup of , , with fresh Week, Month and 3 months views: mail, calendar, Teams chats and the OneDrive and SharePoint files you worked on.” and continue with “It usually takes 10 to 15 minutes …” as above. For a refresh of a setup saved by version 1.6 or 1.7, say plainly that the views are new: “Refreshing your Work IQ sphere from your saved setup of , . Your sphere now has three views you can switch between: Week (the last 7 days, day by day), Month (the last month) and 3 months (the last three months, week by week), all from one read of your mail, calendar, Teams chats and the OneDrive and SharePoint files you worked on.” and continue with “It usually takes 10 to 15 minutes …” as above. After an unusable saved setup, start the opening message with the sentence from rule A8, then continue with the new-build opening.
  2. Progress updates (only these numbers, each straight from script output, with thousands separators such as 1,803; never in years):
    • after step 3: “Update 1 of 6: your mail and calendar for the last 3 months are read. Next: your Teams chats, the slow part (about 5 to 10 minutes).”
    • after step 4: “Update 2 of 6: all of your Teams chats are listed. Next: your OneDrive and SharePoint files.” If listing could not finish: “Update 2 of 6: your Teams chats could not be listed in full, so they won’t be counted this time. Next: your OneDrive and SharePoint files.”
    • after step 5: “Update 3 of 6: your OneDrive and SharePoint files are listed (

      personal files left out). Next: finding what your work is about.” If files could not be read: “Update 3 of 6: your files are not included this time (). Next: finding what your work is about.”

    • after step 6: “Update 4 of 6: emails, calendar events, active group and 1:1 Teams chats and files you worked on in the last 3 months analysed, with the last month and the last 7 days worked out from the same data; escalation numbers found. Next: choosing what goes on your sphere.” (Leave out the files part when files were not read.) When a focus named in the request was not found, add before “Next:”: “I couldn’t find the customer, programme or topic you named in your own mail, calendar, chats or files for the last 3 months, so your sphere covers all your work.”
    • after step 9: “Update 5 of 6: signals counted in the last 3 months ( in the last month, in the last 7 days); entities and relationships chosen, each backed by your own data. Next: building the sphere and the slide.” With a focus, put “ of the 3 months signals are about your focus.” before “Next:” (no name).
    • after step 11: “Update 6 of 6: sphere with its three views, customer-safe copy and slide built and checked against your own data. Saving your setup.” Count entities and relationships the way the slide does: no hub and nothing internal.
  3. When something goes wrong, say the plain-English line from Part F (references/error_table.md), fix it and carry on. Never ask the person to edit code or files.
  4. Final reply: before you reply (step 13), read “Final reply (D4)” in references/workflow_details.md and reply in exactly that shape, numbers table included (numbers from script output written with thousands separators, for example 1,803; safe aliases; no dashes): “Your Work IQ sphere is ready.” with the fingerprint, the Week, Month and 3 months numbers table, “What your Work IQ remembers”, “Your files”, “Check before you present”, “Presenting” and where the setup is saved.

E. Workflow

Work in the session’s working/ folder, never inside <skill> (nothing personal may be written into the skill folder). Run every script as python -B "<skill>/scripts/<script>.py" .... Every step ends with a Check you can verify.

Before anything else: is the skill complete?

Before any build, refresh, small change or demo, and before GetMyDetails or any data read, list the skill folder and confirm these files exist in <skill>: scripts/window.py, scripts/discover.py, scripts/count_signals.py, scripts/build_workiq.py, scripts/fill_slide.py, scripts/wiq_data.py, scripts/focus.py, scripts/collect_files_recipe.py, scripts/guard_data.json, templates/workiq_sphere_template.html, templates/official_workiq_slide.jpg, references/Work IQ Sphere Slide Reference.pptx, references/workflow_details.md, references/error_table.md and examples/example_config.json (for a demo, only the two demo files A3 names). If any is missing, say exactly this and stop, then add one line naming the missing folders (for example “Missing: scripts, templates.”): “Your Work IQ sphere skill is only partly installed, so I’ve stopped before reading any of your data. It needs the whole work-iq-sphere skill, not just SKILL.md. To fix it, add it again in one of two ways: in Copilot Cowork, go to Customize > Skills > Upload skill and choose the work-iq-sphere.zip file (don’t attach the zip here); or extract the zip on your PC (right-click it, then Extract All), delete the work-iq-sphere folder in OneDrive under Documents > Cowork > skills, and drag the whole work-iq-sphere folder into skills. Either way it shows SKILL.md plus four folders: examples, references, scripts and templates. When it has finished uploading or syncing, start a new conversation and ask me again.” Read no data, build nothing and deliver nothing. Never recreate, rewrite or download a missing file from memory, the chat, an attachment or anywhere else: the scripts are what keep real names out of the customer-safe files. Installing a skill from a zip is not part of this skill.

