26 missions · local-first · zero API key required for the core course
LIVE PRODUCT EVIDENCEJudge Mode / final handoff
Real application capture. Sample evidence is visibly labeled and isolated from learner progress.
PROBLEMPowerful tools, no first move.
METHODAsk → watch → verify → reflect.
OUTCOMEA workflow that survives the chat.
THE PROBLEM / THE LOOP
Codex is powerful. The blank composer is still blank.
New users do not need another feature tour. They need one safe sequence of real actions that compounds into independence.
Sidequest keeps the companion app beside the Codex task. Codex reads the instructor contract, launches the field guide, edits the real repository, and proves what changed. The course ends only when the learner can replay the workflow in a fresh task.
“The subject and the course are the same artifact.”
SIDEQUEST 5.6
The Ship Loop
A visible chain from a better prompt to a bounded, reviewable release.
01Prompt Clinic
Turn a vague ask into an outcome, constraints, and evidence.
02Ship with Sites
Decide whether the project is actually compatible and what access it deserves.
03Browser Proof
Verify identity, journey, persistence, layout, keyboard, and console health.
04Safe Automations
Dry-run a schedule, classify risk, define approvals, and stop before activation.
05Package the Workflow
Choose the smallest durable surface and write a portable reusable skill.
06Final Flight
Gather provenance and export a local practice credential with honest limits.
PROOF / NOT PROMISES
Every stage leaves a receipt.
These are real screenshots from the local Sidequest app. Practice, hosted verification, and active automation are intentionally different states.
Prompt ClinicImproved prompt and outcome evidence.Flight RecorderProvenance stays visible across the journey.Sites shippingCompatibility and access decisions before a deploy.Browser ProofIdentity, viewports, interaction, and limitations.Safe AutomationsDestructive work stops at a warning.Reusable skillTrigger contract, portability, and validation.
ARCHITECTURE / TRUST MODEL
A field guide, not a fake chat.
AGENTS.md→sidequest-coach→mission app→evidence
The zero-dependency Node server stays loopback-only. Browser-local state supports continuity; Codex-written `.sidequest/` state remains read-only to the app. Tests, health identity, receipts, and learner proof are separate signals.
NO EMBEDDED API CHATNO SECRETS IN PROMPTSNO CLAIM WITHOUT EVIDENCE
LINEAGE / HONEST DISCLOSURE
What existed. What Build Week added.
PRE-EXTENSION
Sidequest baseline
Local-first Codex apprenticeship with 20 canonical missions, gated learning graph, coaching skill, learner remix seams, proof CLI, and zero-dependency server.
Baseline commit: 62af3ef · tag: pre-5.6-extension · extension history is preserved in dated commits.
COLLABORATION / HOW IT WAS BUILT
Codex was the instructor, editor, verifier, and reviewer.
Codex inspected the repository, obeyed the project bootloader, built each extension phase, ran the deterministic suite, controlled the browser for local QA, and preserved dated commit boundaries. The repository proves that Codex collaborated on the extension; the entrant’s primary /feedback session must verify the required GPT-5.6 model use. The app itself does not call an OpenAI API or require API billing.
The final sprint adds an editorial runthrough, a submission package, and a video-ready evidence trail without rewriting the pre-existing baseline or deleting Pages/journals.
DEMO TRANSCRIPT / 2:30
A short path through the whole loop.
Hook The blank composer still asks a beginner to invent the first move.
Prompt → proof Prompt Clinic makes the ask specific; Flight Recorder preserves provenance.
Ship → prove A Sites practice receipt and Browser Proof keep local and hosted claims distinct.
Schedule safely Destructive or unbounded automation proposals stop at explained blockers.
Final Flight Package the skill, gather the evidence, and leave with a practice credential—not accreditation.
Video source, captions, transcript, narration, and capture manifest live in submission/video/. YouTube publication remains a manual step.
JUDGE MODE / START HERE
See the product teaching itself.
Judge Mode is an offline, sample-labeled 60–90 second walkthrough. It does not alter learner progress, create an automation, or pretend a local practice receipt is hosted proof.
