| name | url-insight |
| description | Auto-triage pasted URLs into signal-tagged, ROI-scored, DoR-compliant portfolio insights. Fetches via nab, classifies (paper/product/repo/blog/docs/benchmark), cross-maps to 21-repo portfolio, dispatches specialist agent panels, runs devil + steelman critiques, scores ROI, auto-files P0-P2 issues via linear skill. Dedupes via hebb. Auto-fires on URL batches, "analyze these urls", "what's interesting in", "should we adopt X", "competitive signal", "worth a spike", multi-URL pastes. |
| triggers | analyze these urls|what's interesting in|should we adopt|competitive signal|worth a spike|triage urls|url batch|read these links|insight from|digest these|landscape scan|what did I miss|portfolio signal |
| effort | medium |
| version | 2026.05.06-v5 |
url-insight — URL batches → portfolio insights → DoR issues
Fires on: multi-URL paste · "analyze these urls" · "should we adopt X" · "competitive signal" · "worth a spike" · arxiv/HN/ProductHunt/GitHub URLs in bulk.
Purpose: compress a firehose of URLs into ranked, evidence-backed actions against the 21-active-repo portfolio. Noise → IGNORE. Signal → specialist agent panel → Linear ticket with devil + steelman reckoning.
Design: Fukasawa (paste URL → pipeline fires) · Rams #10 (8 labels, not 30) · Honest (IGNORE is first-class) · Calibrated (devil + steelman together).
0 · MANDATORY protocol (non-negotiable, like hebb/linear)
- Before fetching any URL: run the dedup pre-check in §0.1. If hebb or Linear finds a prior verdict, skip fetch, skip nab, and report
Already triaged: <MIK-id>, verdict=<ADOPT|IGNORE|INSPIRE|SUPERCHARGE|HUMAN-REVIEW>, ROI=<n>x with the issue link.
- After each URL is classified + scored: write the hebb dedup pin in §0.1 with
project="url-insight" and key schema url-<sha256-of-canonical-url>-<iso-date>.
- Before filing any issue: run devil's advocate (§6) AND steelman (§6.5). Quote both verbatim in issue body.
- For SUPERCHARGE: dispatch innovator agent FIRST (serial gate). If kill → drop signal. If proceed → parallel fan-out: moonshot-architect + business-panel + research patent-prior-art.
- For FAIL-FAST SPIKE: surface user prompt "file issue / also spawn <1h spike agent?" (unless
--auto-spike flag).
- Emit innovation-tracker queue entry per actionable insight (§10).
- Wayback evidence capture (BLOCKING for non-IGNORE): after a successful fetch, call
fulcrum:wayback_save with if_not_archived_within="7d" and capture_screenshot="1"; persist wayback_url or wayback_save_failed=true in the run JSON (§1.1). Capture failure is evidence debt, not a triage blocker.
- Dual output: JSON artifact
~/.claude/data/url-insight/run-TIMESTAMP.json + markdown report.
- Living-document update (BLOCKING for non-IGNORE): append delta line(s) with
[^N] citation(s) to ~/.claude/data/portfolio/market-positioning-living.md per §0.8. No skipping. No "I'll do it later". The doc rots without per-URL writes.
Confidence: V (verified ≥2 sources) · I (inferred from 1) · A (assumed). Mark every ROI claim.
Token budget per URL (hard cap 2M total per batch, 50 URL hard cap):
SUPERCHARGE ~80K · INSPIRE/REPOSITION ~40K · FAIL-FAST/INTEGRATE ~20K · ADOPT ~10K · IGNORE 0.
Override flags: --shallow (skip panels, 1-pass) · --deep (full panel even for INTEGRATE).
0.1 · Dedup pre-check (MIK-3372, FIRST PHASE before nab-fetch)
This is the first phase after URL parsing and before any fetch, nab_fetch, browser, or agent analysis. Duplicate detection must be cheap and deterministic.
- Canonicalize URL with
skills/url-insight/dedup_precheck.py:
- lower-case the host
- strip trailing slash from non-root paths
- remove fragments
- strip
utm_*, utm_source, utm_medium, utm_campaign, utm_term, utm_content, fbclid, gclid, ref, and source
- compute
sha256(canonical_url)
- hebb recall:
- query: canonical hash from step 1
- project:
url-insight
- limit:
3
- matching prior pin key format:
url-<sha256-of-canonical-url>-<iso-date>
- Linear search:
- search title/body for the canonical host and the strongest path or repo component
- example terms from
dedup_precheck.py --url "$URL" --json: linear_search_terms
- Short-circuit on hit:
- print
Already triaged: <MIK-id>, verdict=<ADOPT, IGNORE, INSPIRE, SUPERCHARGE, HUMAN-REVIEW>, ROI=<n>x plus the Linear link
- exit the URL pipeline for that URL
- do not call
nab_fetch, browser fetch, Wayback save, classifier agents, or Linear issue creation for the duplicate
- Pin after successful fresh triage:
- call
mcp__hebb__remember with project="url-insight", tags including url-insight, url-canonical-hash, the signal label, and the canonical host
- key schema:
url-<sha256-of-canonical-url>-<iso-date>
- body must include at least: original URL, canonical URL, canonical hash, Linear ticket identifier if filed or reused, verdict, ROI, and run artifact path
Helper command:
python3 skills/url-insight/dedup_precheck.py --url "$URL" --json
False-positive safety: if the dedup check blocks 3 legitimate re-triages in 7 days, tighten canonicalization and add a --force path before further blocking changes.