New build

Ask nothing before you start: no question card, no choice of time window, no choice of focus. Every build reads the last three months once and makes the Week, Month and 3 months views from that one read. The focus is all their work, unless the request itself names one customer, programme or topic (“build my Work IQ sphere for Contoso”, “about the store pilot”): then step 6 finds it with focus.py. Never ask about a focus and never offer one. A request for one window (“for the last month”, “this week”) still builds all three views; tell the person in the reply that W switches to it, or that adding #month or #week to the end of the file link opens on it. Other windows (the last 6 weeks, since 1 September) are not offered: say “Your sphere has Week, Month and 3 months views, all from the same read of your data.”

  1. Who and when. GetMyDetails gives the UPN, the mail address and the display name. Take today and now in UTC (for example 2026-10-21 and 2026-10-21T11:40:00Z) and work the windows out with the script, never by hand: python -B "<skill>/scripts/window.py" --today <today> --views It prints one JSON line with views.week, views.month and views.quarter, each with start, end (the day after the window), queryEnd, unit and columns (the week: 7 days, shown day by day; the month: 5 weekly columns; the quarter: 14 weekly columns, this partial week included), label (the window label), flags (that view’s window flags for steps 6 and 8), files, period and partial (wording for the messages). The data is read once, for the quarter, which holds the week and the month: in steps 2 to 5, <start>, <end> and <queryEnd> are views.quarter.start, views.quarter.end and views.quarter.queryEnd (the calendar query runs to the end of this week, so meetings later this week can be listed for “prep me for” questions; they are never counted). Send the opening message. Check: "status": "ok"; views.quarter.start is a Monday; views.quarter.end is after today; views.week.unit is day.
  2. Mail (ExecuteSql, Microsoft 365 target): SELECT TOP 5000 internetMessageId, subject, receivedDateTime, sender, parentFolderId FROM messages WHERE receivedDateTime >= '<start>' AND receivedDateTime < '<end>'. If it returns 5000 rows, repeat with receivedDateTime >= '<start>' AND receivedDateTime <= '<oldest receivedDateTime in the previous result>' until a result has fewer than 5000 rows (the scripts remove the overlap). Keep every result file path. Then SELECT id, displayName FROM mailFolders and copy that result file to working/folders.jsonl. If the mailbox is not in English, use code to rename the junk, drafts, outbox and sent folders’ displayName in working/folders.jsonl to exactly Junk Email, Drafts, Outbox and Sent Items (the scripts recognise only these names). If Exchange returns a legacy X500 address (starting /O=EXCHANGELABS/...) as the sender of mail in Sent Items, the scripts count it as sent by the person automatically; never edit the result files or pass the X500 address as --self. Check: the last mail result has fewer than 5000 rows (never present a capped total); working/folders.jsonl has a Sent Items row.
  3. Calendar: SELECT TOP 2000 subject, startDateTime, isCancelled, responseStatus, attendeeCount, attendees FROM calendarView WHERE startDateTime BETWEEN '<start>T00:00:00Z' AND '<queryEnd>T00:00:00Z'. If 2000 rows come back, query each half of the window instead and pass both result files. Check: every calendar result has fewer than 2000 rows. Post Update 1.
  4. Teams chats. In a tool script, call ListChats with no filter and top=50 and follow every next link until there is none (mailboxes often hold 10,000+ chats; never stop at a fixed page count). After every page, save the rows so far and the next link to working/chats_partial.json, so that after a timeout the script resumes from the saved link. Save [{topic, chatType, lastUpdatedDateTime}] for every chat to working/chats.json. Both scripts compute the chat count of each view from this file (group and 1:1 chats active in that view’s window; meeting chats are excluded). If listing cannot finish, leave --chats-file out of every script run and say Teams chats were not counted. Check: the last page had no next link; working/chats.json holds one row per chat listed. Post Update 2.
  5. Files (OneDrive and SharePoint). Run the read-only collection recipe as a tool script, passing the whole text of <skill>/scripts/collect_files_recipe.py as the code: RunScript {"path": ".aether/scripts/collect_files.py", "code": "<the text of <skill>/scripts/collect_files_recipe.py>", "args": {"start": "<start>", "end": "<end>", "out": "working/files.json", "report": "working/files_report.json"}} Add "exclude": "<regex>" to the args for anything the person asks you to leave out (matched on file and site names). Add "onedrive_search": true only if they want documents in their own OneDrive that their recent list missed (a partial sample). The recipe is read-only and reads only the person’s own recent files list; what it reads, leaves out and writes is in “Files recipe (step 5)” in references/workflow_details.md.
    • Graph allows only a limited number of failed calls in one conversation. Never retry a failing call in a loop: the recipe stops its own lookups after a few failures.
