grilling
The core interview primitive: relentless rounds over a design tree. Builds a tree of decisions, computes the frontier (questions whose prerequisites are settled), asks the whole frontier in numbered rounds with recommended answers. Facts = agent's job (sub-agents); decisions = user's job. Done when frontier empty. Use directly when you want the interview with no wrapper; or invoked by grill-me, grill-with-docs, triage, wayfinder, improve-codebase-architecture.
Quellinformationen
- Repository
- Natsummerance/agents
- Letzte Quellaktivität
- 28. August 2026 um 00:28
- Erkannte Sprache von SKILL.md
- Englisch
- Sterne
- 3
- Forks
- 0
Installationsoptionen
Standardmäßig ist der Prompt ausgewählt, der zuerst die Quelle prüft. Sie können zu einem direkten Befehl wechseln oder eine lokale Kopie herunterladen.
Quelldateien prüfen
Lesen Sie SKILL.md und alle von SkillsMP angezeigten Begleitdateien, bevor Sie sich für eine Installation entscheiden.
Datei-Explorer
3 DateienSKILL.md wird angezeigt
SKILL.md
Quellanweisungen · Schreibgeschützte Vorschau- name
- grilling
- description
- The core interview primitive: relentless rounds over a design tree. Builds a tree of decisions, computes the frontier (questions whose prerequisites are settled), asks the whole frontier in numbered rounds with recommended answers. Facts = agent's job (sub-agents); decisions = user's job. Done when frontier empty. Use directly when you want the interview with no wrapper; or invoked by grill-me, grill-with-docs, triage, wayfinder, improve-codebase-architecture.
## R — Reference (≤100 words)
Upstream: `skills/productivity/grilling/SKILL.md` (28 lines). **Design tree**: every decision branches into dependent decisions. **Rounds**: work the **frontier** (decisions whose prerequisites are settled). Ask whole frontier in one round: number each question, give body with choices, give recommended answer. Wait for user answers before next round. **Format**: `❓ QN - <title>: <body>` + `➡️ <recommended answer>` separated by `---`. **Facts = agent's job**: dispatch sub-agents for filesystem/tools; never ask user what agent can look up. Don't block: running exploration = unsettled prerequisite; only downstream questions wait. **Decisions = user's job**: put each to user, wait. **Done when frontier empty**: every branch visited, nothing silently assumed. No action until user confirms shared understanding.
## I — Methodology (in own words)
**Core principle**: Structured thinking as a tree traversal. The frontier is the *current cut* of answerable questions. Each round widens the cut.
**Design tree mechanics**:
- Root = initial topic/goal
- Nodes = decisions (not facts)
- Edges = "depends on" (prerequisite)
- Frontier = nodes whose all prerequisites are answered
**Round protocol**:
1. Compute frontier from current tree
2. For each frontier node: formulate question with context, choices, recommended answer
3. Present all frontier questions in one batch (numbered, separated)
4. User answers all (or defers)
5. Update tree: answered nodes become settled; new nodes unlock
6. Repeat until frontier empty
**Fact/decision separation**:
- Agent *must* find facts (code, docs, web, tools) via sub-agents
- User *must* make decisions (trade-offs, preferences, commitments)
- Never invert: asking user for facts wastes their time; agent deciding violates agency
**Sub-agent concurrency**: Fact-finding for multiple frontier questions runs in parallel. Questions whose facts are still loading stay in next round's frontier.
**Completion**: Frontier empty + user confirms "shared understanding reached". No execution until confirmation.
## A1 — Upstream Typical Applications
- Standalone design session: "grill me on this API design" → tree: auth strategy → rate limiting → versioning → error format
- Embedded in triage: triage uses grilling to sharpen incoming issue into agent-ready ticket
- Embedded in wayfinder: wayfinder uses grilling to resolve each decision ticket on the map
- Embedded in improve-codebase-architecture: uses grilling to design the chosen deepening
- Ad-hoc: "run grilling on this one decision" → single-round tree
## A2 — Trigger Scenarios (3 contexts + language signals + sibling distinction)
| Context | User Language Signals (≥3) | vs Sibling Skills |
|---------|---------------------------|-------------------|
| **Raw primitive access** | "run grilling", "use the primitive", "just the interview", "no wrapper" | vs grill-me/grill-with-docs: those are *named entries* with wrappers (stateless/stateful); grilling is the *bare primitive* |
| **Embedded in other skill** | (internal invocation by triage/wayfinder/improve-codebase-architecture) | vs those skills: they *compose* grilling; grilling has no opinion on *what* is being grilled |
| **Single-decision deep dive** | "grill this one decision", "deep dive on X", "stress-test this choice" | vs grill-me: grill-me runs full session; grilling can run on a sub-tree |
## E — Execution Steps (each with "完成标准:")
1. **Establish root topic** — User states the goal/decision space. Agent creates root node of design tree.
完成标准: Root node captured; initial frontier = {root} (or root's immediate children if pre-split).
2. **Compute and present frontier** — List all nodes whose prerequisites are settled. Format each as numbered question with choices + recommended answer.
完成标准: Round formatted per spec (❓ QN - title: body / ➡️ recommended / ---).
3. **Dispatch fact-finding sub-agents** — For any frontier question needing external facts, spawn sub-agents in parallel. Do not wait; mark those questions as "awaiting facts" for next round.
完成标准: Sub-agents launched; non-blocking questions asked now; fact-dependent questions deferred.
4. **Collect user decisions** — User answers each question. Agent updates tree: answered nodes settled, new children added.
完成标准: All presented questions answered; tree updated; new frontier computed.
5. **Iterate rounds** — Repeat steps 2-4 until frontier empty.
完成标准: Frontier empty; no unanswered prerequisite nodes.
6. **Confirm shared understanding** — Present final decision tree summary. Ask user to confirm "shared understanding reached".
完成标准: User explicitly confirms; no action taken before confirmation.
## B — Boundaries & Limits
**不适用**:
- User wants a wrapper with artifacts (use grill-me or grill-with-docs)
- Bug diagnosis (use diagnosing-bugs)
- Code review (use code-review)
- Implementation (use implement/tdd)
**局限**: No persistence mechanism (caller must handle). No domain-modeling integration (caller composes). Can go deep—frontier may expand faster than user answers. Requires user engagement (decisions are theirs).
## Related Skills
- **grill-me** (stateless wrapper), **grill-with-docs** (stateful wrapper + domain-modeling), **triage** (composes grilling), **wayfinder** (composes grilling), **improve-codebase-architecture** (composes grilling), **domain-modeling** (often composed with grilling), **to-questionnaire** (inverse: interviews user about send)
## Audit
V1✓ V2✓ V3✓ | 2026-08-26 | Source: D:\outcome\self\upstream\mattpocock__skills\skills\productivity\grilling\SKILL.md
Auf GitHub ansehen