0.8 · Living-document update mandate (NON-NEGOTIABLE — ships every URL)
File: ~/.claude/data/portfolio/market-positioning-living.md — the canonical investor / partner / new-agent primer for the portfolio's market posture.
Rule: Every URL classified as anything other than IGNORE / IGNORE-WATCHLIST MUST result in at least one delta line appended to the living document. No exceptions. The living doc decays into stale opinion-art the moment we let an analysis ship without writing back.
Section-mapping (where the delta goes):
| Classification result | Target section in living doc | Delta content |
|---|
| SUPERCHARGE filed | §6 SUPERCHARGE candidates table | new row with Linear ID + EXCEED axis + status |
| New cumulative-axis signal (per MIK-3244) | §5 Cumulative axis signals table | append signal to existing axis row OR create new row |
| Vocabulary critique / external messaging audit | §4 Vocabulary | new AVOID / USE entry with source citation |
| New competitor + product launch | §3 Market thesis (relevant wedge) | competition list update + reference link |
| Honest-weak-spot reveal | §7 Honest weak spots | numbered entry with mitigation + Linear ticket if filed |
| Wedge TAM revision | §3 Market thesis (relevant wedge) | TAM range update with new evidence link |
| Active-ticket re-surfacing (per §4.1 gate) | §11 Source ledger | one-liner cross-ref |
Citation requirement: every delta line MUST include at least one [^N] reference resolving to §12 of the living doc. If §12 does not yet contain the source, add it in the same edit. No claim ships without provenance. A delta with no [^N] is a protocol violation — the user has explicitly directed: "if there is no valid links someone could think it is just a collection of opinions. we are having this fact based + our own conclusions" (2026-05-03).
§12 entries must carry: source URL (V-fetched at addition time), one-line summary, confidence marker (V / I / A), fetch date. Confidence-I and -A entries MUST be upgraded to V within 30 days plus a second corroborating source, OR the dependent claim is removed.
Operational sequence per URL:
- Run pipeline steps 1–13 (§1).
- If signal != IGNORE: identify target section per table above.
Read market-positioning-living.md (last-edit check + section anchor).Edit to append delta line with [^N] reference.- If new source:
Edit §12 with full reference (URL + summary + V/I/A + date).
- Increment doc version at footer (
v1.x → v1.x+1) on every edit, even single-line. Stale version = stale doc.
- Commit message MUST cite the originating URL hash + Linear ticket(s) for audit.
Quarterly consolidation (first Monday Jan/Apr/Jul/Oct): operator (or /loop ticked task) merges duplicate axis-signal rows, prunes >6mo-stale signals with no re-trigger, refreshes TAM ranges, upgrades (I) / (A) markers.
Audit failure modes:
- Delta added without
[^N] → protocol violation; revert + redo.
- §12 reference added without V-fetch → mark
(A) + log a 30-day promotion task in Linear.
- Wedge TAM revised without external evidence → roll back the TAM revision; keep evidence link only.
- Footer version not bumped → CI / pre-commit reformat will flag.
This subsection is BLOCKING the same way §0 is BLOCKING — skipping it leaves the living doc lying to its readers, which destroys its value as an investor primer and as new-agent ramp-up material.
1 · Pipeline (single pass, 13 steps)
1. PARSE URLs from message (regex \bhttps?://\S+\b)
2. DEDUP canonical URL hash + hebb recall + Linear search (§0.1); skip fetch if already triaged
3. FETCH nab_fetch MCP (PREFER); nab_fetch_batch for ≥3 URLs
Fallback: `nab fetch URL` CLI (--cookies brave / --1password)
NEVER WebFetch (~50K token waste; see linear skill §5)
3b. ARCHIVE successful 2xx fetches with fulcrum:wayback_save (§1.1)
4. CLASSIFY URL type: paper · product-launch · blog · repo · docs · benchmark · company · other
5. ROUTE enrichment by type (§3)
6. EXTRACT single-pass: claim · differentiator · numbers · date · surprise
7. CREDIBILITY tier S/A/B/C/D (§5); multiplies ROI probability
8. CROSS-MAP to 21 active repos (§4); v1 descriptions, v2 CLAUDE.md
9. CLASSIFY SIGNAL → 1 of 8 labels (§2)
10. DISPATCH AGENTS by signal (§6); parallel fan-out, normalized JSON return
11. DEVIL + STEELMAN reconciliation (§6.5); calibrated verdict
12. SCORE ROI + compose DoR-compliant issue body (§7) → linear skill (§8)
13. REPORT sorted table + IGNORE list + JSON + queue emit (§9)
Staleness: if published_date > 24mo ago → flag [DATED, check SOTA], cap ROI probability at 0.4.
1.1 · Wayback evidence capture (MIK-3354)
Run immediately after each successful 2xx fetch, before classification:
archive = gateway_execute("fulcrum:wayback_save", {
"url": url,
"if_not_archived_within": "7d",
"capture_screenshot": "1",
})
Capture results are never allowed to block classification, agent fan-out, or Linear filing. They do change evidence quality:
- If the tool returns an archive/timestamp URL, store it as
wayback_url on the URL row in ~/.claude/data/url-insight/run-TIMESTAMP.json.
- If the tool returns a pending job, store
wayback_job_id and follow with fulcrum:wayback_availability; fill wayback_url when availability returns a timestamp URL.