    • If the tool script fails or is refused (for example because the conversation reached that failure limit), or the recipe reports that it could not read the recent list, build without files: leave --files out of every script run, give the plain reason in Update 3 and list files under “Left out” in the reply. Check: the run completed; working/files.json has meta.owner equal to the UPN; if files_report.json says "status": "partial", tell the person the files count may miss the earliest weeks. Post Update 3 with skipped.personal from the report.
  6. Discover, first for the three months: python -B "<skill>/scripts/discover.py" --messages <mail result 1> [--messages <mail result 2> ...] --calendar <calendar result> [--calendar <second half>] --folders working/folders.jsonl --chats-file working/chats.json --files working/files.json <quarter flags> --self <UPN> [--self <mail address>] --self "<display name>" [--own-domains "<d1>|<d2>"] [--exclude-domains "<regex>"] --today <today> --now <now> --out working/discovery.json <quarter flags> is exactly views.quarter.flags from step 1 (for example --start 2026-07-20 --weeks 14). Leave out --chats-file or --files when step 4 or 5 had no result. It prints one JSON status line, then the top meetings, threads, organisations, people, things sent repeatedly, escalation numbers, terms, upcoming meetings, files (only files the person edited or created) and sites (SharePoint and Teams sites holding them). Files the person only opened are counted but never listed: they are not evidence. The printout shows labels only: take ids from working/discovery.json, where top.<kind> lists ids in rank order and items.<id> holds each label and its counts. If the organisation list shows the user’s own company under another domain, add --own-domains. Personal webmail domains (outlook.com, gmail.com, hotmail.com and the like) are left out of the organisations automatically; for other personal domains (family, school, clubs) add --exclude-domains; pass exactly the same values in every run of steps 6 and 8. Then the week and the month views of the same data, once with <view> week and once with <view> month, <view flags> being exactly that view’s flags from step 1 (for example --start 2026-10-15 --unit day --columns 7 for the week): python -B "<skill>/scripts/discover.py" --messages <mail result 1> [--messages <mail result 2> ...] --calendar <calendar result> [--calendar <second half>] --folders working/folders.jsonl --chats-file working/chats.json --files working/files.json <view flags> --self <UPN> [--self <mail address>] --self "<display name>" [--own-domains "<d1>|<d2>"] [--exclude-domains "<regex>"] --today <today> --now <now> --view-of working/discovery.json --out working/views/discovery_<view>.json Apart from the window flags, --view-of and --out, every argument is exactly the three-month run’s (same result files, --chats-file, --files, --self, domains, --today and --now). A view run counts the three-month items in its own window, so the build can size, show or hide each node in that view; never choose entities from it. Check: every run prints "status": "ok"; note the three-month fingerprint, emails, calendar_events, chats_in_window, by_kind.escalation and, with files, files.worked_on and files.edited. If the week or month run says no mail or calendar rows fall inside the window, that view has no activity: leave it out (no step 8 run and no --view for it in step 10) and say so under “Left out”. Post Update 4 (chats = group + oneOnOne). Focus (only when the request itself names one customer, programme or topic; never ask about a focus and never offer one). Find it in their own three-month discovery with the script, never by hand: python -B "<skill>/scripts/focus.py" --discovery working/discovery.json --type <focus type> --name "<the name as they wrote it>" <focus type> is customer for a customer, or programme for a programme or topic (if a programme or topic is not found, run it again with theme, which also looks at their busiest subject terms).
    • "status": "ok" gives name, item and evidence (the discovery ids the focus entity cites in step 7) and flags (the focus flags for step 8).
    • "status": "not_found" means their data does not show it: never ask, never pick one of its closest candidates, and never focus on something their own discovery does not show. Build for all their work (meta.choices.focus {"type": "all"}) and say so in Update 4. Save the focus as meta.choices.focus: {"type": "<focus type>", "id": "<the id of the entity you make for it in step 7>"}. Never show focus.py’s candidates or closest matches in updates or files.
  7. Choose entities, only from the three-month discovery: about 45 to 65 in 7 classes (fewer for a small mailbox; never pad), plus exactly one hub (id me, real = display name, safe = “Me”). One set of entities serves every view: the build hides a node in the Week or Month view when it had no activity there and sizes every node for each view. Each cites "from", and its name must match what it cites (A4); people must cite person: items and escalations esc: items (the build refuses other evidence for those two).
    • customer (sector only when obvious): customer organisations from org: items, up to 25. Skip items marked “looks automated”, newsletters, media, suppliers and personal contacts. Fewer than 12 is fine; never pad. A site: item named after the customer (where the user edited at least 2 files) or a file: item may add evidence.
