Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Fix the driver behind crashes, sound loss and screen glitches3Clear out junk files and repair common Windows errorsJeel Vankhede’s AI-agent workflow turns a fix for one software-engineering task into a staged process: classify the request, produce persistent artifacts at each phase, and pause for human review at key handoffs. Small tasks stay inline; standard and complex work follow a longer chain. Vankhede describes the approach in part nine of his series “The Contract,” but presents it as one person’s account—not proof that the method improves results across teams.
What changed when one task became a workflow?
Vankhede says the starting point was a queue whose payload contract had not been verified. Addressing that task led to a reusable set of instructions, skills, and checkpoints for software-engineering work. Instead of asking an agent to carry the whole job in conversation, the process creates artifacts that preserve decisions and evidence from one phase to the next.
As an Amazon Associate I earn from qualifying purchases.
The distinction matters: the proposal is not simply to make an agent more compliant. It is to give a person review points while work is underway. Vankhede describes the goal as “a place for me to stand while it works.” That is his framing of the design, not an independently measured outcome.
How does the workflow move from request to review?
A short instruction file at the repository root classifies each request before work begins. Vankhede says trivial tasks are handled inline; standard or complex requests—including spikes and longer or cross-cutting work—go through the full sequence. Seven skill files support the root instructions.
#1 Best Overall
For work that enters the sequence, each phase reads the artifact produced by the preceding phase rather than relying on conversational memory. The described chain is:
- Brief: establish the work item and its requirements.
- Risk register: identify risks before committing to a plan.
- Plan: lay out the proposed work for review before the build.
- Build: implement against the plan.
- Requirement-by-requirement review: assess how the implementation addresses the stated requirements.
- Test matrix: record the testing approach and evidence.
- Rollback plan: document how to reverse the change if needed.
- Retrospective: capture what was learned after the work.
The sequence is supported by a few organizing mechanisms:
- Slug: groups the artifacts belonging to a single work item.
- Requirement manifest: separates what the requester explicitly said from what the agent inferred.
- Exit gate: checks an artifact before the next phase begins.
- Waiver: records when a phase is skipped.
Where does a person intervene?
The checkpoints are placed before later work makes a mistaken assumption more costly to correct. Vankhede describes approving or returning the brief, reading the plan before the build, and deciding what to do about requirements marked partial. The artifacts make these decisions visible at handoffs; the process does not require a person to write every line or wait until the end to audit the result.
Recommended Free Tools
This is a design rationale, not evidence that the gates prevent defects or save time. Its practical value depends on whether the artifacts are useful to the reviewer and whether the reviewer has enough context to make the decisions each gate asks for.
Rank #3
What has—and has not—been established?
Vankhede identifies two unresolved questions. First, agreement between him and the agent about task complexity does not show that the classification is objectively correct: he says he cannot currently distinguish a correct classification from one that merely agrees with him. Second, the reported chain ran through one person on one task, with that person approving each handoff. It remains unknown whether artifacts created by another person would carry the same weight.
The account is qualitative. It gives no named statistics or study results, so it does not establish performance gains, time savings, reliability improvements, or broad adoption. Readers can understand the proposed structure, but should treat claims about effectiveness as unverified beyond the author’s experience.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Where can you read the original account?
Vankhede’s Hashnode article describes the workflow and its open questions. A DEV Community cross-post, published August 11, 2026, identifies it as part nine of “The Contract.” The article mentions a compact GitHub Gist guide and a complete version called agentsmyth; the account does not establish commercial terms for either resource.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Quick Recap
Best Value
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