- If capture or availability fails, store
wayback_save_failed=true and wayback_error with a short sanitized reason. Do not log credentials, cookies, request headers, or page bodies.
- For every non-IGNORE Linear issue/comment, include both
Source URL and Wayback URL when wayback_url exists. If it does not exist, include Wayback capture: attempted, failed non-fatally so the evidence gap is visible.
- Report artifacts must include
wayback_url or wayback_save_failed for every non-IGNORE URL.
This is source-of-record preservation, not content validation. The fetched body remains the analytical input; Wayback is the durable evidence trail.
2 · Signal schema — 8 labels (Rams #10)
| Label | Definition | Default Priority |
|---|
| SUPERCHARGE | Combine with existing repo strength → 10x moat. Highest value, often overlooked. | P0 |
| REPOSITION | Competitive landscape shift → update value prop / docs / messaging | P1 |
| INSPIRE | Build our own superior version | P1 |
| INTEGRATE | Wire in as dependency (MCP server, SDK, service) | P2 |
| ADOPT | Use directly (library, tool, API) | P2 |
| FAIL-FAST SPIKE | Cheap <1-day test to validate before commit | P2 |
| RESEARCH | Multi-day empirical validation required | P3 |
| IGNORE | Noise, no portfolio signal. Terminates pipeline — no agent dispatch, no issue, no cross-map. | — |
Rule of least regret: SUPERCHARGE vs INSPIRE → pick SUPERCHARGE (10x beats 1x). INSPIRE vs IGNORE → pick IGNORE (noise kills signal).
Repeat-competitor escalation: same domain ≥3× across batches (via hebb.recall(query=DOMAIN, tag="url-insight")) → auto-add REPOSITION alongside original. "Competitor shipped 3 things this month, that is a market signal."
3 · URL-type routing (enrichment strategy)
| Type | Host patterns | Additional enrichment |
|---|
| paper | arxiv.org, doi.org, biorxiv, researchgate, *.edu | research skill: semantic_scholar (citations+influence), openalex (author h-index) |
| product-launch | news.ycombinator.com (Show HN), producthunt.com | Fetch linked comparison pages / docs |
| repo | github.com, gitlab.com, codeberg.org | gh api /repos/OWNER/NAME → stars, last-commit, topics, license |
| blog | medium.com, substack.com, *.dev, personal blogs | Author credibility (openalex + ror if org-attributed) |
| docs | readthedocs.io, mdbook, .docs, docs. | Version + last-updated |
| benchmark | paperswithcode.com, llm-leaderboard, openreview | SOTA delta vs our numbers if repo-adjacent |
| company | .com home, pitch decks, investor pages | ror_search_organization → Fortune500/FAANG/indie tier |
| other | anything else | credibility tier only; often IGNORE |
4 · Cross-map to portfolio (21 active repos)
Source of truth: ~/.claude/data/url-insight/portfolio.json (generated artifact).
Schema: {generated_at, source, filter, count, repos: [{name, path, description, tagline, visibility, language, pushed_at, commits_30d, has_claude_md, local_clone}, ...]}.
Filter: owned (isFork=false) + not archived + pushedAt ≥ CUTOFF (currently 2026-03-01) + strategic (sites/profile READMEs dropped).
Refresh (weekly; stale if generated_at > 14 days old):
python3 ~/.claude/skills/url-insight/refresh_portfolio.py
gh repo list MikkoParkkola --limit 200 --json name,description,isArchived,isFork,pushedAt,visibility,primaryLanguage \
| python3 -c "import json,sys,pathlib,subprocess,re,datetime; \
raw=json.load(sys.stdin); H=pathlib.Path('/Users/mikko/github'); C='2026-03-01'; \
DROP={'MikkoParkkola','parkkola-website','revaluator-legal','revaluator-website','domain-tracker'}; \
a=[r for r in raw if not r['isArchived'] and not r['isFork'] and (r['pushedAt'] or '')>=C and r['name'] not in DROP]; \
# ... see refresh_portfolio.py for full enrichment logic"
v1 match (always available): token-overlap between extracted claim and description field. Top-3 repos by Jaccard, min-score ≥ 0.15 to qualify.
v2 match (when has_claude_md=true): grep the local CLAUDE.md (first 2000 chars). Weight v2 match 2× v1.
Inline Python helper (Rams #10 — consumes payload["repos"]):
import json, re, pathlib
payload = json.loads(pathlib.Path("~/.claude/data/url-insight/portfolio.json").expanduser().read_text())
portfolio = payload["repos"]
def cross_map(claim: str, top_k=3):
claim_toks = set(re.findall(r'\w{4,}', claim.lower()))
scored = []
for r in portfolio:
desc = (r.get("description") or "").lower()
v1_toks = set(re.findall(r'\w{4,}', desc))
v1 = len(claim_toks & v1_toks) / max(len(claim_toks | v1_toks), 1)
cmd_path = pathlib.Path(r.get("path") or "") / "CLAUDE.md" if r.get("has_claude_md") else None
v2 = 0.0
if cmd_path and cmd_path.exists():
text = cmd_path.read_text(errors="replace")[:2000].lower()
v2_toks = set(re.findall(r'\w{4,}', text))
v2 = 2 * len(claim_toks & v2_toks) / max(len(claim_toks | v2_toks), 1)
score = max(v1, v2)
if score >= 0.15:
scored.append((r["name"], score, r["tagline"]))
return sorted(scored, key=lambda x: -x[1])[:top_k]
Empty result → NOT yet IGNORE. Run the Linear-active-tickets gate below before discarding.
