Stop Prompting. Start Building.
How to build a good customized agent — and hand the repetitive work to it. A practical method
anyone here can use.
The shift is already happening
Across every industry, work is quietly moving from chatbots to agents — assistants configured once to run a job.
40% 60%+ 93%
of enterprise apps will run task- of organizations plan to deploy AI of IT leaders intend to roll out
specific AI agents by end of 2026 agents within two years — up autonomous agents within the
— up from under 5% a year ago. from just 17% today. The next two years.
steepest curve Gartner has
measured.
Gartner, 2026 Gartner CIO Survey MuleSoft / Deloitte
The question is no longer whether — it's who builds first. Today: how to be an early mover, by
building a good customized agent for your own work.
The hidden tax we all pay
~60% of the workweek vanishes into “work about work” — coordination, searching, re-typing. (Asana, Anatomy of Work)
The same weekly report —
THE MATH NOBODY DOES reformatted from scratch, every
week
3 × / month
becomes
The same compliance checks —
re-run by hand, every batch
36 × / year The same enquiry summary —
re-typed, every single time
Do it once by hand, and you've quietly signed up to do it
35 more times. Automation wins back 3.6+ hrs/week — and
79% say their work gets better. (ActivTrak)
The real shift: from user to builder
Don't do the task. Build the agent that does it.
LLM USER AGENT BUILDER
the operator the architect
• Fresh prompt every time; re-explains
• Configures the task once, reuses forever
context
• Output designed once → consistent
• Output = today's prompt → inconsistent
• The process is externalised & shareable
• The process lives in your head
• Value scales with adoption — leverage
• Value scales with your hours — labour
Same tool. Opposite economics — you spend effort, or you invest it.
5 new kinds of work AI is creating
AI takes over the execution — value shifts to these five work-types, across every job title. (Boris Cherny, creator of Claude Code)
Test 10 ideas in the time one
Prototyper tries many ideas fast, keeps what works. used to take.
Your agent, hardened and
Builder turns a working idea into a reliable, shared tool. shared.
Sweeper clears the repetitive mess and tidies the process. Hand the mess to an agent.
One agent → team →
Grower takes what works and scales its reach. department.
Maintainer keeps it running, correct and current. You keep its knowledge fresh.
Notice: as AI does the execution, the sweepers & maintainers rise above the pure doers. Every one starts by building your first
agent.
Two ways to build leverage
Custom GPT Claude Project
the customized AGENT you deploy the WORKSPACE you inhabit
Instructions + knowledge + tools packaged into Persistent instructions + a knowledge base +
one assistant others can run. grouped chats.
Same behaviour for everyone, every time — runs You still drive each session — but from a much
the job hands-off. higher baseline.
Best when: you hand a repeatable task to the Best when: ongoing, judgment-heavy work you
whole team. think inside.
Third option: a Claude Artifact — a deterministic tool you generate once (fixed calculations, checks). Our
focus today → the customized agent. It's the one that scales past you, to the team.
What a good customized agent is made of
ROLE & JOB One sharp identity, one job, a clear definition of “done.”
INSTRUCTIONS How it thinks: process, tone, rules — and what it must never do.
Grounded in your real material (SOPs, style guides, past work) — accurate, not
KNOWLEDGE generic.
TOOLS Only what the job needs: web search, data analysis, or an API action.
TESTS 10–15 real cases it must pass before anyone is allowed to trust it.
Skip any one — and you get a confident, consistent, wrong answer.
One size doesn't fit all — match the build to the job
Instruction-led Knowledge-led
Behaviour is the value — drafting, tone, Answers grounded in documents — FAQ /
structured judgment. SOP assistant.
Build: instructions + examples · report writer Build: knowledge + retrieval · SOP Q&A bot
Action-led Hybrid
It does things — fetches data, triggers a
Chains the above — drafts, grounds, and acts.
workflow via API.
Build: actions + guardrails · data-pull bot Build: all three, end-to-end · triage→log
Pick the archetype first. It decides where your effort goes.
Where most agents fail: the brief
Worked example
A brief, not a wish
a Weekly Ops Summary agent
Role — who it is, one job Role: raw notes → a fixed weekly summary
Inputs — what it receives Input: this week's bullet notes
Rules — the steps + hard constraints Rules: 5 sections · ≤3 lines each · no new facts
Output — the exact format Output: fixed headings · WhatsApp-ready
Edge cases — what to do when unsure Edge: missing item → “not provided”
Refusals — what it must never do Never: give advice or invent numbers
A vague brief makes a confident, useless agent. The brief IS the build.
5 simple ways a custom GPT flops
Every one is avoidable — and every fix is one line.
“Be a helpful assistant.” → It's just ChatGPT with a new name. Fix: Give it ONE sharp job.
No documents uploaded. → It invents a policy that doesn't exist. Fix: Ground it in real SOPs.
No output format set. → Every summary looks different — you still
Fix: Lock the exact format.
reformat by hand.
Never tested on real cases. → Perfect in the demo; breaks on the first Fix: Run 10–15 real cases
messy enquiry. first.
A 3-page instruction dump. → It follows half and contradicts itself. Fix: Short, sharp, no conflicts.
Notice the pattern: each flop is a missing piece of the anatomy — role, knowledge, format, tests, focus.
The workflow: the Build-Once Loop
Run it once per task — then move to the next one.
1 TARGET Pick the right task: high frequency × low judgment × rule-able.
2 BRIEF Spec it — role · inputs · rules · output · edge cases · refusals.
3 BUILD Pick the archetype & tool; create the agent once.
4 TEST 10–15 real cases; fix the failures before anyone trusts it.
5 SHARE Standardise & hand to the team. A team tool is an asset.
Then you spot the next task — and the leverage compounds.
Where to start at Alpro
Good first agent = internal · text-in, text-out · grounded in static docs · judgment-light · no live data.
BEFORE AFTER
Enquiry summaries by hand. Every batch re-
A hybrid agent triages, summarises to a fixed
read, re-typed, re-formatted. Hours a week,
structure, and logs it. Minutes, not hours.
inconsistent, knowledge stuck with one
Consistent every time. Anyone can run it.
person.
Report / summary Internal message Content / script pre- Data clean-up &
SOP / policy Q&A
drafter formatter check reformat
Instruction-led Knowledge-led Instruction-led Instruction-led Action / Artifact
Proof point: automating documentation cuts ~42% of the time — roughly an hour a day back (industry benchmark). Figures illustrative.
Start small. Start safe.
Backend first, not the customer Small model, one task
Enhance your own productivity. Never put Start with a small, scoped agent in your own
an unproven agent in front of customers. area. Prove it, then grow.
No live-data links Company agents → Octopus
Feed it only what the task needs. A data- Enterprise-level agents go through the
linked agent is a leak waiting to happen. governed platform — not personal accounts.
Gartner: ~40% of agent projects will be scrapped by 2027 — mostly weak governance, cost & risk. Small, backend &
governed keeps you in the 60%.
YOUR MOVE
Build one agent — in YOUR area.
1 Pick one repetitive task you own, in your own function.
2 Write its brief — role · rules · output.
3 Build a small backend agent — this month.
One person, one agent. All of us — and the grind disappears across the company.
Stop spending your hours on what an agent can do. Save them for what only you can.