Business judgment on AI products
AI that writes code must design first, test first, and only then claim done, so people stop re-teaching the discipline every chat.
01
Start inside the user's day · Public facts + observable behavior · 2026-08-31
AI that writes code must design first, test first, and only then claim done, so people stop re-teaching the discipline every chat.
Public materials do not yet show how users complete this job today or what they replace.
The product targets friction in this job, but public user evidence does not yet show the cost, frequency, or consequence of leaving it unsolved.
The trend is engineering discipline leaving ad-hoc chat rules and becoming reusable behavior packs. Don't sell a giant prompt. Start with release, incident, and review — steps that must happen in order — and wrap them as actions. Open source; pricing is undisclosed.
Its public repository has 93 stars and 80 forks, showing developer attention; repeat use and payment are not yet verified.
① Whether the fork/star ratio returns to normal (forks well below stars) — 84/85 says copying,; not approval; ② Whether public "ran godmode on a real repo" cases or eval results appear — "production-grade"; needs production evidence; ③ Whether skills get distributed through Claude Code/Codex officia…
Worth dissecting. Its public repository has 93 stars and 80 forks, showing developer attention; repeat use and payment are not yet verified.
when adding an "engineering behavior layer" to any agent product, copy the two design principles — "agent claims are not evidence" (require fresh evidence for done) and "responsibility names over aliases" (literal names prevent drift). For anyone building agent-workflow products, these two principles are free correctness.
Not disclosed. MIT open-source, no pricing page, no hosted service. ① Whether the fork/star ratio returns to normal (forks well below stars) — 84/85 says copying,; not approval; ② Whether public "ran godmode on a real repo" cases or eval results appear — "production-grade"; needs production evidence; ③ Whether skills get distributed through Claude Code/Codex officia…
AI that writes code must design first, test first, and only then claim done, so people stop re-teaching the discipline every chat.
Its public repository has 93 stars and 80 forks, showing developer attention; repeat use and payment are not yet verified.
An English validation note will follow from the public evidence.
The product claims to help users complete: “AI that writes code must design first, test first, and only then claim done, so people stop re-teach”. User evidence has not yet verified pain intensity or the cost of doing without it.
The assessment is recorded; an English explanation is pending.
The assessment is recorded; an English explanation is pending.
The assessment is recorded; an English explanation is pending.
02
Market comparison
Local supply: Emerging
Demand evidence: Not yet verified
Public coverage has been recorded for this market. · 2026-08-31
03
The call and next move come first; the full read retains the evidence and counterevidence.
Engineering behavior for coding agents that already know how to code: a composable catalog of Agent Skills that makes agents design before editing, test before claiming, review independently, and verify with fresh evidence.
thiientv, an individual project. Repository created around 2026-08-12, 12 commits, initial version plus supporting tooling in two days.
Read: packaging software-engineering practice (TDD, review, releases, incident response) as skills isn't a new idea, but the completion here — 14 core workflow skills plus 19 engineering capabilities, a behavior-eval harness, and a repository gate — says the author is polishing it as a product, not posting it as a note.
What it deliberately does not do: no proprietary orchestration runtime — it stays a portable catalog; no claiming compatibility with clients it hasn't actually checked (the README is explicit).
Without it, a coding agent's default is to start writing code immediately: authentication, tests, security and integration get pushed to the end, and it wraps up with a plausible-looking result. Humans catch these by manual review — but as agents multiply and tasks lengthen, review quality collapses.
godmode replaces the old practice of "typing engineering discipline into the conversation" — re-demanding every time that the agent write tests first, design first. It turns discipline from an ad-hoc chat agreement into a reusable skill asset. It doesn't replace the human; it replaces "the human having to re-teach every time."
Not disclosed. MIT open-source, no pricing page, no hosted service.
Read: a skill catalog is hard to monetize directly, but like dsh_workflow it bets that standardized Agent Skills become the public foundation of agent engineering. If that foundation forms, early catalog authors have first-mover advantage.
| Dimension | Call |
|---|---|
| Founder-product fit | A solo systematization of engineering practice; the author is clearly a true believer — but no record of team/real-world validation |
| Product insight | "Agents can code, they can't engineer" is an accurate observation, and the literal skill naming is right |
| Execution quality | Repository gate, behavior evals, compatibility docs — a more complete engineering process than most 12-commit projects |
| Timing | Agent Skills are becoming a standard format; an early high-quality catalog has positioning value; but the format war is undecided |
A serious attempt to package two decades of software-engineering discipline into skills agents can consume, and the naming principle is the part worth stealing.
"Agent claims are not evidence" and "literal responsibility names over borrowed vocabulary" — both map directly to the two real sources of engineering incidents: agents' optimistic false claims, and vocabulary drift across projects. The standard death mode of a skill catalog is "nice-sounding names that never reach concrete actions," and deterministic scripts plus reference files are the plug for that.
The transferable rule: name agent skills and tools by responsibility, not by brand. Literal, task-oriented names (root-cause-debugging, not smart-debug) align an agent's routing decisions with human expectations and avoid being held hostage by one client ecosystem's vocabulary.
The risk is that it hasn't proven who it speaks for. 84 forks out of 85 stars says propagation is by copying, not endorsement. Pre-1.0, no public field record, self-run evals — "production-grade" is currently a self-claim.
① Whether the fork/star ratio returns to normal (forks well below stars) — 84/85 says copying, not approval ② Whether public "ran godmode on a real repo" cases or eval results appear — "production-grade" needs production evidence ③ Whether skills get distributed through Claude Code/Codex official skill marketplaces — decides whether it captures the format dividend
Product logic: when adding an "engineering behavior layer" to any agent product, copy the two design principles — "agent claims are not evidence" (require fresh evidence for done) and "responsibility names over aliases" (literal names prevent drift). For anyone building agent-workflow products, these two principles are free correctness.
Pricing structure: none. Not disclosed.
Ideas present, unvalidated. Disciplined engineering skills are the right direction and the repo quality exceeds what the star count suggests, but the abnormal fork ratio, the absence of field records, and pre-1.0 status all say "production-grade" is still a self-claim. Note it; track field evidence and official marketplace distribution.
04
Evidence trail
05
Use these searches when the official site is missing or the current link is only a lead.