Staleness guard: if payload["generated_at"] > 14d old, warn + trigger refresh before cross-mapping.
Companion context: ~/.claude/data/url-insight/portfolio-context.md holds the strategic + architectural portfolio narrative — cross-repo dep graph, roadmap themes, ADR inventory, companion-bundle pattern, open-issue highlights. Consult when a URL match needs grounding a one-line description cannot provide (e.g. "does this paper relate to the metacognition G4 sibling-wiring gate?").
4.1 · Linear active-ticket gate (BLOCKING before IGNORE)
Lesson learned 2026-04-24 (MIK-2997) and re-surfaced 2026-05-03 (Vision Banana, Vikas Chandra video roadmap): portfolio.json one-line descriptions DO NOT tokenize active research-direction keywords. Repos like botnaut-server describe themselves at the product level; their actual research scope (multimodal, metacog, world-model, mesh-shard, diffusion-paradigm) lives in CLAUDE.md integration matrices AND in active Linear umbrella tickets (MIK-3209 multimodal, MIK-3094 N3 mesh-shard, MIK-3015 cost plane, MIK-3110 sovereign policy, etc).
Rule: BEFORE classifying any URL as IGNORE / IGNORE-WATCHLIST, query Linear for active backlog tickets matching the URL's domain keywords. If any active ticket scope overlaps, classify as REPOSITION / INSPIRE / RESEARCH (with cross-link comment) instead of IGNORE.
Implementation (inline, in pipeline step 8):
domain_kws = extract_domain_keywords(claim)
matching_tickets = linear_search_active(domain_kws, max_results=5)
if matching_tickets:
cross_map_result = matching_tickets
signal = decide_signal(claim, matching_tickets, credibility_tier)
else:
pass
Active-ticket keyword index (refresh quarterly; current as of 2026-05-03):
| Domain keyword | Active umbrella tickets |
|---|
| multimodal / video / vision-foundation / VLM / on-device | MIK-3209 (Qwen Thinker/Talker), MIK-3324 (Vikas roadmap), MIK-3209 sub-area |
| memory / persistent-memory / consolidation | MIK-3323 (stash competitor), hebb portfolio |
| cost / budget / token-spend / quota / cost-control | MIK-3015 (Attested External Cost Plane) |
| sovereignty / sovereign-stack / vendor-lock / closed-API | MIK-3110, MIK-3011 |
| diffusion / generative-pretraining / RGB-output / Marigold | MIK-3311 (zlab/dflash), MIK-3244 (cumulative axis) |
| voice / STT / TTS / speech-tags / WER | MIK-3313 (Grok Omni), MIK-3114 |
| metacognition / CFSR / continuous-microtraining | MIK-3243 (LADIR), MIK-3244, botnaut-metacog |
| mesh-shard / N3 / expert-routing / MoE | MIK-3094 (N3 EPIC), MIK-3097 |
| compute-rationing / capex / GW-scale / EUV | MIK-3015 (Lee.aao evidence), MIK-3011 |
| attestation / signed-state / .state | MIK-2850 (botnaut-attestation), MIK-3110 |
| RL / scaling-laws / RLHF / DPO | MIK-2997, MIK-3243 |
| quantization / NVFP4 / W4A16 / ParetoQ / KV-cache | MIK-3324 (cross-link), nvfp4-mojo repo |
| Apple-Metal / on-device-runtime / ExecuTorch | MIK-3324, bnaut-metal |
Refresh cadence: this index goes stale fast. After each url-insight session, append any new umbrella tickets created. Quarterly: prune closed tickets + re-derive from active backlog.
Failure mode this prevents: classifying a paradigm-shift signal as IGNORE because portfolio.json says "botnaut-server: planner-worker-judge agent server" without tokenizing "multimodal" / "video" / "vision-foundation" — the actual active research direction.
5 · Credibility tier (S/A/B/C/D)
Prevents hype-driven false-positive SUPERCHARGE. Multiplies ROI probability.
| Tier | Factor | Signals |
|---|
| S | 1.0 | Top venue (NeurIPS/ICML/ICLR/Nature) · FAANG/DeepMind/OpenAI · author h-index >50 |
| A | 0.8 | arxiv top-1% cited · Fortune500 eng blog · h-index 20-50 |
| B | 0.6 | arxiv default · well-known company · h-index 5-20 |
| C | 0.4 | personal blog (established author) · small startup · new PhD |
| D | 0.2 | random substack · anonymous · no verifiable affiliation |
Lookups (free, via research skill / fulcrum MCP):
orcid_get_person + openalex_get_author → h-indexror_search_organization → org tier- arxiv category + submission pattern → venue proxy
Default when unresolvable: C (0.4). Don't assume S without evidence.
