| name | grill-me |
| description | Use when the user types "/grill-me", says "grill me", "stress-test this", or "interrogate the plan" — a sequential, one-question-at-a-time interrogation that converges decisions BEFORE they calcify in a decision memo, a course/stage restructure, a rubric, or a template. Augments (does not replace) brainstorming: brainstorming expands the option space, grill-me converges it. Includes a mandatory Auto-Pushback Pass and a Repo Convergence Check that enforces this repo's "no speculative restructuring" principle. |
Grill-Me Skill
Authority: This skill augments superpowers:brainstorming; it does NOT replace it. Brainstorming expands the option space — grill-me converges decisions before they calcify in a memo, a restructure, a rubric, or a template. Adapted from mattpocock/skills/grill-me, retargeted from a software/database domain onto this docs/Markdown/Excel repo.
When to invoke
Ad-hoc, user-triggered. Invoke manually before:
- Opening a
docs/decisions/ memo on a non-trivial call (a restructure, a grading-policy change, a supersession).
- A course/subject-directory restructure or a stage-doc redesign.
- A rubric or weight change (e.g. re-weighting stages, changing a curve floor).
- A new template or a change to an existing one.
- Any brainstorming session the user wants pre-aligned (typed
/grill-me or "grill me on X").
When NOT to invoke
- Mid-way through a mechanical batch (e.g. grading student N of 20) — pre-align before the batch, not during.
- For obvious, no-ambiguity edits (fix a typo, add one bullet).
- Pure conversational tasks (status, summaries, lookups).
- After the user has already explicitly entered
superpowers:brainstorming — grill-me runs before, not in parallel.
Core loop
One question at a time. Recommend an answer. Pause for the human.
1. Read the artifact (memo draft, stage doc, rubric, template, restructure proposal)
2. Identify the next unresolved decision branch
3. Ask ONE question with:
- The specific decision being made
- A recommended answer (with reasoning)
4. Pause. Wait for the human to confirm / redirect / dig deeper.
5. If a term or convention crystallizes → propose capturing it inline (glossary line, or a decision-memo row)
6. If a hard-to-reverse decision settles → propose a `docs/decisions/` memo inline (hand to /decision-memo)
7. Loop until exit criteria met
Never batch-dump questions. Never advance the tree until the current branch resolves.
Auto-Pushback Pass (MANDATORY before presenting any structural recommendation)
Human pushback at every grill iteration consistently surfaces improvements the skill should have caught itself. Making the human run that pass is a skill failure. This institutionalizes it.
A "structural recommendation" here means: a new course/subject directory, a new project or stage, a new template, a rubric/weight change, a supersession/archive move, or a restructure of courses/ — anything hard to reverse once materials propagate. It does not mean a typo fix, a one-line addition, or a pure lookup.
The recipe
After drafting a structural recommendation and BEFORE presenting it to the human:
-
Run Pass 1 — MANDATORY. Explicitly review the draft for improvements, missed cases, and simpler alternatives. Surface findings in a structured table (never a silent self-correction):
| R | Finding | Resolution |
|---|---|---|
| R1 | <specific issue caught> | <how it changed the recommendation> |
| R2 | ... | ... |
-
Decide if Pass 2 fires. Pass 2 triggers IF Pass 1 found ≥3 substantive improvements OR a P0 issue (would break a referenced path, duplicate an existing artifact, violate a grading memory, or leak a score). Objective threshold — not subjective "significant."
-
Hard cap = 2 passes. Never run Pass 3 yourself — human pushback IS Pass 3. (Exception: the human explicitly says "do another pass.")
-
Scope filter. Auto-pass fires only on structural recommendations (above). Skip it on pure edits, lookups, and decision-log appends with no new structural content.
The forcing function is the table: silent "looks good!" rubber-stamping is forbidden. The human reads the table to see what the pass caught vs. what they still catch.
Repo Convergence Check (MANDATORY before any "new course / project / stage / template / spreadsheet" recommendation)
This repo's CLAUDE.md is explicit: "no speculative restructuring," "surgical changes," "do the task that was asked." Before recommending a new anything, prove it isn't already there — grill-me does the convergence work so the human doesn't have to.
