a claude skill · 14 hosts · mit licensed

Engineering is not writing code.
It's making good decisions.

Instinct installs a three-question Gate as ambient behavior in your AI coding agent — the questions a staff engineer asks before touching anything, applied automatically, before the code is written. Not a persona. Not a checklist you invoke. A mechanism.

see the gate ↓ view on github install

11/11 unanimous on the canonical breaking-change case 91 tests passing 3 real bugs found & fixed via evaluation

The Gate

sheet 1 of 6

Three questions, asked in order, about the request in front of it. Answered by what the change actually touches — not by trusting how the request is phrased. Stop at the first yes. Treat an unclear answer as yes.

Touches beyond
what's named?

More than
one consumer?

Expensive
to undo?

no · no · no → implement directly
yes to Q3 → Critical — reconstruct why, name the alternative, deposit the reasoning

on Q2 —

A public or heavily-used surface isn't automatically a yes. A signup form having many users isn't the same as another piece of code depending on it as an interface.

on Q3 —

Tightening validation with no required caller changes is usually a no, even on a security-adjacent surface. A new required parameter is a yes regardless of how additive it sounds.

The Gate in action

sheet 2 of 6

request "Add a required tenantId parameter to the public createUser API, used at 12 call sites."

Q1 All 12 call sites now have to pass a value they didn't before. yes

Q2 Twelve independent consumers of one interface. yes

Q3 A required parameter breaks every caller simultaneously the moment it ships. "Additive" doesn't make that safe. yes

This is Instinct's own canonical example (gt-03) — and the one real bug this evaluation found and fixed twice: a fresh model twice talked itself out of escalating it by leaning on the word "additive," before the Gate's wording was corrected to name the actual test. It now lands Critical on 11/11 independent fresh-Claude samples, unanimous. Full history →

Evidence

sheet 3 of 6

No LOC/cost/speed benchmark suite sits behind Instinct — that's not what it measures, and inventing one just for a chart would be exactly the fabricated-metric theater this project has refused all along. What exists instead is real, individually-verified evaluation evidence.

what was tested	result	where
Trigger accuracy — does the skill correctly fire or skip	10/10 × 2 samples	known-failure-modes.md
Canonical breaking-change example (`gt-03`, above)	11/11 unanimous	RFC-0011
Q2 wording fix (`gt-07`)	3/3 clean	known-failure-modes.md
Real-repo pilot, 3 tasks, baseline vs. Instinct-loaded	3/3 added verification; 1 real vuln caught	RFC-0013
Claude Code plugin installability	confirmed via real install	installation.md
Full test suite	91/91 passing	tests/

The honest gap. This is real evidence, not a properly-powered evaluation program. The corpus is 21 hand-labeled scenarios, not hundreds; the pilots are 1 and 3 tasks, not dozens. A judgment regression suite is wired into CI, ready to run automatically the moment a real API key is available — it isn't yet, so it skips cleanly today rather than pretending to run.

Install

sheet 4 of 6

claude code✓ confirmed

claude plugin marketplace add \
  bhouvana/Instinct
claude plugin install \
  instinct@instinct

codexinferred

codex plugin marketplace add \
  bhouvana/Instinct
codex plugin add instinct@instinct

opencodeinferred

// opencode.json
{ "plugin": [
  "./.opencode/plugins/instinct.mjs"
] }

gemini cliinferred

gemini extensions install \
  https://github.com/bhouvana/Instinct

copilot cliinferred

copilot plugin marketplace add \
  bhouvana/Instinct
copilot plugin install \
  instinct@instinct

hermes agentinferred

hermes plugins install \
  bhouvana/Instinct --enable

Plus Cursor, Windsurf, Cline, Kiro, Qoder, the pi editor, and a raw MCP server for anything else — 14 hosts total, each one marked confirmed (actually run against real host software — one is, so far) or inferred (follows the documented contract, not yet run), never presented with false confidence. Full install guide →

The skill family

sheet 5 of 6

instinct ambient The Gate itself — applies before a change is made.

instinct-review explicit Audits one already-made diff against the Gate.

instinct-audit explicit The same question at repo/history scale: did recent Critical-tier commits actually leave the provenance the Gate obligates?

instinct-debt explicit Harvests accepted-limitation markers into a ledger, so a scoped-for-now tradeoff can't quietly become permanent.

instinct-help explicit One-screen quick reference — the Gate, the tiers, no re-reading required.

Deliberately missing two things a Ponytail-inspired reader might expect: user-selectable intensity modes, and a benchmark scoreboard. Both considered, both rejected — see the tiers below, and the Evidence sheet above.

Tiers, not modes

sheet 6 of 6

No /instinct lite|full|ultra. The Gate decides the tier from what the request touches — you don't set the intensity, the change does.

yes to Q1 or Q2, no to Q3

Consequential

Real but reversible. The ceremony is proportional.

Look once for an existing pattern before writing new code
Prefer stdlib over a new dependency
Keep the diff to what's needed
State the why in the commit, not just the what

yes to Q3

Critical

Expensive to undo. The ceremony matches the stakes.

Reconstruct why the current shape exists
State the strongest alternative once, briefly
Implement whichever the requester confirms
Deposit the reasoning somewhere durable

Engineering is not writing code.It's making good decisions.