6 · Agent dispatch by signal (REPLACES simulated personas)
Claude Code has real specialist agents. Dispatch via Agent tool (parallel fan-out, run_in_background=true). Don't simulate panels.
| Signal | Agents (parallel unless noted) | Token budget |
|---|
| IGNORE | none (skip) | 0 |
| ADOPT | polyglot-code-reviewer | ~10K |
| INTEGRATE | polyglot-code-reviewer + security-auditor | ~20K |
| INSPIRE | innovator + chief-product-officer + fact-checker-investigator | ~40K |
| FAIL-FAST SPIKE | innovator + codex-reasoner (codex designs cheapest kill-test) | ~20K |
| RESEARCH | market-tech-researcher + codebase-intelligence-researcher + data-detective | ~40K |
| SUPERCHARGE | SERIAL innovator (kill/proceed gate) → if proceed: parallel moonshot-architect + business-panel-experts + research patent-prior-art | ~80K |
| REPOSITION | business-panel-experts + brand-customer-experience-lead + growth-hacker | ~40K |
Domain specialist additions (detect via URL extension / arxiv category / page keywords):
- Rust → rust-excellence-engineer
- Mojo / ML kernels → mojo-ml-architect
- iOS / Android → mobile-native-expert
- Distributed systems claims → distributed-systems-architect
- macOS SwiftUI → macos-swiftui-coder
- Tests / QA → comprehensive-test-engineer
Each agent receives: URL + fetched content + cross-map repos + initial claim + credibility tier + hebb-known history.
Each agent returns (normalized JSON, enforced via prompt template):
{
"agent": "AGENT-NAME",
"verdict": "proceed | kill | iterate",
"finding": "one sentence",
"question": "one risk / open item",
"action": "one concrete next step",
"confidence": 0.0
}
Synthesis: majority verdict; ties broken by innovator/ceo/devil. Compose issue body with all panel quotes verbatim under collapsible sections.
6.5 · Devil + Steelman (mandatory for every non-IGNORE)
Run as TWO distinct sections in one agent call (mode-switch, don't collapse). Devil alone biases status-quo (incumbents always find 3 reasons to dismiss disruption: x86 → ARM, MySpace → FB, Blockbuster → Netflix). Steelman alone biases FOMO. Together they calibrate.
Devil's advocate:
- 3 reasons NOT to adopt (concrete, not generic).
- Cheapest way this could be wrong (<1h evidence to kill it).
- Who tried this and failed (research skill: arxiv negative results, semantic_scholar forward citations, HN criticism search).
Steelman ("what if they're right?"):
- If we started today, would we build differently because of this?
- What 3 assumptions of OURS would have to be FALSE for them to be right?
- What evidence are we discarding to keep our current view?
- Who on our team has stake in dismissing this? (NIH / sunk-cost defender?)
- Is this the Christensen pattern — disruptive toy dismissed by incumbent until too late?
- Cost of ignoring IF they're right — asymmetric downside?
Reconciliation table:
| Devil | Steelman | Net action |
|---|
| kill | kill | IGNORE — both agree it's noise |
| kill | proceed | HUMAN-REVIEW tag — NIH bias risk, highest-alpha |
| proceed | proceed | signal stands, confidence +0.2 |
| proceed | kill | signal stands but drop one tier (weak novelty) |
| iterate | iterate | force FAIL-FAST SPIKE (experiment resolves) |
| any other mismatch | | add mismatch-review tag, keep original signal, flag in report |
Mismatch cases are where REAL alpha lives. Most people only do devil → confirmation bias for status quo. Surface mismatches explicitly.
7 · DoR-compliant issue template
Pre-fills the 3 DoR signals (testable AC, priority/ROI, reversibility). Caveat (I): gateway-wrapped Linear calls do not fire dor-issue-guard.py (linear skill §4). Pre-fill is self-discipline.
Template (placeholders marked with __FIELD__ to avoid stub-detector false positives):
## ROI: **NX** | Signal: **LABEL** | Cross-repo target(s): **REPOS**
P**N** — value/cost: $**VAL**/yr / $**COST** → **RATIO**x · Credibility: **TIER**
### Source
- URL: **URL**
- Wayback URL: **WAYBACK_URL_OR_ATTEMPTED_FAILED**
- Author/Org: **AUTHOR_ORG**
- Published: **DATE**
- Fetched: **TODAY**
- Staleness: fresh / DATED Nmo (verify SOTA)
