Suggestive collaborative agent prompting

Suggestive collaborative agent prompting

One-liner

Field consensus: stop treating the agent as a silent implementer — force interview / options / tradeoffs / OK-gates before code, via skills (/grill-me), CLAUDE.md postulates (vibe-27), BLUEPRINT-style paste prompts, or a builder+vetting persona (operating-kit). Pair mode = propose creative paths with a recommended pick; human owns judgment. Anti-pattern: always-on clarifying that never ships (aviralmittal_).

Pin checked 2026-10-02 CST. Primary: X field posts + skill repos. Secondary: Exa docs/blogs. Reddit JSON blocked from box (network policy) — Exa Reddit queries returned thin; gap noted in Sources.

Related: 2026-10-02 Dual-model Sol Codex Claude UI advisor prompts (advisor vs implementer split) · 2026-10-02 Agent screenshot and UI expression (visual pair loop).

---

1. What “suggestive / collaborative” means in the wild

Dumb worker	Pair partner
Implements first guess silently	States assumptions; asks or presents options
One path → code dump	N options + tradeoffs + recommended pick
“Done” after generation	Waits for OK / acceptance criteria
Asks nothing (or asks forever)	Grills decision branches; looks up facts itself

X/skill language clusters around: interview before code, propose smallest slice, present tradeoffs, refuse silent implement, conversation not command.

---

2. Patterns from field posts & packs

A. /grill-me — interview until the design tree is closed

Matt Pocock’s mattpocock/skills (viral pack; DuncanRogoff: “most popular skills just make ai ask you questions” — DuncanRogoff/2103530565735424441; HeyAnjula: blamed the model until skills forced saying what you wanted — HeyAnjula/2097530046479265796; IcyCortex uses grill-me when feature is unclear — IcyCortex/2102843368589914311).

Mechanism (from grilling skill): map a design tree; work in rounds over the frontier (questions whose prerequisites are settled); each Q gets a recommended answer; explore codebase for facts instead of asking; do not act until shared understanding confirmed. User-invoked (disable-model-invocation on /grill-me front door). Sibling: grill-with-docs → writes CONTEXT.md + ADRs.

Luísa Lima workflow split: brainstorm/spike with Claude + grill-me; production with Codex — luisalima/2063893890877231595.

Install: npx skills add mattpocock/skills --skill grill-me (or plugin). Docs: aihero.dev/skills-grill-me.

B. BLUEPRINT / one-slice OK-gates (paste prompts)

cleofasrocha_ copy-paste set (cleofasrocha_/2102008202363576781):

1. BLUEPRINT — clarifying Qs + smallest vertical slice + files + stack + acceptance → wait for OK
2. ONE SLICE — implement only that feature E2E; no drive-by refactors

Same spirit as Spec Kit / SPEC.md loops and Justin Mitchel’s “interview me → write spec.md” AskUserQuestionTool prompt (Exa / justinmitchel.com).

C. Always-on CLAUDE.md behavior: vibe-27

fatalex27/vibe-27 — 27 postulates; highlights for suggestive mode: Editor not typist, Intention over implementation, Think before coding, Conversation not command, Three-level intent (What / Why / What-not). Drop into CLAUDE.md + AGENTS.md.

D. Builder + strategic-vetting persona

Sharrmavishal/operating-kit: not a passive executor — vet plans against code-truth; pause before shared-state mutations; no silent “done” (walk checks visibly); OOB reframe-first. Self-install via BOOTSTRAP-PROMPT.

E. AskUserQuestion / plan-mode interview (Anthropic-adjacent)

HeyAnjula summarizing claude-code-best-practice: plan mode + Ask Claude to interview you using AskUserQuestion tool — HeyAnjula/2071568279274205223. skillcreatorai: ask-questions-if-underspecified installable across agents — skillcreatorai/2006667642854363169.

F. Superpowers / brainstorm-before-code

VibeMarketer_: obra/superpowers “forces your agent to brainstorm before coding” among skills that actually changed workflow — VibeMarketer_/2058910021920944600.

G. Dual-model suggestion vs implementation

Suggestive planning on one model, implement on another (deepwhitman PRD-both-ways; dual Sol/Claude UI advisor note). Keep the suggestive agent in advisor/plan posture so it does not silently rewrite.

---

3. Pasteable snippets (attributed)

Snippet 1 — BLUEPRINT (cleofasrocha_)

Before writing code: ask clarifying questions, propose the smallest vertical slice,
list files you’ll touch, stack + acceptance criteria. Wait for my OK.

Source: x.com/cleofasrocha_/status/2102008202363576781

Snippet 2 — Minimal grill-me body (mattpocock; older minimal form still widely pasted)

---
name: grill-me
description: Interview the user relentlessly about a plan or design until reaching shared understanding, resolving each branch of the decision tree. Use when user wants to stress-test a plan, get grilled on their design, or mentions "grill me".
---

Interview me relentlessly about every aspect of this plan until we reach a shared understanding. Walk down each branch of the design tree, resolving dependencies between decisions one-by-one. For each question, provide your recommended answer.

