Software-Shaped Work Framework from the AI Daily Brief, 29 August 2026, tested against Microsoft 365 Copilot Notebooks. Updated 2 September 2026.

Most work worth handing to AI is not coding. It is the work that is already shaped like software.

A practitioner's framework for spotting that work in your own week, and a straight answer on how much of it the Copilot Notebook you already have can carry.

Where the framework comes from. Nathaniel Whittemore's AI Daily Brief episode "How to start AI coding if you haven't yet" argues that building small software is now a knowledge-worker skill, and offers three lenses for finding the candidates: how a build relates to work that already exists, how durable it needs to be, and six places the candidates hide. Those lenses are one experienced practitioner's observed patterns, not a study. The notebook verdicts are ours, from Microsoft's documentation and first-hand use.

Automate
Same job, same output. You stop making it by hand: renaming files, syncing lists, filling templates, reworking exports.
The tell: whoever receives the output would not notice anything changed, and if the software broke you would go back to doing it manually.
Notebook: the reading, not the loop.It will extract the same facts from many documents or fill a template from sources when asked. It will not watch a folder, run on a schedule or produce a file unprompted.
Upgrade
Same job, new output. A report becomes a live dashboard, a deck becomes a web page, a status email becomes a self-serve page.
The tell: the receiver gets something visibly different and better, and going back to the PDF would feel like a loss.
Notebook: the content, not the surface.It will first-draft the explainer, the comparison and the status narrative from sources. It cannot host a live page, a calculator or a dashboard, and cannot draw a chart.
Invent
New job, new output. Work that was never practical by hand: monitoring hundreds of sources, reading every interview, testing a thousand variants.
The tell: there is nothing proven to copy, so some of what you build will go unused. That is the cost of the category.
Notebook: one of the two archetypes.A pile of minutes and reports nobody re-reads, distilled into registers you can interrogate, is the notebook's natural shape. Anything that runs unattended is not.

The second question is how durable the thing needs to be

Whittemore's four delivery classes are a ladder of audience and consequence. Where a build sits on the ladder decides how much security, documentation and support it needs. A notebook occupies one rung by construction.

Prototype
Disposable. Answers a question, tests an idea, moves a decision one step. Optimised for speed and clear examples, not for security or polish.
Notebook: yes. Load the sources, ask the question, learn in an hour whether the material can answer it.
Personal or team software
Reliable for you or a small team, in an ongoing way. Compromises on interface, permissions and edge cases are allowed because you chose them.
Notebook: this is the rung. Shared with named colleagues, grounded only on what was added, permission-trimmed by the tenant. Still allowed to be disposable when the engagement ends.
Production software
Used by people who are not you, under real load, where failure costs trust or money. Needs support pathways and real access control, but still a known, discrete group.
Notebook: only through an agent. Since February 2026 an agent can be grounded on a notebook and offered to a wider known group. Premium surfaces; the notebook itself stays team software.
Product
Serves an open market. People you will never meet depend on it. Every burden of general-consumption software applies.
Notebook: no. Nothing about the tool is built for strangers, and nothing in this framework asks it to be.

Six places the software-shaped work hides

Whittemore's diagnostic questions, the thing a builder would make in answer, and what the notebook can do with the same problem. The last column is the one to read if you only read one.

Work categories from the episode, with the notebook verdict. Verdicts: filled circle fits, half circle fits in part, empty circle not this tool.
WorkThe tellWhat a builder would makeVerdictIn a Copilot Notebook
Presentation You explain the same thing to different people. You help people compare options. You keep looking up the same rates, rules or contacts. An interactive explainer, a comparison tool, a self-serve calculator, a status page, a lookup tool. In part A shared notebook grounded on the reference material answers the repeated question, with a citation, for anyone with access. That is the lookup tool and the explainer. It is not a page, a calculator or a status surface.
Content You turn long things into short things, transcripts into summaries, one thing into five. A pipeline that does the whole conversion end to end, every time, unprompted. In part Summaries, extractions and derivative drafts from the sources, section by section, on request. The "every day, without asking" part is not available; a person runs the pass.
Data Recurring analysis. The same charts. "How are we doing." Exports from one system reworked by hand for another. A dashboard that also takes questions; a translation pipeline between formats. Weak Spreadsheets upload, but how well the tool grounds on cells is unverified; it cannot draw a chart; a reporting tool's dashboards are not a source type. Use it for the narrative around the numbers, and check every figure against the workbook.
Document The same template filled again and again. The same operation on many files. Reading a lot of documents to find specific things. A template filler, a batch converter, an extractor that reads the pile and returns the facts. Fits The strongest fit. Extraction into numbered registers, template drafting from sources, "which document says this". Batch operations on files themselves (rename, convert, resize) are not this tool.
Inbox You collect things from people: applications, sign-ups, submissions. An intake system, when no vendor tool fits the specific need. Not this tool Nothing in a notebook collects or routes. Buy the form tool, or it is an application.
Admin Named in the episode without elaboration: the chasing, tracking and paperwork around the real work. Small automations that act on other systems. Not this tool Nothing in a notebook acts. It only answers.

