Criteria before conclusions
Define what will count as success before accepting an output.
HOW IT WORKS / VERSION 1.0
The AI Workflow Method turns an objective into stages, checks and a record you can pick up again. The result comes with a way to understand how it was reached.
Drawn from the Research Engine. Portable beyond research.
01 / THE PROCESS
Objective, inputs, constraints and what success means.
Stages, deliverables and the decisions a person must approve.
Do the authorized work. Save sources and real outputs.
Check against criteria. Record defects and correct them.
Share the result, evidence, limits and next action.
A failed check sends work back for correction.
A checkpoint preserves the record; resuming starts by checking it.
02 / WHY A METHOD
A normal chat can use these practices too. The method makes them explicit, repeatable and easy to carry between projects.
Define what will count as success before accepting an output.
“Complete” needs a real artifact and a check. Instructions alone do not complete work.
Keep versions, sources, decisions, corrections and a clear place to resume.
Independent review and approval to publish remain distinct from an AI check.
03 / WHAT EXISTS TODAY
EXISTING RESEARCH ENGINE
It generates research instructions and a downloadable project checkpoint in your browser. You give those to an execution agent. The website itself does not conduct research or run a background job.
NEW IN METHOD v1.0
Schemas, a master template, domain playbooks and an optional local helper record ordered gates, artifact hashes, retries and resumable state. The helper is separately operated; it is not connected to the website.
Review records rely on an honest operator; the helper does not authenticate people or establish scientific correctness. Autonomous execution, live domain integrations and remote synchronization are not included.
04 / A WORKED EXAMPLE
A source-reporting audit checks a few claims in the accessible abstract of Attention Is All You Need. The note records the source version, exact locator and what was checked.
Read the example and limitationsEXAMPLE CHECKPOINT
Source claims were transcribed and checked. Experiments were not reproduced; no completed literature review is claimed.
05 / MAKE IT YOURS
The controller tracks the work. A domain adapter defines the tasks, evidence and human decisions. These adapters are templates, not live service integrations.
Question → evidence → defensible draft. Keep source-linked extraction and independent review.
Approved facts → job fit → application draft. Approve each submission.
Design basis → calculations → checks. Keep units, assumptions and qualified review.
Baseline → focused change → verified preview. Preserve behavior and a rollback path.
Public sources → fit assessment → outreach draft. Review claims and approve sending.
Keep this method's state separate from the Research Engine's existing project JSON. They use different schemas.
Plain files. No account, API key or proprietary runner required for the documentation and local helper.
VERSION HISTORY
November 7, 2026. General method extracted from Research Engine v2.0, with schemas, a reusable template, domain adapters, an honest worked example and a tested optional local helper.
To share: link to this page and include the version. To record a demo: show the objective, handoff, evidence and checkpoint. Review the recording for private data before publishing.
Try the Research Engine