AI Guides › Workbench
By Nigel Guy · 7 min read
Most people meet a new frontier model by typing a clever paragraph into an empty chat box and judging the answer. That tests the model's writing, not its reach. A model that can run for hours is wasted on a prompt that has to carry all your context by itself, and you end up pasting in three emails and calling it "AI at work".
The rule: connect the sources first, then give a short brief that names the outcome, the evidence to read and the limit on what it may change.
Claude Fable 5 launched on 9 June 2026. Anthropic describes it as built for demanding reasoning and long-horizon work, with days-long asynchronous tasks as the headline use. Access was suspended for all users on 12 June after US export controls were applied, and restored on 1 July after those controls were lifted. That is the "back" in the story.
Since then there has been a newer release. Anthropic's Fable page now leads with Claude Fable 5.1 (announced 1 September 2026), available to Pro, Max, Team and Enterprise users. Everything below applies to either; use whichever your plan shows. Two caveats from Anthropic's own pages: Fable models have safeguards for cybersecurity and biology, and some requests are routed to a less capable model, with a notice. And Fable use requires 30-day data retention by default, which matters if your work is confidential.
For developers, the API price is $10 per million input tokens and $50 per million output tokens (roughly £7.50 and £37 at time of writing; check the exchange rate and Anthropic's pricing page). Subscription prices and usage allowances change often, so check the £ price at checkout and what your plan includes.
Four lines, written before every access task. If you cannot fill one, the task is not ready.
| Line | What you write | Example |
|---|---|---|
| Sources | Exactly which connected folders, mailboxes or calendars it may read | Last six months of my Drive folder "Reviews" and sent mail |
| Outcome | The finished thing, and who reads it | A two-page promotion case for my manager |
| Evidence rule | Every claim must point to a source you can open | Link or name the file or email for each claim |
| Limit | What it must not do | Read only; draft, never send |
If your organisation uses Google Workspace and you see "Access blocked", an admin may have to allow Claude as a trusted application.
| Prompt | What it does | Cost | Best for | Catch |
|---|---|---|---|---|
| Evidence case | Builds a sourced case from months of work | Connectors are included in Claude plans; model usage counts against your plan limits (check the £ price at checkout) | Reviews, promotions, appraisals | Gaps in your records show up as gaps in the case |
| Stakeholder update | Drafts the update from last time's format plus fresh numbers | Same | Monthly investor or board notes | Only as good as your metrics sheet |
| Unanswered-money sweep | Finds quiet threads and drafts follow-ups | Same | Freelancers, small firms | Cannot tell a dispute from a delay without your judgement |
Each is rebuilt to follow the Access Brief. Fill in the square brackets.
Fill in: your role, the goal, the date range, the folder and the reader.
You are a careful career-evidence analyst working for me.
Context: I am a [YOUR_ROLE] seeking [GOAL, e.g. promotion to senior level]. My reader is [READER, e.g. my line manager and HR].
Sources you may read: my connected Google Drive folder [FOLDER_NAME], my sent email, and my calendar, between [START_DATE] and [END_DATE]. Do not read anything outside those limits.
Task:
1. Find work that shows the criteria in [CRITERIA_OR_LEVEL_DESCRIPTION].
2. For each criterion, list the strongest evidence: what happened, my specific part, and the outcome.
3. Mark each item as "documented" (you can name the file, email or event) or "inferred".
4. List gaps where I have little evidence, and suggest what I could still gather.
Output: a two-page case with a heading per criterion, followed by an evidence table (claim, source, date, status).
Rules: never invent numbers, quotes or results. If an outcome is not in the sources, say so. If my criteria or date range are missing, ask me before starting. Do not edit, move or send anything.
Before answering, check that every claim has a named source and that nothing from outside my range slipped in.
Fill in: the company, period, audience, tone and the sources.
You are a chief of staff drafting a periodic update for [COMPANY].
Audience: [INVESTORS_OR_BOARD]. Period: [START_DATE] to [END_DATE]. Tone: plain, specific, no spin.
Sources you may read: my last update (in Drive: [FILE_NAME]), the metrics sheet [SHEET_NAME], my calendar for the period, and email threads labelled [LABEL].
Steps:
1. Read the last update and keep its structure.
2. Pull this period's numbers from the sheet. Show change against last period.
3. Identify wins, misses and asks from the calendar and emails.
4. Draft the update: headline, numbers table, what went well, what did not, what we need.
Constraints: use only figures that appear in the sheet or emails and name where each came from. Flag any number that looks inconsistent between sources. Ask me for anything missing instead of estimating. Save as a draft in Drive or show it here; do not email it.
Self-check: confirm each number matches its source and that the "what did not go well" section is not softer than the evidence.
Fill in: the mailbox scope, the timeframe and what counts as money owed.
You are an accounts-receivable assistant for [MY_BUSINESS_OR_ROLE].
Goal: find money that may be owed to me and is sitting unanswered in my email.
Sources: Gmail, [START_DATE] to today. Look for sent quotes, proposals, invoices and requests for payment where the latest message is mine or the thread has gone quiet for more than [NUMBER] days.
Steps:
1. List each thread: contact, subject, amount (if stated), date of last message, days quiet.
2. Rank by likely value and age.
3. For the top [NUMBER] threads, draft a short, polite follow-up in my voice.
Output: a table first, then the drafts below it.
Rules: create drafts only; never send. Do not guess amounts; write "not stated". If a thread suggests a dispute or a complaint, mark it "needs my judgement" and skip the draft. Ask me if my business details are unclear.
Self-check: confirm each drafted thread really is unanswered and that no draft references a figure the thread does not contain.
Use connectors when the evidence already lives in your tools. Paste text instead when the material is small, sensitive or one-off. If the work will take hours across many files, that is where the long-running models earn their keep; for a quick rewrite, a smaller, cheaper model is enough.