### Main claim
**CLAIM_ONE_SENTENCE**
### Why it matters for us
**VISION_MAPPING** (cite CLAUDE.md line/section if v2)
### Acceptance Criteria (testable)
- [ ] **OUTCOME_1**
- [ ] **OUTCOME_2**
- [ ] **OUTCOME_3**
### Agent panel (§6 — verbatim returns)
<details><summary>innovator</summary>__JSON_BLOCK__</details>
<details><summary>chief-product-officer</summary>__JSON_BLOCK__</details>
<details><summary>fact-checker-investigator</summary>__JSON_BLOCK__</details>
### Devil's advocate (§6.5)
**3 kill-reasons**:
1. **REASON_1**
2. **REASON_2**
3. **REASON_3**
**Cheapest way wrong**: **ONE_LINE_REFUTATION**
**Prior failures**: **ARXIV_OR_HN_LINK_OR_NONE**
**Verdict**: keep / downgrade / IGNORE
### Steelman ("what if they're right?")
**3 assumptions of ours that'd have to be false**:
1. **ASSUMPTION_1**
2. **ASSUMPTION_2**
3. **ASSUMPTION_3**
**Evidence we'd have to discard**: **ITEMS**
**Cost of ignoring if right**: $**VAL**/yr asymmetric
**Christensen-check**: disruptive-for-incumbents? yes/no — **WHY**
**Verdict**: kill / proceed / iterate
### Reconciliation
- Devil verdict: **V1**
- Steelman verdict: **V2**
- Net action (per §6.5 table): **ACTION**
- Confidence: **PCT**%
### Reversibility / Rollback
yes/no + how — revert commit / feature flag off / N/A because **REASON**
### ROI breakdown
- value: $**VAL**/yr (**JUSTIFICATION**)
- probability: **PCT**% (credibility-adjusted: base × **FACTOR**)
- cost: **K** tokens + **HOURS**h
- score: **NX**
### Approach
adopt-directly / integrate-as-dep / build-superior / fail-fast-spike / empirical-research / reposition
### Next step
**SINGLE_CONCRETE_ACTION**
8 · Filing — delegate to linear skill
See ~/.claude/skills/linear/SKILL.md for full pattern. Summary:
import json, pathlib
cache = json.loads((pathlib.Path.home()/".claude/data/linear/teams-cache.json").read_text())
mik = cache["teams"]["MIK"]
prio_map = {"P0": 1, "P1": 2, "P2": 3, "P3": 4}
label_ids = [mik["labels"][l] for l in label_keys if l in mik["labels"]]
arguments = {
"teamId": mik["id"],
"title": f"[ROI:{roi}x] [{signal}] {claim_summary}",
"description": dor_body,
"priority": prio_map[priority],
"stateId": mik["states"]["Backlog"]["id"],
"labelIds": label_ids,
"createAsUser": "Claude Elite",
"displayIconUrl": "https://www.anthropic.com/images/icons/safari-pinned-tab.svg",
}
For SUPERCHARGE with no prior art found → also add P0-patent label + comment tagging /!:innovate and /!:discover.
9 · Report format (to user)
Sorted: SUPERCHARGE > REPOSITION > INSPIRE > INTEGRATE > ADOPT > FAIL-FAST SPIKE > RESEARCH > IGNORE.
## url-insight run — N URLs · TIMESTAMP
| # | URL (short) | Signal | Target repo(s) | Cred | ROI | Devil | Steelman | Mismatch? | Issue | Notes |
| --- | ---------------- | ------- | -------------- | ---- | --- | ------- | -------- | --------- | --------- | ---------------------- |
| 1 | arxiv/2507.21474 | INSPIRE | hebb | C | 30x | proceed | proceed | no | (dry-run) | trace-viz narrow scope |
| ... |
### Discarded (IGNORE)
- URL — 1-line reason
### Mismatch / Human-review queue (REVIEW THESE)
- URL — devil=kill steelman=proceed → potential NIH bias, decide manually
### Artifacts
- JSON: ~/.claude/data/url-insight/run-TS.json
- Queue: ~/.claude/data/innovations/queue/HASH.json × N
- Memories: hebb × N (project=url-insight)
- Wayback: archived proof URL per non-IGNORE URL, or non-fatal capture-failure marker
10 · Innovation-tracker queue schema
Emit per actionable insight to ~/.claude/data/innovations/queue/URL_HASH.json:
{
"source": "url-insight",
"url": "FULL_URL",
"wayback_url": "WAYBACK_URL / null",
"wayback_save_failed": false,
"url_hash": "SHA256_PREFIX12",
"signal": "LABEL",
"cross_map": ["repo_a", "repo_b"],
"evidence_url": "PRIMARY_SOURCE_URL",
"claim": "MAIN_CLAIM",
"credibility": "S/A/B/C/D",
"roi": 0,
"priority": "P0/P1/P2/P3",
"linear_issue": "MIK-NNN / null",
"devils_advocate": {
"kill_reasons": ["r1", "r2", "r3"],
"cheapest_wrong": "ONE_LINE",
"prior_failures": "URL / NONE",
"verdict": "keep / downgrade / ignore"
},
"steelman": {
"false_assumptions": ["a1", "a2", "a3"],
"discarded_evidence": "ITEMS",
"asymmetric_cost": 0,
"christensen": "yes/no",
"verdict": "kill / proceed / iterate"
},
"reconciliation": {
"net_action": "STRING",
"confidence": 0.0,
"mismatch": false
},
"agents": [
{ "agent": "NAME", "verdict": "V", "finding": "...", "action": "..." }
],
"timestamp": "RFC3339"
}
Matches innovation-tracker.sh schema (see ~/.claude/hooks/PostToolUse/innovation-tracker.sh). /!:innovate rescores alongside git-commit innovations.
11 · Worked example — arxiv/2507.21474 (V, fetched 2026-04-24)
Input: https://arxiv.org/abs/2507.21474
Dedup: hebb.recall("2507.21474", project="url-insight") returned 0 hits — first analysis. Proceed.
Fetch: nab fetch https://arxiv.org/abs/2507.21474 returned 47,705 bytes in 677ms (cookies auto-loaded). Status 200.
Wayback capture: fulcrum:wayback_save with 7d idempotence and screenshot capture returns a timestamp URL or a pending job. The run JSON records wayback_url when available, otherwise wayback_save_failed=true plus a sanitized reason. This does not block the rest of the analysis.
Classify: paper (arxiv.org).
Route: research skill → semantic_scholar citation count, openalex author h-index.
Extract:
- Claim: ENN (Engram Neural Network) is an RNN with an explicit differentiable Hebbian memory matrix + sparse attention-driven retrieval.
- Differentiator: interpretability via observable memory dynamics (trace visualizations); accuracy comparable to GRU/LSTM on MNIST/CIFAR-10/WikiText-103.
