Brief: vibe-coding

Research brief for vibe-coding · how it was made.

Research brief — vibe-coding

FieldValue
Slugvibe-coding
Primary query / seo_intentReview 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
Date2026-08-20

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)

  1. 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/
  2. 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/
  3. 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 % claimField / contextSourceWhere it will sit
~46% of characters in Copilot-enabled files from the toolAI codingGitHub blogHook / review bottleneck
46% distrust AI accuracy vs 29% trustDeveloper sentimentSO 2025Trust section
~20% of the session (spec + tests + review) → ~80% of ship quality (principle)PracticeLogicIntro

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

ScenarioKeep as Illustrative?Cut instead?
Ten PRs tagged: seven fails were missing tests or an unreviewed auth pathYes

6. Vital few (draft list)

  1. Write the outcome in one paragraph first
  2. A failing test or check before the generate loop
  3. Review auth, money, data, and deletes
  4. One agent / one chat for the task
  5. 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.”

  1. GitHub Copilot 46%
  2. Stack Overflow 2025
  3. 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