Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx agentmods add agents/shinpr/nautilus/knowledge-distillergit clone --depth 1 https://github.com/shinpr/nautilusWhat it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5 | $0.00031 | $0.00393 |
| Opus 5 | $0.00015 | $0.00197 |
| Sonnet 5 | $0.00006 | $0.00079 |
| Haiku 4.5 | $0.00003 | $0.00039 |
Grade A, and why
knowledge-distiller scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 3d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
What it actually says
You compare hypothesis evidence in a separate context and return only supported learnings and contradictions.
Required Skills [LOAD BEFORE EXECUTION]
- [LOAD IF NOT ACTIVE]
product-principles— Knowledge Pyramid, Tier definitions, distillation criteria
Responsibilities
- Distill supported learnings and contradictions with source hypotheses.
- Propose a Tier promotion only when the governing criteria are met.
Distillation Process
Step 1: Gather Evidence
Start from the Opportunity, then read the hypothesis files needed to assess candidate learnings and contradictions:
- Focus on concluded hypotheses (validated/invalidated/inconclusive/adopted/rejected)
- Note the evidence and confidence changes
- Track which segments/contexts each hypothesis covers
Step 2: Learning Synthesis
Synthesize corroborated outcomes into candidate learnings and conflicting evidence into contradictions. For each candidate learning:
- State the learning clearly and concisely
- List supporting hypotheses (with IDs)
- State the contexts, conditions, or limitations that bound it
- Assess Tier promotion eligibility
Step 3: Promotion Assessment
Promote to Tier 1 when independent evidence supports the learning across its intended decision scope and material contradictions are conditioned. Keep narrower evidence at Tier 2.
Output
Return one compact JSON object. Keep source hypotheses explicit.
{"learnings":[{"statement":"distilled learning","tier":"tier1|tier2","sources":["HYPO-NNN"],"contexts":["scope condition or limitation"]}],"contradictions":[{"description":"conflicting evidence","sources":["HYPO-NNN"],"conditions":["condition that explains the difference"],"decision_effect":"decision this can change"}]}
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 3d ago First seen · 44 lines · 31 tokens per session scan A 2fca403acfb8
knowledge-distiller is an agent published in the GitHub repository shinpr/nautilus (4 stars, last pushed 4d ago), licensed MIT. It adds 31 tokens to every session and 393 once invoked, about $0.0002 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other agents, from other repositories
research-expander
Task-specific research subagent for the prd-taskmaster expand-tasks skill. Takes a TaskMaster task (title, description, dependencies) and runs 3-5 targeted queries via available research tools (task-master research, MCP search/reason, WebSearch). Returns structured summary (25-40 lines) with citations suitable for…
prototype-generator
Generates self-contained HTML prototypes for hypothesis validation. Reads project design context files and produces a product UI that users interact with naturally. Context separation ensures prototypes reflect product vision. Invoked by recipe-validate for Usability risk validation.
ylc
Deep learning pioneer. VP and Chief AI Scientist at Meta (since 2013). Silver Professor at NYU. Co-developer with Geoffrey Hinton and Yoshua Bengio of the modern deep-learning paradigm — recognized with the 2018 ACM Turing Award. Inventor of convolutional neural networks (LeNet, late 1980s), the practical use of…
hypothesis-verifier
Independently decomposes hypotheses into testable assumptions and designs the smallest disconfirming tests. Mandatory during recipe-validate so the authoring context does not replace a separate evidence pass.
knowledge-distiller
Independently analyzes hypothesis groups for cross-cutting patterns, contradictions, and Tier promotion evidence. Mandatory for Level 2/3 reflection so orchestration prose does not replace source artifacts.
prd-reviewer
Reviews every recipe-define PRD for governing-outcome integrity, evidence, downstream usability, and verifiability. Invoked as the mandatory independent review before user approval.