You have an idea. You can't explain it well. This skill reads the lossy prompt, recovers what you meant, expands it with corners you didn't think of, catches blind spots, verifies the new claims, and hands you a tight BLUF brief. No clarifying questions by default — it restates for recognition and lets you say "yeah" or correct with one word.
An 11-pass pipeline that runs in any agentic runtime (Hermes, Codex, Claude Code, Cursor, LM Studio — anything with a chat model and optionally web search). No CLI. No Python package. No env vars. The algorithm is markdown; the agent is the executor.
About
plugin/and the GitHub language badge: theplugin/__init__.pyis a Hermes slash-command handler (registers/4eyesand shows the depth picker before any pass runs) — not a Python package. It contains no library code, no installable API, no setup.py / pyproject.toml. The repo language badge reads "Python 100.0%" because the file extension is.py, but functionally the plugin is glue: one file, one slash command, nopip installsurface. If you don't use Hermes, you can delete theplugin/directory and the skill still works in any other runtime.
Trigger: /4eyes in Hermes, or just say "I have this idea but I can't explain it well."
| # | Pass | What It Does | Technique Source |
|---|---|---|---|
| 1 | RESTATE (steel-manned) | The strongest version of what you meant, not what you said | Steelmanning / charitable interpretation |
| 2 | CLARIFY-CHECK | Internal ambiguity check. Asks ONE structured question (2–4 numbered options) only if no interpretation covers ≥60% of the prompt's surface elements. Otherwise proceeds silently. (CLAMBER is the reason the default is "don't ask" — the threshold is the structural test that prevents over-correction.) | Threshold rule (this skill) |
| 3 | EXTRACT (abductive empathic mirroring) | Breaks restatement into 3–5 atomic intents via the unilateral OARS moves (open questions, affirmations, reflections, summaries applied to the text itself, not a live patient) + abductive hypothesis selection (2–3 competing readings, pick best by explanatory power). OARS is a clinical technique designed for two-way human dialogue; we use only the unilateral mirroring moves. | Miller & Rollnick OARS unilateral moves (2013) + Peirce abductive reasoning |
| 4 | DECOMPOSE (first principles) | Breaks idea into 3–5 irreducible relational components for structural search | First-principles thinking (Aristotle → Musk) |
| 5 | PRIOR-ART (structural search) | Searches by mechanism per decomposed component, finds analogs in unrelated domains. Fans out via delegate_task when available. |
Gentner Structural Mapping Theory (1983) — DrugPatentWatch is an operational example, not a technique source |
| 6 | EXPAND (SCAMPER + constraint relaxation) | Transforms the solution (SCAMPER: 7 transforms) AND transforms the rules (constraint relaxation: what if budget/time/platform weren't constraints?). Pick the 2–3 most derivable from Pass 3 intents that aren't already covered by Pass 5 analogs. | SCAMPER (Eberle 1971) + constraint relaxation (ScienceDirect 2025) |
| 7 | INVERT (pre-mortem + lateral hats) | "Imagine this failed in 12 months, why?" + selective Yellow Hat (best-case) + Blue Hat (is this the right problem?) for vision-stage ideas | Klein pre-mortem (2007) + de Bono Six Thinking Hats (1985) |
| 8 | CONVERGE | Picks top 2–3 expansions from passes 5–7 by the Pass-6 criterion (most derivable from Pass 3 intents, not already covered by Pass 5 analogs). Drops the rest. Prevents output bloat. | Selection criterion (this skill) |
| 9 | SELF-CHECK | One reflection: does restate match extraction? Did prior-art find real analogs? Are expansions derivable from intents? Catches recycling. Non-recursive by design. Failure path: any failure is flagged in "What I can't verify" and the pipeline proceeds to Pass 10. | Single-reflection consistency audit (this skill) |
| 10 | SEARCH (parallel) | Evidence-searches surviving claims only. Fans out via delegate_task when available — one subagent per claim. Sequential fallback otherwise. |
delegate_task (Hermes) / subagents (Claude Code) |
| 11 | VERIFY → FORMAT | Confirms/corrects claims against evidence, leads with the bottom line (conclusion first = BLUF principle), then evidence — and keeps total output under 500 words (separate word-budget constraint). | BLUF principle (US Army FM 6-99 / USAF Tongue and Quill) — structural principle, not a word count; word budget is a separate constraint. |
**What I think you're going for:** [steel-manned restatement — the strongest version of what you meant, not a passive mirror]
**The core ask, broken down:**
- [atomic intent 1]
- [atomic intent 2]
- [atomic intent 3]
**Prior art (if found):**
- [1-2 analogs, only if from a domain the user is unlikely to have considered]
**Corners you might not have looked at:**
- [converged derivative 1 — SCAMPER / constraint relaxation / pre-mortem / adjacent-field]
- [converged derivative 2]
**Where I'd push back:**
- [1 specific claim evidence contradicts]
**What I can't verify:**
- [1-2 unverifiable claims]
Sources: [list or omit]
When pass 2 detects genuine ambiguity (2+ equally plausible interpretations), it asks ONE structured question before continuing:
**I think you're asking about one of these — pick one:**
1. [Interpretation A — one sentence]
2. [Interpretation B — one sentence]
3. [Interpretation C — one sentence]
*(Or tell me I'm wrong — one word is enough.)*
- Numbered list, 2–4 options, each one sentence
- User picks a number → pipeline resumes
- Never asks twice — if still ambiguous, picks most likely, states assumption, proceeds
- Only fires when guessing would be irresponsible. Most prompts can be restated without asking.