Six thought-starter projects, tested against the notebook

The episode closes with two example builds per pattern, offered as thought starters rather than best practice. Each is described as Whittemore framed it, then answered for the notebook.

The Friday export

Automate

A spreadsheet ritual you can do with your eyes closed (rename, filter, join, calculate, restyle) becomes a small page where you drop the raw export, preview the transformation and download the file in the old format.

Not this tool. A notebook cannot take a file in and return a transformed file. This is a two-hour build in a coding tool, and it belongs there.

The invoice pile

Automate

Invoices and receipts arrive as PDFs and photos; a tracker is typed by hand. A watched folder reads each one, normalises the fields, flags low-confidence values, catches duplicates, and produces the same spreadsheet for review.

Half. Given the documents as sources it will read them and normalise the fields into a register you review. It will not watch a folder, and it retrieves passages rather than reading the whole pile, so duplicate-catching needs an enumeration prompt with a count.

The live report

Upgrade

Anyone you owe numbers to at a regular interval gets one page instead of snapshots: refreshed from the source, showing when it last updated, answering the common questions before they are asked.

The content, not the surface. A current-state fact register, refreshed by a person each week from the latest status reports and minutes, is the notebook's version of this. Stakeholders ask the notebook instead of asking you. There is no live page and no automatic refresh.

The what-if slider

Upgrade

A projection or staffing model in a deck always reaches "what if we changed price, volume, timing, headcount". A page exposes the variables and shows what moves.

Not this tool. No calculation surface, no interactive state. A spreadsheet already does most of this; a small page does the rest.

The watcher

Invent

A personal research agent that checks sources which change unpredictably (pricing, job posts, grant pages, regulator guidance), removes the noise, keeps a history and sends updates when it matters.

Not this tool. Nothing in a notebook runs unattended: no scheduler, no actions, no memory between sessions beyond its documents. Whittemore himself notes this job may be absorbed by personal-agent platforms rather than bespoke code.

The pattern reader

Invent

A pile nobody has time to read (tickets, transcripts, reviews, survey answers, incident reports) is cleaned, grouped into recurring themes, compared across segments or periods, with every claim linked back to a source passage.

Fits, at engagement scale, with a ritual. Governance minutes and status reports distilled into a decision log, an actions-and-risks register and a change journal, then interrogated at any point. The ritual is the weekly distillation pass a person runs; the caveat is that it retrieves passages, so prompts force complete numbered lists and a count.

Why the verdicts fall where they do

An assistant holds state. A notebook holds documents.

When you prompt a notebook it does not read the documents. It runs a search over them, retrieves a handful of matching passages, and writes an answer from those passages plus a short window of recent chat. Every practical limit on this page follows from that one mechanism, and so does every strength: it cannot wander beyond the sources it was given, which is exactly the property a client engagement wants.

  1. It never reads a whole document. "Summarise the report" summarises whichever passages the search fetched, confidently and with no signal about what it missed. Prompts that demand a complete numbered list, a count, and "sections you could not read" make the gaps announce themselves.
  2. It forgets the conversation quickly. A few turns back is already fading. Every prompt stands alone and names the sources it needs; nothing is built across a long exchange.
  3. It finds what your words match. Retrieval is a search engine. Name the source and use the document's own vocabulary, not a paraphrase of it.
  4. Important outputs go back in. There is no memory between sessions. When a prompt produces something load-bearing (a requirements register, a decision log) it is saved as a source so later prompts can retrieve it. Outputs become inputs; that is how multi-step work chains through a memoryless tool.
  5. It cannot write long, and it cannot act. Output is capped well short of a full document, so drafting is section by section. It answers; it does not run, watch, send or schedule. The person is the orchestrator.

How a notebook is set up to carry this

The worked example is an engagement notebook for IT financial management: a small, knowledge-heavy domain where a curated set of pages fits comfortably inside the tool's source limits. Nothing below is specific to any client; it is the pattern, and it ports to any engagement with a document pile and a team that needs the same facts more than once.

Standing sources

Documents the tool re-reads every session, because a notebook has no memory of its own.

  • Operating instructions in the notebook's settings field: the role, the answer rules (ground in sources, name the page, grade confidence, say "not in the notebook" rather than invent), and the register. Shared and co-edited by the whole team; one copy governs.
  • An index: one line per page, what it answers, when it was last refreshed.
  • A handoff: decisions made, open questions, next step. Re-saved at each session close, read first at each open.
  • Six curated knowledge pages, two pages each, written for retrieval: headings as questions, a source on every number.

Distilled registers

The pattern reader, made of documents. Each is produced by a prompt from the engagement corpus and saved back in as a source.

  • Decision log, from governance minutes.
  • Actions and risks register.
  • Current-state facts, from status reports.
  • Stakeholder map.
  • Business-case baseline: what was assumed, what was approved, the most-asked question of any engagement.
  • Change journal: what moved between passes, so the trajectory is visible, not just the state.

Rituals

The manual substitutes for the automation the tool lacks. Documented, because nothing in a notebook sweeps itself.

