Website · Hire the team · See them work · Read the evidence · 한국어

Questionable Hires

Weird, but employed. Unfortunately, essential.

Eight developer skills for AI coding agents: trace legacy code, prove a fix, challenge weak tests, or review a release. Each hire has a specific job and stopping condition. Built with GPT-6 Astra in mind, using portable skill files.

Development preview. Whole-team quality and cost improvement remain unproven. 0.2.0 development preview and release status · Choose by the job · Run a demo without model usage

Meet the team

Not sure who to ask? Choose by the job: diagnosis, fix verification, test quality, or review →

Hire	Unfortunate personality	Useful engineering instinct
Necromancer	“Their reasons didn't leave with them.”	Trace legacy behavior through callers and history.
Receipt	“You fixed it? Show me the receipt.”	Verify a fix with a real before/after reproduction.
Landlord	“Who's paying rent on this abstraction?”	Make abstractions earn their maintenance cost.
Mother-in-law	“And if I click it twice?”	Reproduce realistic interaction and timing failures.
Exorcist	“Let's test your belief in the cache.”	Distinguish debugging hypotheses with experiments.
Hostage Negotiator	“Release the button. The architecture stays.”	Deliver focused changes without optional scope creep.
Con Artist	“Your mock is impressed with itself.”	Find tests that let broken behavior pass.
Friday	“Can Monday-you undo this?”	Review version compatibility and recovery paths.

Hire one

With Node.js/npm and Git, choose a skill and supported agent:

npx skills add SoonGwan/questionable-hires

For one project-local Codex skill without prompts:

npx skills add SoonGwan/questionable-hires --agent codex --skill mother-in-law --copy -y

These commands use the independent skills CLI. Installation, updates, removal and the bundled Python installer · Small standalone archive for offline handoffs · Dated public installation verification.

In a new Codex CLI or IDE thread:

$necromancer Can we remove this workaround?
$receipt Verify that this fix actually prevents duplicate submissions.
$friday Review this release's rollout and rollback plan.

Automatic selection is enabled. Installing eight skills does not invoke all eight on every task. Existing project tests come first; helper tools and their limits are optional. No lifecycle hooks, background services, telemetry or model-setting changes.

Try a real example without a model account

The Con Artist example contains a test that checks only a successful acknowledgment. Removing the write, or writing twice, still passes that test. Run the helper on trusted sample code with Python 3.9+ on POSIX:

python3 -B skills/con-artist/scripts/audit.py \
  --source examples/con-artist-batch \
  --spec examples/con-artist-batch/recipe.json

For both faults, the existing test passes; the stronger stored-record assertion passes on the correct code and fails on the faulty copy. The original files stay intact. This is an executable helper demonstration, not a new model benchmark. Inspect the recipe, observations and limitations.

Does it actually work?

Integration07 — 2026-09-28, measured resource 1be35120: total tokens +1.67%, CLI elapsed −9.63%; only 2/8 roles reduce both. Bounded task outcomes were preserved; quality superiority was not established. These exposed development tasks have one session per condition, shared host/cache and unequal checks. All costs · Original review and limitations · Current decision index.

The frozen confirmation below measures one role on its own cases:

Latest frozen confirmation: mother-in-law met 5/5 reviewed interaction-QA targets versus baseline 5/5, with 0 skill false positives. Across 5 new cases committed before execution, the skill used 93.2% normalized total tokens and 71.3% normalized elapsed time. This is one fresh session per arm and task, not a whole-team or repeated-sample claim.

Inspect the frozen cases, raw values, method, and limitations.

Historical integration06 and earlier team comparisons, adverse outcomes and per-tool experiments remain available. Local tests, hosted checks and model measurements establish different things; a favorable case does not prove whole-team savings.

One example

In a Con Artist audit, a test that checked only save(...)["ok"] still passed after the actual write was removed in a disposable copy. Checking the stored record caught the defect. Production code stayed intact; the no-skill baseline also found it. Compare the original runs and try the helper.

Contribute

A hire needs a distinct job, a realistic example and a case where it should leave things alone. Contribution guide · Security reporting · Design rationale · Release readiness · Roadmap · Changelog · MIT license.

Inspired by Ponytail: memorable characters carrying concrete engineering habits. Our team, instructions and fixtures are original.