    • person: 6 to 10 top collaborators and customer sponsors from person: items.
    • programme: 3 to 6 initiatives from recurring meeting: items, busy thread: items, group chat: items, or a site: or file: item whose name shares a real word with the programme.
    • theme: 4 to 6 recurring concerns from term: and thread: items (not file: or site:: file names already count in terms and keyword series).
    • product: Microsoft products named in their subjects or file names (term: items or a p_ series).
    • uat (escalations): every esc: item in the window. Only 6-digit Action or UAT numbers are found; none is fine.
    • asset: things the user or their agents produced: sent: items, file: items (every listed file is one the user edited or created), or stated:.
    • Size every node. Node size and labels on the sphere follow the entity’s signal, else its numeric weight; an entity with neither is drawn too small to be labelled. So every entity without a "signal" gets a "weight", computed with code from discovery: a person round(min(8, 4 + (meetings + mails) / 60), 1) from its person: item; a programme, theme or asset without a signal 5. The build prints a WARNING naming any entity that has neither. Give every customer, person and escalation a safe alias (B1), and every entity whose name contains a customer’s, person’s or organisation’s word (B3). Each entity has a one-line detail (tokens only, its counts as placeholders, step 9) and, where a keyword series exists, "signal": "<key>" (bubble size follows it). Leave out personal, family and leisure items (focus time, private appointments). Add 40 to 60 relations [from, to, verb] (leads, runs at, raises, escalates, answers, adopts, for, works at, co-sells with, manages, supports, evidences, prepared for, edits). A focused sphere (meta.choices.focus has an id) shows the focus and what connects to it, nothing else:
    • Make the focus entity first: the name focus.py printed, citing its evidence, class customer for a customer focus and otherwise programme or theme as chosen, with the focus id. It gets a safe alias like any customer (or B3’s alias when its name holds a site, file or organisation word).
    • Choose only entities whose own evidence is about the focus: meetings, threads, chats, sent items, files or sites whose label names it or (a customer) that involve its mail domains, people at its meetings or from its organisation, and themes and products whose keyword series has hits about it (signals.focus.grid, after step 8). Expect about half the numbers above; never pad.
    • Link every entity to the focus, or to something linked to it (2 relationships at most; links through me do not count), with relationships the data supports, and add ["me", "<focus id>", "focuses on"].
    • A customer focus shows no other customer and nothing that cites another organisation or is named after one: other customers stay off. A programme or topic focus may show the customers that its own evidence involves.
  8. Keywords and counting. Write working/keywords.json from the user’s own discovery terms (case-insensitive regular expressions): c_<customer> for the biggest customers, t_<theme> for themes and programmes, p_<product> for Microsoft products. Product names go only under p_. Catch-all patterns, and keywords copied from the sample that the user’s own subjects hardly match, are refused. Then count the three months: python -B "<skill>/scripts/count_signals.py" --messages <mail result 1> [--messages ...] --calendar <calendar result> [--calendar ...] --folders working/folders.jsonl --chats-file working/chats.json --files working/files.json <quarter flags> --keywords working/keywords.json [<focus flags>] --self <UPN> [--self <mail address>] --self "<display name>" [--own-domains "<d1>|<d2>"] [--exclude-domains "<regex>"] --now <now> --out working/signals.json Use exactly the same result files, --chats-file, --files, --self and domain flags as step 6. With a focus, <focus flags> is exactly the flags focus.py printed (for example --focus-type customer --focus-regex '\bcontoso\b' --focus-domains 'contoso.com'): signals.json then also holds focus, the emails, calendar events, chats and files about the focus per column and every keyword series inside them (a customer: sent from or attended by its domains, or its name in a subject, chat title, file or site name; a programme or topic: its words there). Matching is by name, domain and keywords, so these counts show direction, not a census. Then the week and the month counts (only the views step 6 made), once with <view> week and once with month, with that view’s <view flags> and the same keywords, focus flags and everything else: python -B "<skill>/scripts/count_signals.py" --messages <mail result 1> [--messages ...] --calendar <calendar result> [--calendar ...] --folders working/folders.jsonl --chats-file working/chats.json --files working/files.json <view flags> --keywords working/keywords.json [<focus flags>] --self <UPN> [--self <mail address>] --self "<display name>" [--own-domains "<d1>|<d2>"] [--exclude-domains "<regex>"] --now <now> --view-of working/signals.json --out working/views/signals_<view>.json Check: every run prints "status": "ok" with the fingerprint, emails and calendar_events of the same view’s step 6 run; with files, files_worked_on equals that run’s files.worked_on; with a focus, the three-month status line’s focus totals are below the totals for all your work and above 0 (0 means the focus flags do not match their data: run focus.py again). Put the whole three-month signals.json into the config unchanged as signals, and its weeks list as the config’s weeks: signals.meta.digest covers every number, so an edited, extended or pasted signals block is refused. The week and month files stay exactly as count_signals.py wrote them: the build reads them in step 10.
