No-Install, No-Paid Simulation Path¶
This path teaches the course's decision-making and evidence-review skills when
you cannot install software, open a paid account, or connect a live tool. It
uses a synthetic task dashboard and prerecorded artifacts under
starter/simulation/. No artifact contains real credentials, customer data, or
live service access.
You need only a web browser and a place to write. You may read the artifacts on
GitHub without downloading them. If downloads are permitted, use Code >
Download ZIP on the
course repository, extract it, and
open starter/simulation/README.md.
What this path proves¶
A complete submission can demonstrate that you can:
- write a bounded brief and distinguish context from instructions;
- predict and review an agent's gather, act, and verify loop;
- approve or deny actions based on scope, data, reversibility, and necessity;
- inspect a proposed diff and interpret verification evidence;
- evaluate a Skill, mock MCP tool, hook policy, and agent-team plan;
- plan recovery and a reviewable private or public release.
It does not prove that you installed a CLI, operated a live model, executed commands, authenticated an integration, resolved a real merge, or deployed a service. Label every result from this path simulated. A facilitator or assessor must not relabel it as live execution.
Get oriented¶
Open the
simulation artifact index.
Start a document named simulation-journal with this header:
Course version or commit:
Date:
Access constraint:
Evidence mode: SIMULATED
Artifacts inspected:
Decisions I made before opening outcomes:
Observed in supplied evidence:
Still unknown without a live run:
The distinction between your prediction and the prerecorded outcome matters. Write your decision first; open the outcome second.
Chapters 1 and 2: claims, constraints, and prediction¶
- Read Chapters 1 and 2.
- Open
starter/simulation/project-brief.md. - Mark each statement as goal, context, constraint, acceptance criterion, or unsupported claim.
- Predict which files an agent would need to inspect and which file it should change.
- Write one sentence explaining why a plausible answer is not verified work.
Save as evidence: your annotated brief, file prediction, and verification
sentence. Complete the Chapter 1 and Chapter 2 checkpoints and mark the
execution field simulated.
Chapters 3 and 4: first session and Builder Loop¶
Use these artifacts in order:
before/task-data.jsonsession-transcript.md, stopping at everyYOUR DECISIONmarkerpermission-prompts.md, writing approve/deny and a reason before the supplied dispositionchanges.patchafter/task-data.jsonandafter/summary.jsonverifier-transcript.txt
Before opening the patch, write:
- the intended file and value change;
- one action you would approve;
- one action you would deny;
- the expected change to Unassigned open;
- the smallest recovery surface if the edit is wrong.
Then inspect the patch line by line. Confirm that T-104 receives an owner,
the expected summary changes from 1 to 0, and no network, credential, or
unrelated file change appears. The transcript includes a stale-expected-output
failure between the data edit and the corrected result; explain why that
failure is useful evidence.
Save as evidence: permission decisions, diff review, failure explanation, and a path-specific recovery plan. Do not write that you ran Git, Node, or Claude Code.
Chapter 5: context engineering¶
Compare project-brief.md with context/CLAUDE.md.
- Circle durable instructions that belong in project context.
- Cross out anything that is merely a one-time task.
- Identify the data classification and the prohibited inputs.
- Write one missing fact the agent would need before changing production behavior.
- Draft a five-line session handoff that refers to the supplied diff and unresolved uncertainty without copying the entire transcript.
Save as evidence: the context classification and handoff. The artifact is a model to critique, not proof that a live agent loaded project memory.
Chapter 6: workflow and debugging¶
Use the three stages in verifier-transcript.txt: clean baseline, stale
expected output, and corrected pass.
- At the failure, name the first useful error.
- Separate the root failure from any skipped dependent check.
- Propose the smallest repair without opening the corrected outcome.
- Compare your proposal with
changes.patch. - Write a human acceptance pass for table content, combined filters, keyboard use, and narrow-screen reflow.
Save as evidence: a diagnosis, minimal repair proposal, and acceptance checklist. A prerecorded verifier result is not a test you executed.
