SDFlow
A specification-driven development workflow framework that inserts a multi-agent review checkpoint between specification writing and code generation. Multiple AI agents in different roles cross-review the specification document, while humans are responsible only for key decisions.
Timeline (last 90 days)
The SDFlow process inserts a multi-agent review gate after the spec is written and before coding begins, letting multiple AIs cross-review from different angles while humans only make the key decisions.
SDFlow's review design references and extends the G-Stick framework of YC president Gary Tan, with core practices including switching to a clean context, parallel multi-angle review, and dispatching additional reviewers on demand.
SDFlow's review roles include CEO review, engineering review, developer experience review, code review, domain review, adversarial review, and an independent second opinion from another model.
The reviews SDFlow always runs are the CEO review, engineering review, code review, another-model review, and two adversarial reviews; the main session judges 4 of 16 situations as high-risk, and when high-risk is hit it dispatches another model for re-review and additional adversarial reviews.
In the real case of notifying on-call personnel in a DingTalk group, the adversarial review found that if the push code placed in a goroutine panics without recover, the entire process would exit; two of the three adversarial reviews reported it, and the engineering review and Go domain review also reported it, so the spec ultimately added a panic recover requirement.
For review findings, the main session first merges duplicates and records where multiple reviewers reported the same issue, then verifies citations (checking file line numbers and original text), then rules one by one on whether the problem would occur during implementation, leaving uncertain ones for humans to judge.
In one SDFlow review run, several review streams reported 50 issues, 28 remained after merging duplicates, and ultimately 18 were adopted to directly improve the spec document, with only one requiring a human decision.
One SDFlow review run takes about 17 minutes and consumes more than 15 million tokens, 90% of which is cache reads.
All Facts (7)
The reviews SDFlow always runs are the CEO review, engineering review, code review, another-model review, and two adversarial reviews; the main session judges 4 of 16 situations as high-risk, and when high-risk is hit it dispatches another model for re-review and additional adversarial reviews.
50%UnverifiedSDFlow's review design references and extends the G-Stick framework of YC president Gary Tan, with core practices including switching to a clean context, parallel multi-angle review, and dispatching additional reviewers on demand.
50%UnverifiedSDFlow's review roles include CEO review, engineering review, developer experience review, code review, domain review, adversarial review, and an independent second opinion from another model.
50%UnverifiedThe SDFlow process inserts a multi-agent review gate after the spec is written and before coding begins, letting multiple AIs cross-review from different angles while humans only make the key decisions.
50%UnverifiedIn one SDFlow review run, several review streams reported 50 issues, 28 remained after merging duplicates, and ultimately 18 were adopted to directly improve the spec document, with only one requiring a human decision.
50%UnverifiedFor review findings, the main session first merges duplicates and records where multiple reviewers reported the same issue, then verifies citations (checking file line numbers and original text), then rules one by one on whether the problem would occur during implementation, leaving uncertain ones for humans to judge.
50%UnverifiedIn the real case of notifying on-call personnel in a DingTalk group, the adversarial review found that if the push code placed in a goroutine panics without recover, the entire process would exit; two of the three adversarial reviews reported it, and the engineering review and Go domain review also reported it, so the spec ultimately added a panic recover requirement.
50%