Research Note
Bounded AI Assisted Build Framework
This framework supports a local learning project with no production systems, payments, personal data, regulated data, secrets, external users, or irreversible infrastruct
In this article
Bounded AI Assisted Build Framework
Intended use
This framework supports a local learning project with no production systems, payments, personal data, regulated data, secrets, external users, or irreversible infrastructure.
Build record
| Phase | Record | Passing evidence |
|---|---|---|
| Outcome | One user flow and explicit exclusions | Scope fits one short project |
| Acceptance | Observable success and failure checks | Reader can run each check |
| Baseline | Tool versions, clean repository, initial test state | Starting point is reproducible |
| Plan | Files, interfaces, sequence, risks | Plan can be reviewed before edits |
| Slice | One end-to-end behavior | Diff is small enough to explain |
| Verify | Automated and manual evidence | Expected and failure paths pass |
| Learn | Reader explanation and manual change | Reader can work without copying |
| Preserve | Commit and build notes | Coherent state can be restored |
Working method
Begin with the behavior, not a technology wish list. Choose one vertical slice. Ask the agent to inspect the repository and propose a plan before modifying files. Confirm exact paths and project root. Add tests or checks before or with the smallest change.
After each slice, read the entire diff, run the focused checks, run broader project checks, explain the data flow, and make one small manual modification. If the learner cannot explain the change, the project has produced output without transferring understanding.
Stop conditions
Stop when the task introduces authentication, money, personal data, secrets, public deployment, destructive migration, production access, broad dependency changes, or a domain the reviewer cannot validate.
Stop when the agent is confused about paths, repeatedly changes unrelated files, edits tests to hide failure, or cannot produce an understandable plan.
The safe next move is to narrow the slice, restore the known baseline, or involve a qualified reviewer.
Sources
Follow the evidence.
- daltonanderson.net: how i built a go app in 4 hours with cursor aidaltonanderson.net
- owasp.org: www project top 10 for large language model applicationsowasp.org
- go.dev: getting startedgo.dev
- csrc.nist.gov: finalcsrc.nist.gov
- cursor.com: auto reviewcursor.com
- daltonanderson.ghost.io: how i built a go app in 4 hours with cursor aidaltonanderson.ghost.io
- cursor.com: 1 7cursor.com
- Spotify episodeopen.spotify.com
- cursor.com: teamscursor.com
- youtu.be: n4 J1tDwreMyoutu.be
- owasp.org: www project code review guideowasp.org
- cursor.com: privacycursor.com
- cursor.com: securitycursor.com
- slsa.dev: v1.2slsa.dev
- git-scm.com: v2git-scm.com
- cursor.com: agent best practicescursor.com
- cursor.com: data usecursor.com