  9. Story content, computed with code from the three-month signals.json and discovery.json:
    • Counts in text every view shows are placeholders. The entities’ detail, the Skills questions, answers and grounding lines, the outputs (title and text) and text.exampleChain show in all three views, so they never type a count, a date or a month: the build fills a placeholder for each view, and refuses a typed-in count there, or a date or month outside the last 7 days. Write {#<item id>.<field>} for an item of working/discovery.json, with a field its kind has (meeting: count, weeks, first, last; thread: count, in, out, first, last; sent: count, first, last; org: count, mails, meetings, people, first, last; person: count, mails, meetings, first, last; escalation: count, first, last; chat: count, last; term: count; file: count, last; site: count, edited, last); {#signal:<key>.<stat>} for a keyword series (sum its hits in the view, max its busiest day or week, last today or this week, nonzero the days or weeks it was active, peak “in the week of 12 Oct” or “on Mon 19 Oct”); {#period} (“the last 7 days”, “the last month” or “the last three months”), {#unit} (day or week) and {#current} (today or this week). A count can carry its noun, {#sent:<id>.count|time|times} (“1 time”, “3 times”), or |unit for days or weeks ({#signal:<key>.nonzero|unit} gives “4 days” or “1 week”). first and last fill in a date; a sentence whose date does not exist in a view (nothing happened there) is left out of that view. For example a customer’s detail: "{#org:contoso.com.mails|email|emails} and {#org:contoso.com.meetings|calendar event|calendar events} with {#org:contoso.com.people|person|people} there."
    • lanes: up to 8 keyword series (t_ or p_ keys), each {key, name, color}; a lane’s name must describe what its keyword finds.
    • memory: 3 to 6 callouts of the three months {text, lane, week} (lane and week are positions in lanes and weeks, counting from 0): “constant every week”, “spiked in the week of X”, “rose about Nx in ”, comparing averages with code. Only the 3 months view and the slide show them, so they keep numbers typed from code; the build works out the Week and Month callouts itself. Keep each under about 110 characters so the slide can show one whole (or set text.slideMemory).
    • questions: 4 Skills questions the user would really ask Copilot, each naming at least one of their entities with a {token}, each a different kind with from listing the discovery items that kind rests on: prep (an upcoming: item; upcoming meetings list their attendees), latest (a meeting:, thread: or signal: item for their busiest programme), cares (an org:, thread:, meeting: or signal: item for a named customer), change (the signal: lane behind a callout), escalations (an esc: item), draft (a sent: item they send repeatedly), involve (a meeting: item), promised (a sent: or thread: item), document (a file: item: where the latest version of a document they edit lives and when it changed). Each has highlight ids, 3 to 6 answer lines built only from discovery and signals, and a grounding line with counts (placeholders). Two questions that share most of their words are refused. Write in your own words. Optional fifth question with "internal": true.
    • outputs: 3 to 5 real things produced in the three months, each with "from" (sent:, thread: or file: items, or stated:); titles use a token or a generic name, never a raw file name. Fewer is fine; never pad. {"title": "This visual", "text": "...", "from": ["builtin:this-visual"]} is allowed.
    • portals: 1 to 4 Copilot experiences evident in their data (for example Microsoft 365 Copilot, Copilot Studio, a named agent of theirs); omit it to use the official slide’s three.
    • text, for the 3 months view and the slide (the build writes the Week and Month ones): kicker, headline (a short line in your own words for three months), period (views.quarter.period), windowLabel (exactly views.quarter.label; the build checks it), sparkRange, chatLabel, orgLabel with orgMetric ("orgs" external organisations, the default; "pub_orgs" UK public sector bodies), memoryNote (name views.quarter.partial, the partial week), exampleChain (tokens and placeholders only), optional uatLabel (the name the user’s organisation gives escalations) and slideMemory. Every number and date in the text fields, including slideTitle, slideSubtitle, slideMemory and slideSkills, is checked against the user’s own counts and window, and portal names must be backed by their data.
