survey
/survey
Prowadzi badanie rozwiązań w branży dwoma agentami, badaczem i sceptykiem, porównując narzędzia, opcje i kompromisy.
- Autor
- cookys ↗
- Repozytorium
- autopilot ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 14.11.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 1
- Skrypty
- 0
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 0
Brak sygnałów. Repozytorium zmienione 14.11.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
Co robi i kiedy
Prowadzi badanie rozwiązań w branży dwoma agentami, badaczem i sceptykiem, porównując narzędzia, opcje i kompromisy.
Wyzwalacze z opisu
Oryginalny opis (EN)
Research what the industry does — dual-agent (researcher + skeptic) tech investigation. Use when: "research X", "survey", "investigate options/tradeoffs", "what do others use", "industry standard for X", "compare X vs Y vs Z", "I'm not sure which to pick", "業界怎麼做", "調研一下", "別人怎麼處理的". Not for: strategic priority decisions (→ think-tank), brainstorming designs, or creating implementation plans.
Jak działa
- Trigger
- Boundary
- Flow
- Output Format
- Error Handling
- Signal Detection (for other skills)
Podgląd SKILL.md
---
name: survey
description: >
Research what the industry does — dual-agent (researcher + skeptic) tech investigation. Use when:
"research X", "survey", "investigate options/tradeoffs", "what do others use", "industry standard
for X", "compare X vs Y vs Z", "I'm not sure which to pick", "業界怎麼做", "調研一下",
"別人怎麼處理的". Not for: strategic priority decisions (→ think-tank), brainstorming designs,
or creating implementation plans.
---
# Survey -- Technology Research
> Routing overlap? If this intent better matches a sibling skill, redirect per [references/routing-tiebreaks.md](../../references/routing-tiebreaks.md) (prefer external evidence research over internal priority debate).
Two independent agents (researcher + skeptic) search in parallel, bringing different perspectives to avoid anchoring bias and recency bias.
## Trigger
| Method | When |
|--------|------|
| Manual | User says "survey", "research", "investigate", "what does the industry do" |
| Signal suggestion | Plan Mode / brainstorming detects tech selection / architecture decision / uncertainty -> suggest (user confirms before running) |
**Suggestion template** (when signal detected):
> "This decision involves tech selection. Suggest running `/survey` to research industry practices first. Proceed?"
## Boundary
- **Produce recommendation, marked as suggestion** -- attach reasoning and preconditions so user can judge applicability
- **No code** -- research only
- **No judgment-only auto-trigger** -- a survey runs without a Proceed prompt only when the unknown-escalation ladder probe (plan `project/plans/_archive/2026/09/2026-09-07-unknown-escalation-ladder.md`, ships with the ladder) recommends U2 with budget left, in every mode; the caller records the climb in the decision ledger. Judgment-only suggestions, knob-off and budget-exhausted cases keep the suggest-then-confirm template above
## Flow
### Step 1: Scope the Research
Extract from user input:
- **Topic**: what to research (e.g. "WS compression options")
- **Constraints**: known limits (e.g. "must support C++20", "latency < 1ms")
- **Existing bias**: does user already lean toward something? (researcher should specifically probe that option's weaknesses)
If input is vague (e.g. "look into cache"), ask one clarifying round before dispatch.
**Mode `issue-search`** (unknown-escalation ladder rung U2 for a `why` unknown — a bug that resisted two hypotheses): the input is an exact error / finding string plus the runtime and version tuple, not a topic. The researcher searches the literal string **as written** (at least one query verbatim — familiarity with the error is not a reason to skip the search), the skeptic checks whether each hit applies to *our* version; the report's Options table gains a "matches our version?" column. Budget: one round, no second generation; if nothing matches, the explicit "no public data" marker is the result and the caller climbs or stops per the probe. Callers: `debug` step 4, the foreman round end. After the report, the caller stamps the climb: `node scripts/probe-unknown.js receipt --ledger <ledger> --rung U2 --unknown-type why --terms <terms> --signals <ids>`.
### Step 2: Parallel Dispatch
**Model routing**: Read `.claude/model-routing-config.md` if exists; otherwise defaults from [references/model-routing.md](references/model-routing.md). Survey agents map to `researcher` → default: `model: "sonnet"` (needs web search tools, so NOT plan mode). **Effort floor: `medium`** on **anthropic-family seats**, both seats, every mode — at `low` effort Claude is less likely to call a search/retrieval tool and more likely to answer from memory (Anthropic "Prompting Claude Fable 5.1"), which turns a survey into a recollection. Do NOT apply this floor to families whose vendor documents low effort as search-capable — e.g. OpenAI's current reasoning guide (developers.openai.com/api/docs/guides/reasoning) lists low effort as "Ideal for use cases requiring tool-use, planning, search, or multi-step decision making" for their latest model; a forced bump there buys nothing and only adds cost/latency.