When the runtime supports subagent delegation (Hermes delegate_task, Claude Code subagents, etc.):
- Pass 5 (PRIOR-ART): One subagent per decomposed component. Each searches a different domain. Results merge.
- Pass 10 (SEARCH): One subagent per surviving claim. Each searches for evidence. Results merge.
When the runtime does NOT support delegation (LM Studio, Ollama, minimal Cursor):
- Sequential mode with concrete turn budgets
- Pass 5: 2–3 searches batched in one turn
- Pass 10: 1–5 searches batched in one turn
- Same algorithm, slower wall time
/4eyes I want to build something like notion but for backend engineers
Or just say:
I have this idea but I can't explain it well — [fuzzy idea]
Copy SKILL.md and the references/ directory into your skill directory. Load the skill, follow the 11 passes, use whatever web search tool your runtime exposes.
Works end-to-end without web search. The "Where I'd push back" and "What I can't verify" sections degrade to "based on model knowledge only." The expansion still works.
| File | Contents |
|---|---|
SKILL.md |
Frontmatter + triggers + 11-pass overview + pitfalls table (18 entries) + cost profile + runtime adapters |
plugin/ |
Optional plugin/__init__.py — Hermes slash-command handler that registers /4eyes and shows the depth picker before any pass runs. No pip install surface; safe to delete on non-Hermes runtimes. |
references/algorithm.md |
Full pass-by-pass algorithm with worked end-to-end example (11 passes, "Notion for backend engineers" prompt) |
references/output-template.md |
Output template + clarify-check interactive format + display rules + anti-patterns + word budget |
references/research-findings.md |
15 research sections + landscape analysis of 7 comparable GitHub projects + design-choice evidence table |
findout |
grounded-research |
4eyes |
|
|---|---|---|---|
| User knows what they want | Yes (verify claim) | Yes (research topic) | No (fuzzy idea) |
| Direction | Model → web (verify) | Web → model (build) | Lossy input → structured output |
| First action | Generate answer | Web search | Restate for recognition |
| Expansion built in | No | No | Yes (SCAMPER + constraint relaxation + pre-mortem) |
| Blind-spot coverage | No | No | Yes (inversion + lateral hats) |
| Clarify when needed | No | No | Yes (one structured question, rare) |
| Parallel search | No | Yes (parallel web search fan-out) | Yes (delegate_task fan-out) |
| Best for | "Is X true?" | "What's the state of X?" | "I have an idea but can't explain it" |
Mental model: findout = checker. grounded-research = builder. 4eyes = interpreter.
Papers / projects surveyed while scoping 4eyes. Star counts are not shown — most entries are research papers, not maintained repos, and a single star number would be cherry-picked (the one maintained repo, AutoPrompt, has ~3k; everything else is a paper or a one-shot). The "vs 4eyes" column is the point.
| Project | Kind | What It Does | vs 4eyes |
|---|---|---|---|
| AutoPrompt (Eladlev) | Repo (~3k stars) | Intent-based prompt calibration via synthetic boundary cases | Optimizes machine input, assumes user knows intent |
| ClarifyGPT (ACL 2024) | Paper | Detects ambiguity in code prompts, asks clarifying questions | Code-only, asks questions (anti-pattern for fuzzy persona) |
| Clarify-AI (sirutisb) | Repo | Ambiguity detection → human-in-the-loop | Interactive (what we avoid by default) |
| R-Bot (VLDB 2024) | Paper | LLM query rewrite with self-reflection for retrieval | Retrieval optimization, not intent recovery |
| CLAMBER (ACL 2024) | Paper + benchmark | 12K-query benchmark: LLMs are bad at ambiguity detection | Validates "don't ask, restate" as the default; threshold rule is the structural check against over-asking |
| OpenClaw SCAMPER | Repo | Single-pass SCAMPER agent skill | No recovery, no prior-art, no verify — a toy |
| cross-domain-analogy | Repo | Embedding-based structural analogy discovery | Mechanism for our pass 5; we use language instead |
Bottom line: No existing project combines no-questions intent recovery + multi-pass expansion + evidence verification. The gap is real.