    • owner (display name, or a role such as “A Microsoft specialist” to keep their name off the files), title (“My Work IQ, <views.quarter.label>”; the build shows each view’s own label), source (one sentence naming the person’s sources and window, for example “Source: ’s Outlook mail (junk and duplicate copies removed), calendar, Teams group and 1:1 chats and the OneDrive and SharePoint files they worked on, <views.quarter.label>.”; name only the sources actually read), meta (A6). Owner, title and source may not name anyone else, and hold no placeholders.
    • Optional internal block (only when asked): nav, kicker, headline, intro, 4 kpis, up to 4 cards, note, total, totalLabel, legend (needed when pillars use live), pillars [{name, value, group, live}], from a document the user names (read it with ReadFileContent, cite it as "source:<document>", note its sensitivity label).
    • With a focus: lanes are keyword series with hits about the focus (the sphere draws them, and sizes nodes, from signals.focus.grid); compute the callouts from signals.focus.grid too and say so (“peaked in the week of 21 Sep, 5 signals about my focus”); every question names the focus with its {token} or highlights it (prep for an upcoming meeting with it, cares for a customer focus, latest for a programme); outputs are things produced about it; headline names it with its token (“Three months of {}, understood.”), and so does slideTitle when you set one, so the customer-safe files show its alias. Business data and the slide show the focus counts by themselves (“145 emails about (of 999 in all my work)”), labelled as matched by name, domain and keywords. A small mailbox may support fewer entities, callouts and questions; the slide needs at least one callout and one customer-safe question with a grounding line. Assemble working/workiq_config.json with the structure of examples/example_config.json. Post Update 5.
  10. Build both spheres with every view step 6 made (leave out the --view of a view that was left out): python -B "<skill>/scripts/build_workiq.py" --config working/workiq_config.json --discovery working/discovery.json --view week working/views/discovery_week.json working/views/signals_week.json --view month working/views/discovery_month.json working/views/signals_month.json --out "working/Work IQ Sphere.html" python -B "<skill>/scripts/build_workiq.py" --config working/workiq_config.json --discovery working/discovery.json --view week working/views/discovery_week.json working/views/signals_week.json --view month working/views/discovery_month.json working/views/signals_month.json --out "working/Work IQ Sphere (customer-safe).html" --customer-safe python -B "<skill>/scripts/fill_slide.py" --config working/workiq_config.json --discovery working/discovery.json --out "working/Work IQ Sphere Slide.pptx" Internal slide, only on request: the same fill_slide.py command with --internal-names --out "working/Work IQ Sphere Slide (internal).pptx". Check: each prints one JSON line with "status": "ok"; both spheres’ views hold week, month and quarter (less any view left out), each with signals equal to that view’s step 8 total_signals; the second has "customer_safe": true; the slide has "html_links" above 0, "target": "Work%20IQ%20Sphere%20(customer-safe).html" and numbers.total equal to the three-month total_signals (files included when counted; then numbers.files equals files_worked_on). Then read the slide’s fitted and warnings: if the title (WIQ title) or subtitle was cut or shortened, set text.slideTitle (or text.slideSubtitle) to a shorter one in your own words with the same tokens; if the output titles (WIQ outputs) were shortened, shorten those outputs titles to a few words each; if a memory callout was cut, set text.slideMemory to a shorter callout with the same numbers; if warnings names a missing field, add it; then rebuild all three.
  11. Own-data check. Copy the Python block under “Own-data check (step 11)” in references/workflow_details.md exactly into working/check_own.py (never change it), then run python -B working/check_own.py "<skill>" working <UPN>. It checks that everything is stamped for the signed-in user, that nothing from the fictional sample or an earlier version’s sample appears unless their own data has it, that every customer, person and escalation has an alias that differs from its real name, and that the customer-safe files hold no real name, address, number, file or site name. Check: it prints OWN-DATA CHECK: OK. On FAIL, rewrite or remove each listed item from the user’s own data and repeat steps 10 and 11. Post Update 6.
  12. Deliver and save. Publish Work IQ Sphere.html, Work IQ Sphere (customer-safe).html and Work IQ Sphere Slide.pptx (and the internal slide if asked) side by side in one output folder with the usual file delivery. On a refresh, replace the earlier copies; if the delivery refuses to replace a file, delete the earlier copy of that one file and publish again. Save working/workiq_config.json to the user’s own OneDrive at Documents/Cowork/work-iq-sphere/workiq_config.json (create the folder if needed, replace the earlier copy, and note whether an earlier copy was there). Never publish or save discovery.json, signals.json, anything in working/views/, keywords.json, folders.jsonl, chats.json, chats_partial.json, files.json, files_report.json or check_own.py: they hold raw subjects, names and file names.
  13. Reply in the D4 shape: read “Final reply (D4)” in references/workflow_details.md first.

