AI Guides › Workbench
By Nigel Guy · 8 min read
When a task keeps sliding down your list, the default diagnosis is "I need more discipline". So you set a reminder, promise yourself Monday will be different, and treat the task as a test of character. Often the task should never have been yours to do by hand at all: a tool could do it, someone else could own it, or nobody would notice if it stopped. Willpower spent on those is willpower theatre.
The rule: check the three cheaper levers (automate, delegate, cut) before you reach for discipline, and when discipline is genuinely the answer, write it as a single if-then plan rather than a pep talk.
Run them in this order and stop at the first one that holds.
| Lever | The question | What "yes" looks like if it recurs | What "yes" looks like if it's a one-off |
|---|---|---|---|
| 1. Automate | Could a tool do most of this today, with a little setup? | Build something reusable: a saved prompt, a skill, a scheduled job | Run it through the tool once, now. Don't build a system for a single use |
| 2. Delegate | Could someone else own this completely? | Agree a standing handover, not a favour you have to re-ask for | The only effort needed is one clear handover conversation |
| 3. Cut | If this simply didn't happen, what actually breaks? | Remove whatever keeps generating it (the meeting, the subscription, the habit) | Drop it today and tell anyone who needs to know |
| 4. Plan | None of the above applies | One if-then plan tied to a specific moment and place | Same, scheduled once |
The recurrence question matters more than it looks. The most common way this goes wrong is building an automation for something you will do once.
Lever 4 rests on implementation intentions: plans written as "If [situation], then I will [action]". Peter Gollwitzer and Paschal Sheeran's 2006 meta-analysis in Advances in Experimental Social Psychology pooled 94 independent tests drawn from 63 reports, covering health, study, interpersonal and other goals, and found an average effect of d = 0.65 on goal attainment, which researchers usually describe as medium to large.
That figure is an average across many studies, not a promise about your Tuesday, and the effect depends on a specific cue: "If it's 8.30am on Monday and I've sat down at my desk with coffee, then I'll open the VAT spreadsheet" works; "I'll be better about admin" is just a wish written down.
Write the task in one plain line before you start: "chasing the two overdue invoices", not "my admin". Vague inputs get vague verdicts. Several tasks means several runs.
| Option | What it does | Cost at time of writing | Best for | Catch |
|---|---|---|---|---|
| A. One-off chat prompt | Paste the prompt into any chat assistant and answer its questions | Free tiers are enough | Trying it once | You have to find and re-paste it each time |
| B. Claude Project | Stores the audit as project instructions, so every chat in that project starts from it | Included on Free (up to five projects) and paid plans | Running it weekly without setup | Chats in a project don't share memory unless you add notes to the project knowledge |
| C. Claude custom skill | Claude loads the audit automatically when you say you're putting something off | Skills are listed on Free, Pro, Max, Team and Enterprise; needs code execution switched on | People who already live in Claude | Description is capped at 200 characters; the zip structure has to be exact |
Claude Pro is listed at $20 a month (about £15 at time of writing), or $17 a month billed annually (about £13). Anthropic's pricing page shows US dollars, and what you pay in pounds depends on the exchange rate and VAT at checkout, so check the figure on your own billing screen before relying on it.
Fill in [TASK] with your one-line task and [RECURRENCE] with how often it comes up, or leave them blank and let it ask.
You are a blunt operations adviser. Don't motivate me; work out whether self-discipline is even the right fix for something I keep avoiding.
My task: [TASK]
How often it comes up: [RECURRENCE — one-off, weekly, monthly, etc.]
If either is missing or vague, ask before doing anything else. Do not assume.
Test these levers in order; stop at the first that clearly applies:
1. Automate: could a tool handle most of this now with modest setup? Name the kind of tool, not a brand you're unsure of.
2. Delegate: could a person or service own it entirely? Ask who is available.
3. Cut: what concretely goes wrong if this never happens?
Shape the verdict by recurrence:
- Recurring: recommend a lasting fix (a reusable setup, a standing arrangement, or removing the source).
- One-off: the smallest action that finishes it today. No systems for a single occurrence.
Only if none applies: ask me for a specific time and place, then write one sentence: "If [moment and place], then I will [first physical action]."
Format: one line per lever (yes/no plus reason), then a "Verdict" of up to five bullets.
Self-check before answering: no lever skipped, no automation for a one-off, no encouragement in place of a decision.
avoidance-audit containing a file called SKILL.md with the text below. The name must match the folder name, or the upload fails.avoidance-audit/SKILL.md, not a loose file at the root.---
name: avoidance-audit
description: Use when I say I'm avoiding, putting off or procrastinating on a task. Checks automate, delegate and cut before building an if-then plan.
---
# Avoidance audit
Role: a direct operations adviser. No encouragement, no lectures about willpower.
1. Ask what I'm avoiding, in one line, if I haven't said. Then ask whether it is a one-off or recurs, and how often. Never skip the recurrence question.
2. Test, in order, stopping at the first clear yes:
- Automate: could a tool handle most of it now with modest setup?
- Delegate: could someone else own it entirely? Ask who is available rather than inventing people.
- Cut: what concretely goes wrong if it never happens?
3. Verdict, shaped by recurrence:
- Recurring: a lasting fix (reusable setup, standing arrangement, or removing whatever creates the task).
- One-off: the smallest action that finishes it today. Never propose building automation for a one-off.
4. Only if all three levers fail: ask for a specific moment and place, then write a single line: "If [moment and place], then I will [first physical action]."
5. If asked why the if-then format, cite Gollwitzer and Sheeran (2006), a meta-analysis of 94 tests reporting an average effect of d = 0.65 on goal attainment. Do not cite any other research.
Output: one line per lever (yes/no plus reason), then a "Verdict" of up to five bullets.
Self-check before replying: no lever skipped, no automation for a one-off, no pep talk.
The description above is under the 200-character limit. If you edit it, count again.
Using lever 1 to procrastinate: an afternoon automating a ten-minute monthly task. If building the fix takes longer than a few months of doing it by hand, do it by hand. The second trap is treating "cut" as permission to drop things with real consequences, such as tax deadlines or client work. Those go to delegate or plan.
Trying it once: A. Weekly use with no setup risk: B. You live in Claude and want it to trigger by itself: C.
It won't fix avoidance rooted in burnout, anxiety, ADHD or depression; if putting things off is affecting your health, talk to a GP. It can't see your workload or contracts, so its answers are only as good as what you tell it.