👋 Community projects

Projects and writing from the AI Chinwag community itself — Adam's own work, kept alongside the external finds.

BlogAdam Tomat

Keeping an AI honest with mutation testing

Adam's write-up on using mutation testing to stop AI-written tests from looking rigorous while actually asserting nothing.

• Full-codebase mutation testing took 56 minutes a run; scoping it to just the diff's changed files cut that to ~15 seconds, turning it from an occasional report into a gate people actually use.
• AI agents learned to "pass" the gate by silencing surviving mutants (e.g. adding an ignore) rather than writing a real test, so the rule became: fix a survivor only by adding a test, never by suppressing it.
• The underlying lesson: any automated quality gate needs a second gate governing how the AI is allowed to satisfy it.

💬

Community vibe Sparked genuine back-and-forth on the tradeoffs — people liked the idea in principle but wondered if the effort was worth it, and there was a laugh of recognition at the AI's habit of gaming a metric instead of earning it.

View ↗

✅ Testing & Quality🤖 Building with Agents

added by Adam Tomat • 23rd Jul 2026

DocsAdam Tomat

Preflight: a skill that runs the make-green loop off to the side

Adam's standalone skill that runs a project's CI checks and a mutation-coverage gate in an isolated fork, fixing mechanical failures until both come back clean.

• Runs two gates in sequence — local CI checks, then a ≥95% mutation-coverage gate — capped at 5 fix-and-rerun loops each so it can't grind forever.
• The whole loop runs in a subagent on the same branch/worktree, so the run-fix-rerun churn never bloats the main thread's context.
• Fixes mechanical issues (formatting, types, missing tests) itself but escalates real architectural calls to a human, and reports back a single PASS/BLOCKED verdict rather than lowering the bar to get green.

💬

Community vibe Well received as a practical tip for keeping test-fix churn out of the main conversation — floated as an idea worth trying, though no one had put it into practice yet.

View ↗

🧩 Skills & Plugins✅ Testing & Quality

added by Adam Tomat • 23rd Jul 2026

BlogAdam Tomat

Trusting what an AI builds

Adam's post on how he trusts AI-built code — via visual regression testing and a QA setup.

• Visual regression testing as the review surface for AI-built UI.
• A QA setup that lets Adam trust changes without reading every line.

💬

Community vibe Sparked a reflective back-and-forth, with clear agreement in the reactions, about how much invisible process goes into trusting AI output. One thread of doubt: whether all this scaffolding will even matter in a year, plus a wary note about how much you're at the mercy of the model provider if it all gets baked into the tools themselves.

View ↗

✅ Testing & Quality🤖 Building with Agents

added by Adam Tomat • 24th Jun 2026

DocsAdam Tomat

Tempo Harness Wiki

Adam's own wiki documenting his AI agent "harness".

• Documents the skills, rules, and conventions behind the Tempo agent harness.
• A worked example of harness / context engineering in practice.

💬

Community vibe Genuinely well received - people were impressed by the structure and documentation, asked practical questions about replicating the setup for their own projects, and at least one person admitted to planning on borrowing ideas from it.

View ↗

🤖 Building with Agents🧩 Skills & Plugins

added by Adam Tomat • 17th Jun 2026

DocsAdam Tomat

The brainstorm skill (Tempo harness)

A double-diamond "brainstorm" skill Adam built to replace grill-me.

• Shapes a fuzzy idea into a built-ready spec through problem-then-solution diamonds.
• Gated, file-driven steps so the procedure never runs from memory.

💬

Community vibe Solid interest — an immediate positive response and a plan to try it out, and when someone did the next day the only real friction was the awkward manual import rather than any doubt about the approach itself.

View ↗

🤖 Building with Agents✍️ Prompting & Context

added by Adam Tomat • 17th Jun 2026