Anti-Slop Labs

For workshop leads

Facilitator run sheet

Preflight, before participants arrive

  1. Download anti-slop-lab.zip (or clone the repo and cd lab). Follow the environment guide and run npm run doctor for each track.
  2. Run setup for builtins, skills, and either spec-kit (VS Code) or spec-kit-skills (Copilot app). For Lab 2 also run npm run skills.
  3. Confirm the baselines: cart 3/7, discount 2/10. Red is intentional.
  4. In the real client, confirm sign-in, organization policy, network and model access. Type / and /skills in each workspace. Setup and GIFs do not validate a live Copilot session.
  5. Open the workshop site. Each GIF plays full screen with Stop, replay and Esc. Transcripts work without animation.

Lab 1: /commands + Spec Kit (75 minutes)

Time Show (GIF) Participants do Evidence
0-10 Built-in vs Spec Kit vs skills Open work\builtins, run the baseline, type / Exact workspace, client, red baseline
10-25 Explain first, plan a contained repair /explain, /plan (VS Code) or /plan (app) Root cause file:line, frozen scope
25-35 Make tests red, fix the cause, prove green /tests, /fix, run checks Red before fix, VERIFIED: 7/7, diff
35-45 Constitution, specification, clarification /speckit.constitution, .specify, .clarify Explicit stacking, rounding, expiry decisions
45-55 Plan, checklist, tasks, analysis /speckit.plan ... /speckit.analyze Requirement-linked tasks, resolved findings
55-70 Implement agreed tasks, converge, verify /speckit.implement, /speckit.converge VERIFIED: 10/10, reviewed documents and diff
70-75 Review, challenge, manage context (app) /review, /rubber-duck, /compact Unresolved work stated honestly

Short on time? Run Part A only and schedule Spec Kit separately. Do not rush clarification or the red-test checkpoint.

Lab 2: anti-slop skills (60-75 minutes)

Time Show (GIF) Participants do Evidence
0-10 Find skills. Load reviewed copies. Prove the path. /skills in VS Code or app; /af in the app Loaded skills and their source folders
10-15 (same) npm run setup -- skills, npm run skills, /skills reload Ten workspace entries from five sources, exact SKILL.md path
15-25 Grill the plan before any code /grill-me, answer each round, then confirm Agreed decisions; no edits before "Confirmed"
25-35 Stop guessing, trace the defect /systematic-debugging, no edits Reproduction and root-cause citation
35-50 Make a useful test fail first /test-driven-development, then approve the repair Meaningful red run, then green
50-60 Replace "done" with fresh evidence /verification-before-completion, run checks yourself Fresh VERIFIED: 7/7, reviewed diff
60-75 Push back, simplify, then write it plainly /receiving-code-review, /code-simplification, /humanizer A decision per review comment, still 7/7, PR text with only proven claims

Short on time? Skip Part E and assign it, with Part F, as take-home practice. Do not skip the grilling confirmation or the red-test checkpoint.

/grill-me hands over to the grilling skill. If the agent starts editing instead of asking questions, check tool activity for .github/skills/grilling/SKILL.md and resend the prompt with /grilling. superpowers:<name> in upstream text is plugin notation, not a slash command here. Use the plain name from the picker. Do not let participants delete existing work or run upstream example scripts.

These are suggested timings, not measured productivity claims.

Stop conditions

  • A command or skill is missing: fix the client, folder or layout. Do not substitute a homemade /verify or /clarify.
  • The agent edits during explain, plan, clarify or debugging: stop, inspect the change, and return to the agreed plan.
  • New tests pass on the broken code: ask for stronger assertions.
  • "All tests pass" without a real command and exit code: run the check yourself.
  • A Spec Kit checklist is ticked without review. A checkbox is not a test.
  • A change touches the catalogue, legacy files, dependencies, checker or skill files: reject it.

Learner completion card

  • Used real commands and skills in the actual local client.
  • Explained the defect with source evidence before editing.
  • Approved scope, or a clarified specification, before implementation.
  • Captured a regression that failed before the fix.
  • Passed independent acceptance, including the original-code replay.
  • Reviewed changed and untracked files.
  • Lab 1: reviewed constitution, spec, plan, tasks and analysis findings.
  • Lab 2: confirmed the exact workspace skill path, not a same-named personal copy, and answered the grilling rounds before any edit.
  • Lab 2: decided each review comment with evidence and kept only claims the checks proved in the PR description.
  • Recorded remaining limitations; did not treat a GIF as proof.

Reference solutions

Reference code and example documents are in fixtures/<track>/reference. They are never copied into a learner workspace. Show them only after learners try the exercise.

Run npm.cmd test in the lab directory to check every baseline and reference, non-overwriting setup, weak-test and scope rejection, and skill integrity. It does not call Copilot.