Refresh (“refresh” or “update my Work IQ”)

Before you start, read “Refresh (steps 1 to 6)” in references/workflow_details.md and follow it. In short:

  1. Read the saved setup under rule A8 and copy it to working/previous_config.json; if it is missing or unusable, do a New build.
  2. Never show a question card and never reuse the saved window or focus: all three views, all their work, unless the refresh request itself names one customer, programme or topic (then focus.py in step 6). Send the opening message (the refresh form, D1), then run steps 1 to 6: every source again, files included.
  3. Keep a saved entity only if the new three-month discovery still gives it evidence, with new from ids and its safe alias. Keep a role alias such as “My manager” only when the manager lookup in B1 confirms it again. Keep a stated: item only after asking the user again. Re-stamp meta (A6: version 2.0.0, "window": "quarter", the focus only if named).
  4. Run steps 8 and 9 fresh (never reuse old numbers, callouts, questions or outputs; write every kept detail again with placeholders).
  5. Build the first sphere with the saved setup as --previous: python -B "<skill>/scripts/build_workiq.py" --config working/workiq_config.json --discovery working/discovery.json --view week working/views/discovery_week.json working/views/signals_week.json --view month working/views/discovery_month.json working/views/signals_month.json --out "working/Work IQ Sphere.html" --previous working/previous_config.json, then the rest of steps 10 to 13.
  6. Add “Changed since ” to the reply, from its CHANGES: line.