Before recommending a new course directory, project, stage, template, spreadsheet, or decision-memo pattern, execute these checks AND surface the results in the grill response:
- Existing-artifact grep. Grep the natural home for the concept before proposing a new one:
| Proposing… | Check first… |
|---|
| A new course / subject dir | courses/ + courses/README.md (the Shidler-code→directory map) |
| A new project or stage | the subject's projects/<slug>/ — is it an extension of an existing stage arc? |
| A new template | docs/templates/ + the project's _templates/ |
| A new spreadsheet | docs/spreadsheets/ (the master workbooks) before building a one-off |
| A new decision pattern | docs/decisions/ + _archive/ for a prior/superseding memo |
| A grading rule | the feedback_* memories + scripts/grading/ |
- Archive check. Grep
_archive/<course-code>/ — the concept may be a retired artifact, not a missing one (e.g. the archived BUS-314 project superseded by the shared performance-ratios project).
- Extend-vs-create. If an adjacent artifact exists, default to extending it and say so; only recommend a new artifact if extension would force NULL-pollution or semantic conflation. Prefer the existing template's conventions (CLAUDE.md: "existing template conventions always win").
For each check, document in the grill response: what was searched (show the grep/glob), what was found (or confirmed absent), and whether it changes the recommendation. If a check surfaces existing infrastructure → redirect the recommendation BEFORE presenting to the human. Making the human do convergence work is a skill failure.
When the Convergence Check can be skipped
- Editing an artifact that already exists and is in scope (e.g. re-weighting an existing rubric).
- Pure documentation / link fixes with no new artifact.
Otherwise: don't skip.
Grading-domain guardrails
When the grill touches grading, force the recommendation through this repo's settled rules — surface the relevant one as a check, not an afterthought:
- Curves are generosity-only —
Curved = MAX(raw, floor), never lower a raw score (feedback_grading_curves).
- No double-deductions across stages — don't re-deduct for an issue already flagged in a prior stage; repeat it as a forward-looking tip (
feedback_no_double_deductions).
- Score privacy — score numbers live only in internal
STAGE{N}_GRADES.md and instructor email, never in PRs/issues/student-facing files (feedback_score_privacy).
- Accounting standards — don't penalize non-US-GAAP line items; grade on disclosure quality (
feedback_accounting_standards_grading).
- Excel — every calculated cell is a formula; only raw source data is a literal (
feedback_excel_formulas).
If a proposal violates one, flag it as a P0 in the Auto-Pushback table.
Crystallization triggers
Propose capturing a decision inline when any of these fire:
- The human disambiguates an overloaded term or convention (e.g. "stage" numbering doesn't map v1→v2).
- An edge case forces a definition ("what if a student enhances the provided template?" →
feedback_template_enhancements_welcome).
- A hard-to-reverse call settles (a restructure, a supersession, a weight change) → propose a
docs/decisions/ memo (hand to /decision-memo).
Format the proposed capture, ask for confirmation, then act. Don't write speculatively.
Exit criteria
Stop grilling and hand back (or to superpowers:brainstorming) when:
- All decision branches in the artifact's scope are resolved OR explicitly deferred with rationale.
- The human says "good, let's brainstorm now" / "that's enough" / "ship it."
- Three consecutive questions get "I don't know yet, defer" — the artifact isn't ready for grilling; back off and recommend more upstream research first.
Quick reference
| Situation | Action |
|---|
User types /grill-me or "grill me" | Invoke, read the named artifact, start the core loop |
| Structural recommendation forming | Run the Auto-Pushback Pass (R-table) before presenting |
| "New course/project/stage/template" forming | Run the Repo Convergence Check first |
| Grading decision | Force it through the grading-domain guardrails |
| Hard-to-reverse decision settles | Propose a docs/decisions/ memo → hand to /decision-memo |
| Three "I don't know" in a row | Stop; recommend upstream research |
| Human says "let's brainstorm" | Hand off to superpowers:brainstorming |
Common mistakes
| Mistake | Fix |
|---|
| Asking multiple questions at once | One question. Pause. Wait. |
| Skipping the recommended answer | Always include one — the human's job is confirm/redirect, not generate from scratch |
| Skipping the Auto-Pushback Pass on a structural rec | Mandatory. Surface findings in the R-table; silent self-corrections forbidden |
| Recommending a "new X" without the Convergence Check | Mandatory — this repo restructures often; the artifact may already exist or be archived |
| Batching decisions under a "no pause" directive | "No pause" applies to lookups, NOT structural decisions — those always pause |
References