Spawn researcher + skeptic **simultaneously**. Skeptic does NOT wait for researcher -- independent search finds different angles.
```
Agent tool:
name: "survey-researcher"
subagent_type: "general-purpose"
model: "sonnet" # from model-routing config (researcher role)
run_in_background: true
prompt: -> references/prompts.md #Researcher (issue-search mode: #Issue-search Researcher)
Agent tool:
name: "survey-skeptic"
subagent_type: "general-purpose"
model: "sonnet" # from model-routing config (researcher role)
run_in_background: true
prompt: -> references/prompts.md #Skeptic (issue-search mode: add the #Issue-search Skeptic addendum)
```
> Full agent prompts with variable substitution: [references/prompts.md](references/prompts.md)
### Step 3: Merge Report
After both agents complete:
1. **Deduplicate** -- merge overlapping sources
2. **Cross-validate** -- researcher's pros vs skeptic's risks, same row per option
3. **Gap-fill** -- option with only pros and no risks (or vice versa) -> mark "incomplete data"
4. **Rank** -- by combined pros, risks, industry adoption
5. **Generate report** -- fixed format below
### Step 4: Present to User
Present report, clearly mark "decision is yours". If user wants to dive deeper into one option, run follow-up search.
## Output Format
```markdown
## Survey: {topic}
> Date: YYYY-MM-DD
### Background
{Why this research, 1-2 sentences}
### Options Comparison
| Option | Pros | Risks | Industry Adoption | Fit |
|--------|------|-------|-------------------|-----|
| A | ... | ... | ... | 3/3 |
| B | ... | ... | ... | 2/3 |
### Recommendation
{Recommended option + reasoning + preconditions (under what assumptions this holds). Decision is yours.}
### Sources
**Theory/Standards**
- [{title}]({url}) -- {one-line summary}
**Production Practice**
- [{title}]({url}) -- {one-line summary}
**Benchmark / Demo**
- [{title}]({url}) -- {one-line summary}
**Adoption Cases**
- [{title}]({url}) -- {one-line summary}
**Risk/Failure Cases**
- [{title}]({url}) -- {one-line summary}
```
## Error Handling
| Situation | Action |
|-----------|--------|
| WebSearch unavailable | Use LLM knowledge + WebFetch known URLs. Mark "no live search" in report |
| Agent timeout | Use partial results. Mark "{agent} incomplete" |
| No sources found for option | Keep in table, mark all fields "no public data" -- this is a risk signal |
## Signal Detection (for other skills)
| Signal | Example |
|--------|---------|
| Tech selection | "X or Y?", "which library?" |
| Architecture decision | "canonical approach for this?" |
| Novelty / no consensus | Computed, not phrased: the unknown-escalation ladder probe reports `unknown_type: how` (task terms with zero knowledge/memory/repo hits, or an unrecognised / fast-moving name) or `whether` (think-tank consensus LOW). A plan that reaches review carries no `TBD` (plan-template placeholder scan), so plan wording is not a signal |
| New domain | Tech/protocol/pattern team hasn't used |
| User bias | "I think X is good" -- worth verifying with survey |
Wersje
| Repozytorium | Wynik | Licencja |
|---|---|---|
| Ta strona · cookys/autopilot | 100 | MIT |
| akillness/jeo-skills | 100 | MIT |
| cookys/autopilot | 100 | MIT |
| epoko77-ai/vibestrata | 100 | MIT |
| mathbullet/skills | 100 | MIT |
| genkovich/sdd | 100 | MIT |
| quantumbfs/sci-brain | 100 | MIT |
| quickchatai/quickreview | 90 | MIT |
Migawka bazy z 13.11.2026 · wersja paczki 656d75edd10d (skrót SHA-256 całej paczki)
Pobrania w ostatnich 30 dniach: 0 · ranking
Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.
Podobne skille
- 95think-tank
Zestawia perspektywy ról: architekta, operacji, QA, produktu, UX i finansów, by ocenić kompromisy wybranej decyzji. - 100think-tank
Prowadzi debatę ról, np. architekta, operacji, QA, produktu i finansów, nad decyzją strategiczną i kompromisami. - 100gds-domain-research
Prowadzi badanie domeny i branży gier, tworząc raport badawczy na wskazany temat rynku lub gatunku. - 100research
Bada otwarte pytanie: opcje, rozwiązania, poprzednie prace i kompromisy, a potem pisze raport z rekomendacją, który nie zobowiązuje zespołu. - 100think-tank
Prowadzi wirtualny think tank, czyli ustrukturyzowaną debatę wielu person przed planowaniem lub decyzjami architektonicznymi, projektowymi i strategicznymi, w tym wyborem technologii i oceną kompromisów.