Small changes after a build

“Rename …”, “remove …”, “add the agent I built”, “change the headline” or “make the internal slide” in the same session: edit working/workiq_config.json, then steps 10 to 13 (one to two minutes, no data re-read), and add one line at the top of the D4 reply: “Changed: .” Something the data does not show goes in as "stated:<their words, naming it>" (A4). “Which alias is which?”: first say “This shows real names, so don’t share your screen”, then list alias and real name. Changing the focus in the same session by naming it (“focus on instead”, “all my work again”) keeps the data already read: run focus.py for the named focus (step 6), then steps 7 to 13 with the new meta.choices.focus and focus flags (step 8 counts every view again with them). “Make it the last month” or “show me this week” needs no rebuild: every sphere has Week, Month and 3 months; tell them W switches the view, and adding #week or #month to the end of the file link opens on it. In a new session working/ is gone, so do a Refresh.

F. Errors in plain English

Whenever a script or tool prints ERROR:, WARNING: or NOTE:, a Python Traceback or OWN-DATA CHECK: FAIL, or a tool call fails or is refused, read references/error_table.md and find the row whose first column matches what was printed. Say its middle column to the person, then do its last column. Never show them the raw error and never ask them to edit code or files. Never patch meta, and never hand-edit script output, the scripts or the template to get past an ERROR (A7). If a file the skill needs is missing (the completeness check fails, or No such file or directory / can't open file for a path inside <skill>), give the partly-installed message from “Before anything else” and stop: read no more data, deliver nothing and never recreate the missing files.

G. Presenting (keys match the renderer)

When the person asks how to present, or about a key, a view, search, kiosk mode or the slide’s links, read “Presenting (G)” in references/workflow_details.md and answer from it.

H. Sharing safely

  • Customers get only Work IQ Sphere (customer-safe).html and the customer-safe Work IQ Sphere Slide.pptx. Never send Work IQ Sphere.html, the internal slide, workiq_config.json or anything from working/.
  • The customer-safe files still show the presenter’s own name as owner; set owner to a role and rebuild to hide it.
  • Microsoft’s official Work IQ slide is built into the HTML and the slide. The presenter checks it is cleared for the audience; --official-slide "" leaves it out of the HTML, and in PowerPoint delete hidden slide 2 and the “Microsoft’s Work IQ slide” button.
  • The slide keeps the sensitivity label that the skill’s slide template carries (Office shows it); the presenter changes it in PowerPoint if their organisation needs another.
  • Many mail systems block .html attachments: share a OneDrive link to the customer-safe file and ask recipients to download it and open it in Edge or Chrome.
  • To pass the skill on, share it from Copilot Cowork (Customize > Skills > Share) or share the original skill zip, never a skill folder that has been used or anyone’s outputs.

Guardrails

  • Never build, refresh or show a sphere from anyone’s data but the signed-in user’s own (A1, A2).
  • Never invent evidence, numbers or dates, and never hand-edit script output, the scripts or the template to get past an ERROR (A4, A5, A7).
  • Never send, post or share anything for the person; customers get only the customer-safe files (H).
  • Never delete files or data except the earlier copy of a delivered file that a refresh replaces (E12); “remove …” requests only take items off the sphere.
  • Translate only subject words, to name a node, and keep the original word in “aka” (A4); never translate documents or other content.

When not to use

  • A generic deck or slide with no personal Work IQ data: use the pptx skill.
  • Work IQ features, licensing or roadmap without the user’s own data: answer directly.
  • Anyone else’s data, team or org-wide analytics, or ranking colleagues: decline (this skill only shows the requesting user’s own work). Editing someone else’s sphere or slide: offer to build the user’s own instead.
  • Scheduled refreshes: use the automation skill, which can run “Refresh my Work IQ” on a schedule.

Files in this skill

  • scripts/: window.py (the windows of step 1), discover.py (step 6), focus.py (a named focus), count_signals.py (step 8), build_workiq.py (the fresh-analysis guard and the sphere, --customer-safe, --previous), fill_slide.py (the slide, same guard, keeps the template’s sensitivity label; also --add-sphere-link and --add-official-slide for another deck, see its --help), wiq_data.py (shared reading, ownership and sample checks, windows, fingerprints), collect_files_recipe.py (the read-only tool script of step 5) and guard_data.json (one-way fingerprints only, of the 1.4.1 sample and of the fictional sample, so the scripts refuse them; nothing from anyone’s current work).
  • templates/workiq_sphere_template.html: the renderer (never edit it during a build); templates/official_workiq_slide.jpg: Microsoft’s official Work IQ slide (--official-slide "" leaves it out).
  • references/Work IQ Sphere Slide Reference.pptx: the slide template (also the demo slide); references/slide_style.md: the slide’s design; references/workflow_details.md and references/error_table.md: the detail SKILL.md points to.
  • examples/example_config.json: the fictional Sam Taylor config (structure reference); examples/Work IQ Sphere (customer-safe).html: the demo sphere.

Rate & discuss this skill

Give this skill a 👍 with your GitHub account. Ratings power the gallery's "Top rated" sort (refreshed on each deploy).