Ask the questions one at a time.

If a question can be answered by exploring the codebase, explore the codebase instead.

Canonical front door now delegates to grilling (rounds/frontier). Repo: github.com/mattpocock/skills

Snippet 3 — vibe-27 collaborative core (fatalex27)

## Conversation, not command
- Prefer dialogue: propose → explain failure modes → refine → verify.
- Ask what happens when dependencies fail, input is garbage, or the happy path is skipped.
- Do not declare "done" until acceptance criteria are explicit or confirmed.

## Think before coding
- State assumptions explicitly. If uncertain, ask — never hide confusion.
- Present tradeoffs briefly when they matter; push back when a simpler path exists.

## Intention over implementation
- Clarify what and why before how.
- Never invent product requirements silently. If a choice changes UX, data, or scope — ask or state the assumption explicitly.

Source: github.com/fatalex27/vibe-27

Snippet 4 — Interview → SPEC.md (Justin Mitchel pattern; Exa)

read the spec.md file and interview me in detail using the AskUserQuestionTool about
literally anything: technical implementation, UI & UX, concerns, tradeoffs, etc.
but make sure the questions are not obvious

be very in-depth and continue interviewing me continually until it's complete,
then write the output spec to the file

Source: justinmitchel.com/posts/vibe-interview (mirrored widely on Medium/LinkedIn)

Snippet 5 — Options + recommend (synthesized field pattern)

Do not implement yet. Propose 2–3 approaches with tradeoffs (cost, risk, UX, migration).
Pick one as your recommendation and why. Ask me only the decisions you cannot infer
from the repo. After I pick, propose the smallest vertical slice and wait for OK.

---

4. Anti-patterns

Anti-pattern	Why it hurts	Field note
Always-on clarifying default	Agent thrash: “should I ask or answer?” after you once said “just answer”	aviralmittal_ — 2097049453123482051
Passive “agreed” grilling	Session feels productive; you rubber-stamp the agent’s plan	aihero: conversation not interview; push back
Ungrillable questions	Feel/layout taste can’t be talked into existence	Stop → prototype/screenshot → resume (aihero)
Grill on typo tasks	Overhead > value	Use for new/side-effectful work only
Codex under-grills / Claude over-grills	Same skill, different depth	dprophecyguy: Claude drills; Codex jumps after 2–3 Qs — 2066548832603914662; use grill-me-plus / explicit “keep drilling”
Analysis paralysis	Frontier never empties on mega-scope	Break into slices; grill each (aihero)
Plan mode eagerness	Rushes to write a plan asset	Prefer grill session first; leave plan mode off during grill (aihero)

---

5. Practical wiring for Noa

1. Default always-on (light): vibe-27 postulates 1–5 + 11 in project CLAUDE.md / AGENTS.md — suggestive without 40-question sessions.
2. Heavy design: invoke /grill-me (or /grilling) in a fresh chat; handoff to /to-prd or implement session without clearing context.
3. Feature start ritual: paste BLUEPRINT → OK → ONE SLICE.
4. Persona kit: operating-kit if you want builder+vetting + pause gates baked in.
5. UI taste decisions: ungrillable → visual loop (2026-10-02 Agent screenshot and UI expression).

---

Sources

X

• DuncanRogoff — mattpocock skills / grill-me
• HeyAnjula — skills fixed “wrong thing”
• HeyAnjula — AskUserQuestion interview
• IcyCortex — grill-me for unclear features
• cleofasrocha_ — BLUEPRINT / ONE SLICE
• aviralmittal_ — clarifying over-ask anti-pattern
• luisalima — grill-me spike / Codex prod
• dprophecyguy — Claude vs Codex grill depth
• VibeMarketer_ — superpowers brainstorm-before-code
• skillcreatorai — ask-questions-if-underspecified
• UnCorped — mattpocock grill-me / teach-me
• leeked — grill-me-plus
• hisevenih — grill-me拷问需求

Repos / docs (Exa)

• mattpocock/skills · grilling SKILL · aihero grill-me
• fatalex27/vibe-27
• Sharrmavishal/operating-kit
• Justin Mitchel interview→spec · obra/superpowers · GitHub Spec Kit clarify phase

Reddit / HN

• Reddit: box API blocked; Exa site:reddit.com returned no usable threads this pass — gap.
• HN secondary: Playwright MCP TIL (Simon Willison) more relevant to visual digest; grill-me mostly X/GitHub discourse.

Gaps

• Could not re-fetch full threads for vibe-27 Twitter handle (invalid from:), ssojet “Interview me before any code”, or full HeyAnjula grill thread bodies beyond cited posts.
• operating-kit matched (Sharrmavishal); confirm if parent meant a different “operating-kit” brand.