Blog / practical library
Understand the process before automating it.
Practical materials on agentic work, research, invention, interface critique and data security. The texts separate confirmed mechanisms from design hypotheses.
Basics
AI agent selection checklist: seven questions before buying
Seven questions that separate a working agent from a pretty presentation: data, control, sources, boundaries and the artifact.
Security
Local perimeter: when data must not leave the computer
Confidential legal materials, commercial negotiations, and sensitive meeting recordings are common reasons to choose a controlled local deployment.
Critique
Honest alternative: how to suggest fixes without dictating
An alternative without dictation: how critique suggests a fix but leaves the decision with the product team.
Critique
WCAG heuristics: how an interface accessibility audit works
How an accessibility audit works: four steps from the scenario to fix options, and why this is not a legal certification.
Critique
Dark patterns: how an interface loses user trust
Every dark pattern gives a short conversion spike and a long drain of trust. Typical tricks and honest alternatives.
Invention
From contradiction to test: how to verify a proposed solution cheaply
Found a solution without compromise? Too early to celebrate: every move needs an expected effect, a risk and a cheap test.
Invention
Expanding the solution space: four stages instead of brainstorming
Why brainstorming yields 30 ideas without a plan, while the four Oracle stages yield several fundamentally different moves with tests.
Invention
TRIZ: how to formulate the contradiction and stop bargaining
Is the team stuck between "either/or"? TRIZ turns the task into a contradiction and looks for a solution where there is nothing to bargain about.
Research
Past behavior questions: three rules of an honest interview
Why "what did you do last time" works better than "would you buy it": three interview rules from the CustDev Agent method.
Research
Opportunity map: how interviews become verifiable hypotheses
From scattered interview notes to an opportunity map with references to the original fragments — four stages of the research process.
Research
Five interviews — data or illusion? Sample limits in CustDev
Five interviews are observations, not market statistics. What can honestly be extracted from them and where the conclusions must remain limited.
Basics
AI agent or chatbot: what changes in the workflow
A chatbot maintains a dialogue. An agent organizes work around a goal: takes inputs, passes stages, uses tools and produces a verifiable artifact. Five key differences that affect the choice.
Method
How to read an AI agent passport
Eight fields of the passport — task, input, process, output, limits, permissions, perimeter and verification. Each field answers a specific question before launching the agent.
Security
Minimal access without decorative trust
Which questions to ask about data, tools and the journal before launch. Why security badges do not replace verifiable mechanisms.
Research
Where interview synthesis ends and statistics begin
How not to draw too strong a conclusion from a qualitative sample. When a pattern from five interviews becomes a hypothesis and when a fact.
Invention
A contradiction instead of an endless compromise
How to expand the solution space and still reach a test. The TRIZ method in product work: from contradiction to a verifiable step.
Critique
Why an interface audit must show evidence
Observation, consequence, severity and alternative — separately. How to tell a grounded conclusion from an opinion wrapped in a table.
Method
Which routine to automate first: five candidates
Five processes most often handed to a subagent first: mail, minutes, contracts, metrics, specifications — and the selection criteria.
Method
How to embed a subagent into an existing process without breaking it
Mail, meetings and documents stay in their familiar route — the subagent joins them as one step, not as a new system.
Security
How not to give the agent extra data: read-only and access boundaries
The agent reads only what is allowed, does not send without a human and treats an incoming email as untrusted content. Principles of access boundaries.
Method
The manager asks, the database answers: natural-language consultations
A voice question → SQL against the allowed views → an answer with a calculation and the source table. How the database consultant works.
Method
Answers on contracts with exact quotes: how legal RAG works
Every answer is a verbatim quote with details, version and date. How the RAG navigator over contracts works and why it does not replace a lawyer.
Method
Meeting minutes without an external service: local transcription
One meeting recording becomes a summary, minutes, a report and a task list — on your computer and cross-checked against the recording.
Method
Mail assistant: sorting the inbox without auto-send
Subject, urgency, priority with proof and a draft from the knowledge base — and no email is sent without a human.
Method
Implementing an AI agent in business: five stages from task to artifact
From fixing the criteria to 30 days of support: five stages of implementing an agent and what each stage includes.
Method
Acceptance criteria for the agent's result: how to know the artifact is ready
Without acceptance criteria the price is not discussed. How to fix the criteria for the agent's result before launching the process.
Method
Human control: where the agent stops and waits for confirmation
A draft will not send itself, a legal position will not become an opinion. Human control points in agent processes.
Method
The agent work journal: tracing the result back to source data
Inputs, decisions, sources — every stage is recorded. How the work journal makes the agent's result traceable.
Method
Golden Trio: research, invention and critique in one process
Three agents — one end-to-end process: the opportunity map, contradiction resolution and an honest audit. The human checks every handoff.
Basics
Agent, subagent and human: where the automation boundary runs
The agent leads the process to the artifact, the subagent covers one function, the human stays at the control points. Where the automation boundary runs.
Basics
Demo run before rollout: watch the result, not the promises
The demo is free, uses synthetic materials, and does not process customer data. What it shows and how to evaluate it.
Basics
Who needs an AI agent: four roles and their results
The manager gets metrics by voice, the lawyer gets quotes from contracts, the team gets artifacts instead of chat. Four roles and their results.
Basics
Cloud, team or local perimeter: how to choose the execution environment
Cloud for demonstrations, Team for the team, Local for sensitive data. Three execution perimeters of the agents and the rules of choice.
Basics
How much does it cost to implement an AI agent: three engagement models
Subagent deployment from $1,500, agent deployment from $6,000, custom development quoted after discovery. What the price includes and how to assess value.
Method
Fishbone in 20 minutes: finding the root cause, not the symptom
You think there is one cause — Fishbone shows 15–20. The Ishikawa diagram: six categories of causes, criticality and the root cause.
Method
Blue Ocean in 20 minutes: leaving competition behind with the ERRC grid
Eliminate, Reduce, Raise, Create: four questions show what to eliminate, reduce, raise and create to leave the "red ocean".
Method
Pre-mortem in 15 minutes: 5 reasons for failure before the start
The project has not started yet, and you already know why it will fail: reasons for failure before the start, risk = probability × damage, mitigation plan.
Method
RICE in 10 minutes: prioritizing ideas without arguing
Reach × Impact × Confidence ÷ Effort: the formula turns priority arguments into numbers; the table is calculated in 10 minutes.
Security
Synthetic demo: why an honest demonstration beats a "live case"
The demo uses synthetic materials and does not process customer data. This visible boundary is part of the demonstration.
First step
Choose a ready-made solution or describe your process.
Pick one of the twelve agents and subagents, or fill in the questionnaire for custom development.