- Numbers: 3 benchmarks; converges to similar accuracy/perplexity as baselines (no SOTA claim).
- Date: 2025-07-29 (9mo old, fresh).
- Surprise: LOW on idea (Hebbian+NN is classic territory); MEDIUM on execution (explicit differentiable memory + sparse retrieval + transparency angle in one architecture).
Credibility: Daniel Szelogowski — solo author, no org listed on abstract page, arxiv-only venue → tier C (factor 0.4).
Cross-map:
- v2 (hebb/CLAUDE.md exists): keywords {memory, hebbian, engram, recall, decay, plasticity, neuroscience, trace} overlap strongly → top match is hebb.
- v1 candidates (weaker): botnaut-server (neural inference), metacognition (reasoning).
- Verified: hebb's CLAUDE.md line 5 reads "world-class memory system built to the quality bar" + "Hebbian synaptic plasticity, surprisal gating, reconsolidation".
Initial signal: between INSPIRE and SUPERCHARGE.
- hebb already has BGE-M3 + HNSW + BM25 + RRF, thermodynamic decay, surprisal gating — far more sophisticated than ENN's memory matrix.
- ENN's training-time differentiable memory + observable trace visualizations is the novel piece hebb doesn't have.
- Not SUPERCHARGE because ENN's core mechanism is weaker than hebb's existing stack — no 10x moat.
- Initial classification: INSPIRE (build superior version of trace-viz angle).
Agent dispatch (INSPIRE budget ~40K)
Three agents in parallel, each receives URL + fetched abstract + cross-map=[hebb] + credibility=C:
innovator -> "proceed", finding: "trace-viz interpretability is net-new for hebb, narrow scope means low-cost spike", action: "branch hebb-trace-viz, render activation heatmap"
chief-product-officer -> "proceed", finding: "interpretability is a stated hebb moat — visualizing recall decisions strengthens the I1 trust-chain story", action: "ship as feature behind --debug flag"
fact-checker-investigator -> "iterate", finding: "no peer review, no GitHub link, 0 citations checked via semantic_scholar — high reproducibility risk", action: "verify ENN claims with PyTorch reimpl before adopting any architecture"
Synthesis: 2 proceed + 1 iterate → proceed with iterate-modifier (narrow scope further).
Devil's advocate
- Solo-author arxiv-only paper, no peer review, no GitHub code linked — reproducibility risk HIGH.
- "Comparable to RNN/GRU/LSTM" is not a SOTA win. 2026 baselines are Transformers/Mamba/Hyena — ENN isn't benchmarked against them.
- Training-time Hebbian memory is not hebb's inference-time architecture — different problem class, trace-viz concept may not transfer.
Cheapest wrong: 1h semantic_scholar check — if 0 citations 9mo post-submission, signal dies.
Prior failures: Hebbian-in-NN graveyard (Hopfield nets, Kanerva memory, DNC). Must explain how ENN avoids DNC's read-head-inspection-nobody-uses failure mode.
Devil verdict: proceed (no kill-class evidence; risks are about scope, not feasibility).
Steelman ("what if they're right?")
- 3 assumptions of ours that'd have to be false:
- "hebb's BGE-M3 + HNSW retrieval is sufficient for interpretability" — false if users can't trace WHY a memory ranked top-3.
- "Surprisal-gate logs are enough debugging telemetry" — false if a visual heatmap surfaces patterns logs hide.
- "Inference-time vs training-time memory are unrelated" — false if ENN's trace mechanism transfers as a debugging UI primitive regardless of when memory was learned.
- Evidence we'd discard: hebb users repeatedly ask "why did this memory rank?" in issues/Slack — log-based answers are unsatisfying.
- Asymmetric cost if right: ~$5K/yr in user trust + a sticky differentiator vs Mem0/MemGPT competitors.
- Christensen-check: YES — interpretability is the kind of "toy feature" incumbents (vector DB vendors) dismiss until users defect to the visible-thinking alternative.
Steelman verdict: proceed.
Reconciliation
- Devil = proceed, Steelman = proceed → signal stands, confidence +0.2 (INSPIRE confirmed, no mismatch).
- Net action: INSPIRE — narrow-scope trace-viz spike.
- Confidence: 70%.
ROI
- value: $5K/yr (hebb recall debugging UX, interpretability feature, Christensen-defense).
- probability: 30% × 0.4 (credibility C) = 12% — but +0.2 from clean reconciliation → about 25%.
- cost: ~2K tokens (fetched) + 4h dev for spike.
- score: ($5000 × 0.25) / ~$40 compute equiv ≈ 30x → P2.
Composed issue body (would be filed, not actually sent)
## ROI: 30x | Signal: INSPIRE | Cross-repo target(s): hebb
P2 — value/cost: $5000/yr / $40 → 30x · Credibility: C
### Source
- URL: https://arxiv.org/abs/2507.21474
- Wayback URL: WAYBACK_URL_OR_ATTEMPTED_FAILED
- Author/Org: Daniel Szelogowski (solo, no affiliation)
- Published: 2025-07-29
- Fetched: 2026-04-24
- Staleness: fresh (9mo)
### Main claim
Engram Neural Network (ENN): RNN variant with explicit differentiable Hebbian memory matrix + sparse attention retrieval, achieving comparable accuracy to GRU/LSTM on MNIST/CIFAR-10/WikiText-103 while offering observable memory dynamics (trace visualizations) for interpretability.