| Technique | Primary source | Supplementary |
|---|---|---|
| Articulation barrier | Nielsen, J. (2024). "The Articulation Barrier: Prompt-Driven AI UX Hurts Usability." UX Tigers. https://www.uxtigers.com/post/ai-articulation-barrier | — |
| Steelmanning | Steelmanning: The Lost Art of Arguing Against the Best Version of an Opponent. Future of Thought. https://futureofthought.substack.com | Also: debateladder.com, umbrex.com |
| Clarify-check threshold | Threshold rule defined in this skill (≥60% surface coverage) — the operational test for when the model is allowed to ask. The principle ("don't ask, restate") is borrowed from the CLAMBER finding; the specific 60% threshold is a 4eyes invention, not from CLAMBER itself. | arXiv 2505.13360 — underspecified prompts 2x regression (general evidence that asking is expensive) |
| OARS empathic mirroring (unilateral moves only) | Miller, W. R. & Rollnick, S. (2013). Motivational Interviewing: Helping People Change (3rd ed.). Guilford Press. | Used unilaterally (mirroring moves applied to the prompt text, not a live patient). Two-way clinical use is not what's happening here; we're explicit about that. |
| Abductive reasoning | Peirce, C. S. — Inference to the Best Explanation (pragmatic maxim, 1878) | Harman, G. (1965). "The Inference to the Best Explanation." Philosophical Review 74(1): 88–95. |
| First-principles decomposition | Aristotle, Physics (Book I, §1) — "the way of inquiry… proceeds by analysis" | Operationalized in modern product thinking (popularized by Elon Musk, 2013 interview; documented by FourWeekMBA). The Aristotle reference is the philosophical lineage; the Musk reference is the popular shorthand. |
| Structural prior-art search | Gentner, D. (1983). "Structure-Mapping: A Theoretical Framework for Analogy." Cognitive Science 7(2): 155–170. https://doi.org/10.1207/s15516709cog0702_3 | Operational example: DrugPatentWatch (commercial patent DB) demonstrates the structural-search principle at scale. We use language; the patent DB uses embeddings — same principle. |
| SCAMPER | Eberle, B. (1971). Scamper: Games for Imagination. D.O.K. Publishers. | Teaching guide: IMD. "SCAMPER Design Thinking Technique: Complete Guide & Examples." https://www.imd.org/blog/innovation/scamper-method-design-thinking/ — useful for prompts, not for the underlying theory. |
| Constraint relaxation | Constraint-relaxation research: "Neural dynamics of constraint relaxation and problem representation," ScienceDirect 2025. https://www.sciencedirect.com/science/article/pii/S0166432825004000 — empirical work on the "Aha!" transition from impasse to insight | Cross-domain analogical reasoning: arXiv 2509.09381 (2025). The behavioral pattern (relax a constraint → reframe → "Aha!") is documented across cognitive science. |
| Pre-mortem | Klein, G. (2007). "Performing a Project Premortem." Harvard Business Review. https://hbr.org/2007/09/performing-a-project-premortem | Munger inversion: Poor Charlie's Almanack (2005). McKinsey has operational pre-mortem guides. |
| Six Thinking Hats (selective) | de Bono, E. (1985). Six Thinking Hats. Little, Brown. | The Decision Lab — "Six Thinking Hats: A Practical Guide." |
| Convergence (selection from a list) | Internal rule (Pass 6 selection criterion, applied in Pass 8) — no external citation. "Pick the 2–3 most derivable from Pass 3 intents not already covered by Pass 5 analogs." We do not cite Double Diamond here; the British Design Council framework is a full product design process (discover / define / develop / deliver), not a list-selection rule. | — |
| Self-check (single reflection) | Internal rule — one consistency pass, no loop. Self-Refine (Madaan et al. 2023, arXiv 2303.17651) is about iterative self-improvement and does NOT apply to a non-recursive single reflection. Earlier drafts cited Self-Refine here; that was wrong and has been removed. | |
| BLUF principle (structural, not a word count) | US Army Field Manual 6-99 (Army writing) + USAF "Tongue and Quill" (communication doctrine) — BLUF means "Bottom Line Up Front," i.e. lead with the conclusion, then evidence. It is a structural principle, not a word count. The word budget (~150 / 300 / 500 by mode) is a separate constraint on evidence economy, derived from user-experience research on the fuzzy-explainer persona, not from military doctrine. | Civilian adoption: Mind Tools "BLUF" communications guide. (Earlier draft cited MIL-STD-2525, which is battlefield map symbology, and JP 2-01, which is intel prep of the operational environment — neither is a prose-writing standard. Corrected.) |
Full citations + quoted passages in references/research-findings.md.
MIT — see LICENSE. Copyright (c) 2026 Akash Mukherjee.
Scope of this license: the MIT license covers the implementation of the pipeline in this repo (the markdown spec, the algorithm, the output template, the per-mode placeholder guide, and any code in plugin/). It does not extend to the underlying techniques the pipeline orchestrates (SCAMPER, pre-mortem, OARS, Six Thinking Hats, Double Diamond, BLUF, etc.) — those are referenced as prior art and remain under their original authors' copyrights. See references/research-findings.md for the full provenance.