feature-torture

feature-torture

A decision-discipline skill for Claude Code (and compatible AI coding assistants). Pressure-tests one roadmap feature per session and emits a 500–900-word report a human can act on — verdict + ADR + adjacency map + spec stub when shipping.

What this is not

feature-torture is not a stress-test, load-test, fuzzer, or chaos-engineering tool. It does not run code, exercise APIs under load, or break a deployed system to see what breaks. It runs before implementation, on a roadmap entry. The "torture" in the name is the interrogation of an idea — does the slice make sense, what's the better cut, when should it ship — not the abuse of a built artefact.

If you landed here looking for stress / load / chaos testing, you want one of: Gatling, k6, Locust, Chaos Mesh, or Toxiproxy.

Why

A roadmap entry is a headline. A sprint plan is a commitment. The gap between them is where unverified assumptions hide. feature-torture forces the assumptions out before the sprint inherits them.

Most AI feature-review sessions drift: you ask "what should we do with X?", the model lists pros and cons, and you leave with a longer paragraph but no decision. feature-torture treats the AI as a decision facilitator, not a participant — it runs a structured diamond-loop process, picks 2–4 techniques from a closed list, and forces the verdict into one of six labels:

• 👍 ship — proceed as scoped, or with named amendments
• ✂️ reshape — keep the goal, change the slice
• ⏸ park — right idea, wrong cycle
• 🧬 split — entry hides multiple features; spawn children
• 👎 kill — drop the entry; reasons are load-bearing
• 🤷 defer-decision — only when a single probe would flip everything

Each non-default verdict triggers a post-converge cross-label challenge: the report has to argue why the two nearest-neighbour verdicts don't fit. Refutations live in Choice, making the verdict defensible at a glance.

When the input isn't a roadmap feature

When you point the skill at something that isn't a feature row — a release-cut block, a policy / semver question, a coupled bundle of decisions — it doesn't improvise a refusal. It detects the shape and emits a structured pick:

• (a) Reframe as feature — name an F-ID this stands in for, and the skill runs as normal against that row.
• (b) Torture as a single decision — keeps the diamond loop and the 6-label verdict, drops the F-ID / spawned-children / "make me dream" artefacts, writes policy-<slug>.md. Hard cap: one decision per session — bundles are refused.
• (c) Bail to /brainstorm — when the input is divergent ideation rather than pressure-test-to-verdict shape.

Detection is automatic; there is no flag. Standard feature-row runs are unchanged.

Install

Claude Code (plugin channel)

/plugin install feature-torture@bastien-gallay/feature-torture

Claude Desktop or claude.ai web (.skill archive)

Download the latest feature-torture-vX.Y.Z.skill from the Releases page and:

• Desktop: double-click — .skill is OS-registered and opens an Add to library dialog.
• Web: Settings → Capabilities → Skills → upload.

vercel-labs/skills

npx skills install bastien-gallay/feature-torture

Quickstart

1. Bootstrap a config for your project. Copy the closest example from examples/ to .feature-torture.md at your repo root, edit the placeholders.

   cp ~/.claude/skills/feature-torture/examples/lucid-lint.md .feature-torture.md
   $EDITOR .feature-torture.md
   

2. Invoke the skill in Claude Code:

   /feature-torture
   

   The agent reads the config, picks one unstarted feature at random, runs the start-state check, and writes a report to {{REPORTS_DIR}}/<F-id>.md.

3. Read the report. If the verdict is 👍 or ✂️ with a spec stub, hand the spec to the next pair-programming session. If it's ⏸ / 🧬 / 👎 / 🤷, the report is the deliverable.

Configuration

feature-torture reads project config from, in order:

1. .feature-torture.md at project root.
2. .personal/feature-torture/config.md (gitignored variant).
3. Inferred values from AGENTS.md / CLAUDE.md / README.md.

The config supplies six placeholders: PROJECT_NAME, ROADMAP_PATH, PICK_SCRIPT (optional), ARCH_DOCS_PATH, REPORTS_DIR, and UNSTARTED_RULE. The examples/ directory ships five real configs covering five different roadmap shapes: