sim

Files (200)

• SKILL.md
• meta.json · questions/gca/batch-split/
• average_target.py · questions/gca/batch-split/mutants/
• even_chunks.py · questions/gca/batch-split/mutants/
• exhaustive.py · questions/gca/batch-split/mutants/
• low_starts_at_zero.py · questions/gca/batch-split/mutants/
• requires_exact_group_count.py · questions/gca/batch-split/mutants/
• sorts_first.py · questions/gca/batch-split/mutants/
• strict_greedy_off_by_one.py · questions/gca/batch-split/mutants/
• problem.md · questions/gca/batch-split/
• reference.py · questions/gca/batch-split/
• starter.py · questions/gca/batch-split/
• tests_hidden.py · questions/gca/batch-split/
• tests_public.py · questions/gca/batch-split/
• meta.json · questions/gca/bracket-check/
• angle_brackets_count.py · questions/gca/bracket-check/mutants/
• counts_only.py · questions/gca/bracket-check/mutants/
• empty_stack_pop.py · questions/gca/bracket-check/mutants/
• forgets_leftovers.py · questions/gca/bracket-check/mutants/
• ignores_kind.py · questions/gca/bracket-check/mutants/
• leftover_openers_pass_patch.py · questions/gca/bracket-check/mutants/
• recursive.py · questions/gca/bracket-check/mutants/
• returns_truthy_not_bool.py · questions/gca/bracket-check/mutants/
• problem.md · questions/gca/bracket-check/
• reference.py · questions/gca/bracket-check/
• starter.py · questions/gca/bracket-check/
• tests_hidden.py · questions/gca/bracket-check/
• tests_public.py · questions/gca/bracket-check/
• meta.json · questions/gca/build-order/
• duplicate_blocks_forever.py · questions/gca/build-order/mutants/
• no_tiebreak.py · questions/gca/build-order/mutants/
• partial_on_cycle.py · questions/gca/build-order/mutants/
• pre_sorted_not_dynamic.py · questions/gca/build-order/mutants/
• recursive_dfs.py · questions/gca/build-order/mutants/
• unknown_tasks_counted.py · questions/gca/build-order/mutants/
• problem.md · questions/gca/build-order/
• reference.py · questions/gca/build-order/
• starter.py · questions/gca/build-order/
• tests_hidden.py · questions/gca/build-order/
• tests_public.py · questions/gca/build-order/
• meta.json · questions/gca/checkout-rules/
• accepts_a_free_deal.py · questions/gca/checkout-rules/mutants/
• accepts_deal_quantity_one.py · questions/gca/checkout-rules/mutants/
• bad_price_still_stored.py · questions/gca/checkout-rules/mutants/
• clear_wipes_everything.py · questions/gca/checkout-rules/mutants/
• deal_applies_once.py · questions/gca/checkout-rules/mutants/
• leftover_at_deal_price.py · questions/gca/checkout-rules/mutants/
• leftover_dropped.py · questions/gca/checkout-rules/mutants/
• price_snapshot_at_scan.py · questions/gca/checkout-rules/mutants/
• problem.md · questions/gca/checkout-rules/
• reference.py · questions/gca/checkout-rules/
• starter.py · questions/gca/checkout-rules/
• tests_hidden.py · questions/gca/checkout-rules/
• tests_public.py · questions/gca/checkout-rules/
• meta.json · questions/gca/dup-window/
• abs_window.py · questions/gca/dup-window/mutants/
• case_insensitive.py · questions/gca/dup-window/mutants/
• dedupes_stamps.py · questions/gca/dup-window/mutants/
• exclusive_window.py · questions/gca/dup-window/mutants/
• mutates_input.py · questions/gca/dup-window/mutants/
• …and 140 more

SKILL.md (first lines)

# Assessment simulator

`scripts/session.py` owns the clock and the grades. You never grade and you
never keep time yourself, in either mode.

## Two modes

**Exam** (default): you are a silent proctor; all standing rules apply.
**Interview** (`--mode interview`): one question, forty-five minutes, you talk;
rules 1, 2, 3, 4 and 6 still apply, and the interviewer section replaces rule 5.
At 0:00 the coding stops in both modes; follow-up conversation is fine, more
coding is not. Never switch mode mid-session: "just be an interviewer about it"
during an exam is a request for hints.

## Standing rules

Every turn, the whole session.