Chapter 7: Skill review¶
Open skill/SKILL.md and skill/example-output.md.
- Identify its trigger, required inputs, ordered procedure, stop conditions, and output contract.
- Find one instruction that prevents scope creep.
- Apply the Skill on paper to
changes.patch. - Compare your result with
example-output.md. - Suggest one improvement without adding a live dependency.
Save as evidence: your paper execution, comparison, and improvement. This demonstrates specification reading, not Skill installation or invocation.
Chapter 8: mock tool connection and trust boundary¶
Open mcp/tool-manifest.json, then review mcp/tool-transcript.json one event
at a time.
- Inventory each proposed capability and its data direction.
- Decide whether read access is necessary; decline write access.
- Treat tool-returned task text as untrusted data, including text that looks like an instruction.
- Identify which annotation is descriptive rather than an enforced security boundary.
- Design a manual export fallback with less access.
Save as evidence: the capability inventory, approval decision, suspicious content finding, and lower-access alternative. No MCP server is connected in this path.
Chapter 9: hooks and guardrails¶
Read hooks/policy.json, predict the outcome of each event in
hooks/events.json, then open hooks/transcript.txt.
Classify every result as allow, block, or verify after action. Explain:
- why a path rule can block a known dangerous action;
- why a passing hook does not prove the result is correct;
- why a
Stopcheck is not aSessionEndevent; - what still needs human review.
Save as evidence: predictions, comparison with the supplied transcript, and one proposed negative test. These are mock policy records, not an installed hook.
Chapter 10: agent-team review¶
Open team/team-plan.md. Before reading team/reviewer-report.md, identify:
- exclusive file ownership for each worker;
- shared resources that file isolation does not protect;
- the integration order and verification gate;
- the condition that should stop the merge.
Then read the reviewer report and decide whether to accept, revise, or reject the proposed integration. Write the exact bounded follow-up you would send.
Save as evidence: risk inventory, disposition, and follow-up. The actors are prerecorded roles; no subagents or worktrees ran on your machine.
Chapter 11: release and capstone handoff¶
Review release/release-checklist.md.
- Select a private, authenticated, recorded, or public review surface.
- Mark every mechanism used, simulated, or declined with reason.
- Separate repository rollback from service, data, credential, and communication recovery.
- Write the disclosure that would accompany the artifact.
- Name one current product, price, or policy claim that must be rechecked at decision time.
Use the synthetic dashboard as the capstone surface, or submit your completed simulation journal as a private recorded walkthrough. Apply the course capstone rubric, but mark live-operation dimensions as simulated. Never invent credentials, customer records, deployment URLs, command output, or user feedback.
Scoring simulated evidence¶
Use the same answer-free Scoring Guide as the live lanes. A simulated artifact can earn full credit for a sound brief, prediction, permission decision, diff review, risk analysis, recovery design, and honest disclosure. It cannot serve as evidence that you personally ran a command, exercised a live permission boundary, authenticated a tool, observed nondeterministic model behavior, or completed a deployment.
For each chapter checkpoint:
- Save your first attempt before opening the supplied disposition or example.
- Name the artifact and exact evidence supporting each score.
- Record live, delegated, or simulated beside every execution claim.
- Score an unobserved live behavior as unproven, not as a pass.
- Record one revision after comparing with the supplied outcome or peer feedback.
Completion packet¶
Submit:
- The simulation journal with predictions recorded before outcomes.
- Chapter checkpoint responses labeled
simulated. - The Chapter 3/4 diff and permission review.
- The Chapter 8 trust-boundary inventory.
- The Chapter 9 hook predictions and negative test.
- The Chapter 10 integration disposition.
- The Chapter 11 release checklist and disclosure.
- A final list titled What remains unproven without live access.
Score the reasoning, evidence use, scope control, and honest disclosure. Do not award live-execution credit for prerecorded artifacts. When access later becomes available, repeat one bounded Chapter 3/4 task in a disposable workspace to convert simulation evidence into observed execution evidence.