Brief: vibe-coding
Research brief for vibe-coding · how it was made.
Research brief — vibe-coding
| Field | Value |
|---|---|
| Slug | vibe-coding |
Primary query / seo_intent | Review the Few Lines the Model Cannot Own |
| Template (A–E) | E |
| Evidence tier (1–3) | 3 |
| Audience (one line) | Developers who already paste into Cursor or Copilot - shop-talk, a little dry humor; not a paper on LLMs |
| Public byline (human / editorial) | 8020.in Editorial |
| Reviewer (expert or practitioner) | 8020.in Editorial — engineering desk pass |
| Date | 2026-08-20 |
1. Concentration claim (one sentence)
In vibe coding, a small share of moments - the spec, the tests, the review of risky AI output - decides whether the feature ships; more prompting does not.
2. Hard anchors (2–5)
- GitHub, Copilot-enabled files — average 46% of code built with Copilot across languages (61% Java); counts accepted suggestions in enabled files, not all repos. https://github.blog/news-insights/product-news/github-copilot-for-business-is-now-available/
- Stack Overflow Developer Survey 2025 — 84% use or plan to use AI tools; among those with an opinion, more distrust accuracy (commonly cited 46%) than trust it (~29%). https://survey.stackoverflow.co/2025/
- Software reliability adage — a minority of modules / defects create most of the failures (Juran “vital few”; commonly reported in defect logs). Hedge as industry pattern, not a vibe-coding RCT.
2b. Field 80/20 examples
| Approx % claim | Field / context | Source | Where it will sit |
|---|---|---|---|
| ~46% of characters in Copilot-enabled files from the tool | AI coding | GitHub blog | Hook / review bottleneck |
| 46% distrust AI accuracy vs 29% trust | Developer sentiment | SO 2025 | Trust section |
| ~20% of the session (spec + tests + review) → ~80% of ship quality (principle) | Practice | Logic | Intro |
3. Original observation (only-on-8020 seed)
Five bottlenecks with named drills: spec first, tests before vibe, review the risky 20%, one-tool default, revert when lost.
4. Ignored majority (named)
Prompting the same bug five ways, generating extra files, skipping tests because the demo looks right, pasting secrets into the chat.
5. Composite policy
| Scenario | Keep as Illustrative? | Cut instead? |
|---|---|---|
| Ten PRs tagged: seven fails were missing tests or an unreviewed auth path | Yes |
6. Vital few (draft list)
- Write the outcome in one paragraph first
- A failing test or check before the generate loop
- Review auth, money, data, and deletes
- One agent / one chat for the task
- Revert and rewrite when the vibe is mush
7. Device budget reminder
Template E: ≤5 bottlenecks, one drill each. ≤2 example/move pairs.
7b. Viz & misreads
D3 viz? none
Misreads: “If it compiles, ship it.” “More agents means more quality.”
8. Outbound links planned
- GitHub Copilot 46%
- Stack Overflow 2025
- Internal: software-development, learning-programming, chat-gpt
9. Sign-off
- ☑ Brief complete — ready to outline
- ☑ Reviewer has agreed to expert/practitioner pass
- ☑ No fake citations planned