My engineering method

AI Engineering Workflow

I use AI agents within a development process that connects a clear request to scoped implementation, meaningful verification and an explicit delivery decision.

I keep responsibility for the problem, architecture, review and acceptance. I adjust the depth of planning and checking to the risk, and use specialist roles only when they add useful review or independent ownership.

I share process concepts and synthetic examples here. I keep internal system identities, environment configuration, access procedures and operational records outside the public library.

From request to acceptance

How I move work through the system

I use the same decision points for a small fix and a larger feature. I classify risk from L0 to L3 and choose lite, medium or hard documentation, keeping the depth proportional to the change.

  1. Understand

    I read the request and inspect the relevant behavior before choosing a solution.

  2. Scope

    I identify the task type, risk, owner, constraints and acceptance criteria.

  3. Clarify and plan

    I resolve material unknowns and agree what evidence will demonstrate success.

  4. Implement

    I make a focused change and preserve unrelated work.

  5. Verify

    I exercise the affected behavior, failure paths and persistence boundaries.

  6. Review and deliver

    I review the diff and evidence before any separately authorized release action.

  7. Observe and accept

    I distinguish installation, observed operation and requester acceptance.

I return to the affected step when a check fails. I retain human decision points for scope, risk and release rather than treating every transition as automatic.

Different questions, different evidence

Three paths through the workflow

I first establish which kind of work is needed. Expand a path to see what changes.

Bug Explain the failure before fixing it

I compare expected and observed behavior, trace the affected layers and reproduce the failure when possible.

I use a regression assertion to connect the correction to the original problem. When reproduction is blocked, I keep that uncertainty visible.

Feature Check what already exists

I verify whether the capability already works for the relevant role and conditions before proposing new implementation.

I define the missing behavior and acceptance criteria, then review scope and design before building it.

Documentation Connect the words to real behavior

I check the instructions against the intended and observed application behavior.

I route a broken application to bug work and new behavior to feature work. I verify the revised steps and links for the intended audience.

Synthetic example · no real system or incident

A saved preference disappears after reload

I use this invented settings example to explain the method. It contains no real records, environment details or measured project outcomes.

What I would investigate

I would compare the form value, submitted input, application validation, stored state and reopened view in an isolated test environment.

I would establish which boundary loses the value before changing code.

What I would change

I would correct the demonstrated cause and preserve the previous valid state when a save fails.

I would keep unrelated behavior outside the change and demonstrate the original failure with a regression assertion when practical.

Illustrative verification plan — not an executed test report
ScenarioEvidence I would require
Valid saveI observe the expected state through the interface, application and persistence layers.
Invalid inputI observe a clear error, a usable recovery path and preservation of the previous valid value.
Permission boundaryI verify that the relevant role can perform only the intended action.
Reload and reopenI confirm that the saved value survives a new read rather than only a temporary interface update.
Unavailable dependencyI report the blocked scenario explicitly instead of turning a partial result into a pass.

I would report local verification, review, release, operational observation and acceptance as separate states. This example demonstrates the reasoning; it is not a claim that these tests ran.

Claims I can explain

How I keep evidence and authority explicit

Scope and ownership

I agree the outcome, preserve unrelated changes and separate exploration from delivery work. I use independent owners only for independently writable tasks.

Meaningful verification

I select checks from changed behavior. I distinguish passed, failed, blocked and untested scenarios; a screenshot, status code or signed record alone cannot prove a complete workflow.

Configured safeguards

I use project-specific checks for protected paths, secret exposure, scope and evidence where those checks are configured. I verify their wiring before claiming automatic enforcement.

Human release decisions

I keep tracker updates, commits, remote writes, releases and operational changes within their separate authorizations. I do not equate a merge with a working deployment.

Learning from failures

I use retrospectives and synthetic evaluation cases to target repeated mistakes. I prefer a focused check or clearer existing rule over another permanent process layer.

Selective supporting tools

I use dependency graphs and bounded local-model triage when they reduce uncertainty or context size. I keep sensitive material out of reusable prompts and examples.

Reusable source

Explore the skills behind the process

I keep the library inspectable: search by name or category, select a skill and read or download its source. These generic files contain no private project overlay or access runbook.

Read my source overview

ai-workflow Reusable skills catalog
GitLab
M↓ ai-workflow-skills-plan.md
↓ □

--- ai-workflow-skills-plan.md

Markdown Plan Source

The source below is loaded from the same `.md` file used for download.

Loading Markdown source...
skills/*/SKILL.md 24 reusable skills interactive filter GitLab source linked

Activity

Skill Credits

External and community-derived prompts

roast, grill-me, grill-with-docs, and grilling are shown here as adapted workflow skills. Original upstream credit and license notes should stay attached in the source repository when those notes are available.

Repository-maintained skills

The remaining catalog entries are maintained in ai-workflow. Imported third-party material is not presented as original authorship.