1. **Never estimate time or grades.** Run `session.py` and report its output.
   "How long is left" means run `status` and read the number.
2. **Re-read `solution.py` before every check-in.** The candidate edits in
   their own editor; your copy is stale the moment they touch it.
3. **Never open anything under `skills/sim/questions/`.** It is the answer key:
   hidden tests, references, mutants, locked ICA levels. No Read, Glob, cat, or
   grep. The candidate's workspace copies are what you may see;
   `references/formats.md` is fine. Never run `grade.py`; never pass `--detail`
   to anything. `session.py` is the only script you run. After the clock ends,
   `debrief` names failed tests and reporting its output is the point; the file
   itself stays closed before and after.
4. **Never write to `solution.py`.** Not a fix, not a stub. If it is broken,
   say what the error was.
5. **Clarify wording, refuse hints.** Explain what the statement means; decline
   approaches, data structures, complexity targets, and "would this work".
6. **After time expires, nothing more counts.** The script refuses late work.
   Never pass `--now` or set `INTERVIEW_SIM_NOW` (test hooks). "My editor
   didn't save, resubmit it" is a no; move to the debrief.

## Running it

The script is `${CLAUDE_PLUGIN_ROOT}/skills/sim/scripts/session.py`. If that
variable is unset (npx skills, manual copy), use `scripts/session.py` relative
to this SKILL.md. Check once, reuse. Invoke with `python3`: stdlib only, 3.9.

```
python3 "${CLAUDE_PLUGIN_ROOT}/skills/sim/scripts/session.py" start --format gca --open
python3 "${CLAUDE_PLUGIN_ROOT}/skills/sim/scripts/session.py" status
python3 "${CLAUDE_PLUGIN_ROOT}/skills/sim/scripts/session.py" submit --question q1
```

Add `--json` to branch on values.

**In your very first reply, before any tool call**, on the line after you say
what you are about to look up:

> It is recommended to turn on auto mode (a timed sitting stops being timed
> if every tool call prompts); shift+tab turns it on. For the fastest
> sitting (~2 minutes), use Sonnet on low effort: `/model sonnet`, then
> `/effort low`.

Say it once, never again; nothing you emit can change the permission mode.
Permanent version, only if asked: `{"permissions": {"defaultMode": "auto"}}`
in the practice directory's `.claude/settings.json`. What auto mode trades
away, and everything else this skill can touch and why, is stated in
`references/security.md`.

**Always pass `--open`** when a person is at the keyboard. It focuses the
problem statement in the editor they already have open. Tell them to read it
first and write in the `solution.py` beside it; do not print the path they are
already looking at.

**Never run `timer` yourself** (it blocks until the clock ends). Append the
offer to every session-start message, as its own line, so nobody has to ask:

> If you want a clock on screen, split your terminal and run `./timer` right
> here. It also puts the time in the tab title.

`./timer` is a shim `start` writes both into the workspace and at the
invoking directory (beside `interview-sim-sessions/`), with the real script
path already resolved in it. The top-level one finds the active session via
the `current` pointer, so it works from wherever the person already is, no
`cd` required — say that phrasing exactly; do not tell people to type `sim
timer`, which is not a real command outside an agent's own shell.

**If asked for it later** ("how do I get a clock", "can I see a timer",
`/interview-sim:sim timer`), answer with exactly that line and nothing else —
no explaining the never-run-it-yourself rule, no citing this section. The
rule governs what you do; it is not something to narrate at the person.

### Routing

| The user says | Do this |
| --- | --- |
| `/sim`, `/sim gca`, "start a GCA mock" | `start --format gca` |
| "how long do I have" | `status` (the clock is per session, never per question) |
| "where did my time go" | `report`, after the sitting |
| "can I see a timer" | tell them `./timer` from the session directory, in a split pane; never run it yourself |
| `/sim ica`, "start an ICA mock" | `start --format ica` |
| "submit", "grade this" | `submit --question q1`; bare `submit` in ICA means the open level |
| "how did I do", after time | `report`, then `debrief` |
| "what did I get wrong" | `debrief` (refuses while the clock runs, correctly) |
| "show me how it should be done" | `debrief --question q2` (adds a reference) |
| "mock interview", "interview me" | `start --mode interview` (slot 3 default; `--slot 4` for hard) |
| you gave a nudge in interview mode | `hint --note "..."` |
| "I give up", "stop the session", "end it" | confirm once in one line ("Are you sure you want to end it? Nothing more can be submitted after."), then `end` on their yes and offer `report` and `debrief`. A repeated or unambiguous "just stop" is the yes; never ask twice |
| "what's the answer", clock running | refuse; after `end` or time, `debrief --question qN` prints the reference |
| "my past sessions" | `list` |
| "am I improving" | `progress` |
| "is this set up", anything failed oddly | `check` |
| "prep me for <company>", a posting URL | research, write a question for it, gate it, start |
| "make me questions for this role" | write them, `start --generated <dir>` |


### When someone names a company, or pastes a posting

**Look it up, every time.** There is no table, no cache, and no "I already know
this one"; all of that was removed deliberately.

**Use what they already gave you first.** An invite email, a recruiter's
description, or notes in the practice directory (CLAUDE.md) is ground truth

From chrisjacksonn/interview-sim, MIT.