  • Session open: "Read the operating instructions and the handoff; state the open questions before we start."
  • Session close, whenever state changed: redraft the handoff, save it out, replace the old one.
  • Weekly pass: distil new minutes and reports into the registers; prune sources that no longer earn their slot; re-upload any copy whose original changed.
  • One steward owns the instructions field and is the only hand that deletes. Deletion is permanent.

Build, buy, or just ask the notebook

Whittemore's own discipline is to look for an existing tool before building, because a vendor whose whole job is that problem has more capacity to do it well than you do as the sixty-eighth item on a to-do list. The notebook version of that rule: if the job is grounded reading, extraction or a first draft from documents the team already holds, the notebook is the existing tool. If the job has to run without a person, take a file in and return a file, or show a live surface, it is a small build or a purchase, and the notebook is the wrong shape for it.

What never gets delegated

The notebook extracts, structures and first-drafts. People decide.

  • Strategy, positioning and the angle: a notebook holding only the client's documents will reproduce the client's documents. Insight has to be captured as a source before it can be retrieved.
  • Pricing, figures and anything bound for a stakeholder: checked against the workbook by a person, every time.
  • Final compliance or sign-off traceability: the tool's citations point at a document, not a clause. Carry identifiers inline.
  • Judgement about the client.

Facts and limits the verdicts rest on

Enterprise tools change quarterly, so every fact carries a date and a confidence grade. H means documented by Microsoft or observed first-hand; M means a single secondary source or an inference that has not yet been checked on a device.

Capability facts for Microsoft 365 Copilot Notebooks as at the date shown. Re-verify before relying on any line older than a quarter.
FactDetailAs atGradeSource
Available on both licence tiersNotebooks are included with Copilot Chat (the tier bundled with Microsoft 365) as well as the paid Microsoft 365 Copilot seat. Rollout to Chat users completed mid-June 2026.Jul 2026HFirst-hand; S2
Source limits differ by tierPaid seat: up to 300 references ground an answer. Chat tier: 50 references, all grounded.Aug 2026HS3
Source types differ by tierPaid seat adds tenant-hosted sources: Teams content, SharePoint folders, OneDrive files, internal links. Chat tier takes uploads (Word, Excel, PowerPoint, PDF), Copilot Pages, images, audio and external links.Aug 2026HS3
Grounded only on what was addedThe notebook answers from its curated sources alone; no reach into mail, files or the web beyond them. Linked folders contribute the most relevant items per prompt, permission-trimmed.Aug 2026HS2, S3
Supported reference formatsWord, PowerPoint, Excel, PDF, Loop, Copilot Pages, plain text, rich text and OneNote pages. Markdown sources were rolling out July to August 2026.Aug 2026HS3; Markdown rollout M
Instructions fieldA notebook-level custom-instructions field exists in settings, is shared with collaborators, and measured at 8,000 characters. Microsoft's own example packs several hard rules into it.Jul 2026HFirst-hand; S4
No images, no chartsNo image generation or data visualisation inside a notebook. Audio overviews are available.Aug 2026HS2
Reporting dashboards are not a sourcePower BI reports cannot be added; the underlying workbook can be uploaded, and grounding quality on spreadsheet cells is unverified.Aug 2026HS3; spreadsheet quality untested
Output and input ceilingsThe per-response output length and the prompt-box limit are undocumented. Widely repeated figures come from a different consumer product with the same name.Aug 2026MAbsence finding
Deletion is permanentNo recycle bin; neither users nor administrators can recover a deleted notebook.Aug 2026HS5
GovernancePurview labels on source documents are honoured; DLP can block prompts and labelled files; eDiscovery and audit are supported. Information Barriers are not supported, and Microsoft advises organisations that depend on them to disable Pages and Notebooks.Jul 2026HS5, S6
Agents on a notebookAn agent can be grounded on a notebook's references and shared, which is the route from team software toward a wider known audience.Feb 2026MMicrosoft announcements, secondary mirrors

Sources

  1. Nathaniel Whittemore, "How to start AI coding if you haven't yet", The AI Daily Brief, 29 August 2026. Official transcript. aidailybrief.ai/e/2026-08-29. The three build patterns, four delivery classes, six work categories and six example projects are his; treated as practitioner opinion.
  2. Microsoft Support, "How Microsoft 365 Copilot Notebooks works". support.microsoft.com.
  3. Microsoft Support, "Add references to your Microsoft 365 Copilot Notebook". support.microsoft.com. Note: the page states in one place that external links can be added on the Chat tier and elsewhere that they cannot; verify on a device.
  4. Microsoft Support, "Provide custom instructions for your Microsoft 365 Copilot Notebook". support.microsoft.com.
  5. Microsoft Learn, "Copilot Pages and Copilot Notebooks compliance summary". learn.microsoft.com.
  6. Microsoft Learn, "Microsoft Purview data security and compliance protections for Microsoft 365 Copilot". learn.microsoft.com.
  7. First-hand observation on a Microsoft 365 tenant, July to August 2026: Notebooks present on the Chat tier; instructions field present, shared, and measured at 8,000 characters. Facts about the tool only.
Fits Fits in part, usually with a person running the loop Not this tool