### Why it matters for us
hebb's vision (CLAUDE.md line 5) is a "world-class memory system built to the quality bar". hebb has superior retrieval (BGE-M3 + HNSW + BM25 + RRF) vs ENN's matrix, but lacks ENN's trace-visualization angle. The interpretability-via-observable-dynamics idea could become a hebb recall-debugging UI feature — net-new, not a retrofit of our memory core.
### Acceptance Criteria (testable)
- [ ] Render memory activation heatmap for a single hebb recall query on a sample WikiText prompt.
- [ ] Overlay surprisal-gate decisions on the heatmap (green=stored, red=suppressed).
- [ ] Produce 1 debugging insight not previously visible (e.g. "stored memory N had near-zero activation but ranked top-3 — why?").
### Agent panel (verbatim)
<details><summary>innovator</summary>{"verdict":"proceed","finding":"trace-viz interpretability is net-new for hebb, narrow scope means low-cost spike","action":"branch hebb-trace-viz, render activation heatmap","confidence":0.7}</details>
<details><summary>chief-product-officer</summary>{"verdict":"proceed","finding":"interpretability is a stated hebb moat, visualizing recall decisions strengthens the I1 trust-chain story","action":"ship as feature behind --debug flag","confidence":0.65}</details>
<details><summary>fact-checker-investigator</summary>{"verdict":"iterate","finding":"no peer review, no GitHub link, citation count not yet verified, high reproducibility risk","action":"verify ENN claims via reimplementation before adopting architecture","confidence":0.55}</details>
### Devil's advocate
1. Solo-author arxiv-only paper, no peer review, no GitHub linked — reproducibility risk HIGH.
2. "Comparable to RNN/GRU/LSTM" is not a SOTA win. 2026 baselines are Transformers/Mamba — ENN isn't benchmarked against them.
3. Training-time Hebbian memory is not hebb's inference-time memory — trace-viz concept may not transfer.
**Cheapest wrong**: 1h semantic_scholar check — if 0 citations 9mo post-submission, signal dies.
**Prior failures**: Hebbian-in-NN graveyard (Hopfield, Kanerva, DNC). Must explain how trace-viz beats DNC's read-head inspection (which nobody uses).
**Verdict**: proceed (risks are scope, not feasibility).
### Steelman ("what if they're right?")
**3 assumptions of ours that'd have to be false**:
1. hebb's BGE-M3 + HNSW retrieval is sufficient for interpretability.
2. Surprisal-gate logs are enough debugging telemetry.
3. Inference-time vs training-time memory are unrelated.
**Evidence discarded**: hebb users repeatedly ask "why did this memory rank?" — logs are unsatisfying.
**Cost of ignoring if right**: ~$5K/yr in user trust + sticky differentiator vs Mem0/MemGPT.
**Christensen-check**: YES — interpretability is a "toy feature" incumbents (vector DB vendors) dismiss until users defect.
**Verdict**: proceed.
### Reconciliation
- Devil verdict: proceed
- Steelman verdict: proceed
- Net action: signal stands, confidence +0.2 → INSPIRE confirmed, narrow-scope trace-viz spike.
- Confidence: 70%
### Reversibility / Rollback
Yes — spike on `hebb-trace-viz` branch. No merge if spike fails. Revert = delete branch. Zero prod impact.
### ROI breakdown
- value: $5K/yr (hebb recall debugging UX, Christensen-defense)
- probability: 25% (12% base × +0.2 reconciliation bonus)
- cost: ~2K tokens + 4h dev
- score: 30x
### Approach
fail-fast-spike — 1-day implementation on `hebb-trace-viz` branch.
### Next step
Branch `hebb-trace-viz` off main. Implement activation heatmap for `hebb recall "<sample query>"`. Eyeball debugging value. Kill / promote within 4h.
Report table (single-URL run)
| # | URL | Signal | Target | Cred | ROI | Devil | Steelman | Mismatch? | Issue | Notes |
| --- | ---------------- | ------- | ------ | ---- | --- | ------- | -------- | --------- | --------- | ---------------------- |
| 1 | arxiv/2507.21474 | INSPIRE | hebb | C | 30x | proceed | proceed | no | (dry-run) | trace-viz narrow scope |
Counterfactual SUPERCHARGE path (illustrative)
If credibility had been S (e.g. DeepMind authors, NeurIPS venue) AND the paper had introduced a memory primitive hebb lacks (not just trace-viz):
- Initial signal would have been SUPERCHARGE (their idea × hebb's existing strength = 10x moat).
- SERIAL dispatch: innovator first. If kill → drop to INSPIRE.
- If innovator proceeds → parallel fan-out: moonshot-architect + business-panel-experts + research-skill patent-prior-art search.
- If patent-prior-art returns empty → escalate to P0 [PATENT], tag
/!:innovate /!:discover.
- If prior art exists → downgrade to INSPIRE, name competitor.
Token budget would have been ~80K vs the ~40K spent here.
12 · Worked example 2 — product launch (illustrative, not fetched)
Input: https://www.producthunt.com/posts/mem0-long-term-memory-for-ai-agents
Classify: product-launch. Route: fetch comparison page if linked, credibility via ror on Mem0 Inc.
Cross-map: hebb (direct competitor, long-term memory for AI agents).
Initial signal: REPOSITION (hebb must update value prop vs Mem0 messaging). Possible secondary INSPIRE (what does Mem0 do that hebb doesn't?).
Agent dispatch (REPOSITION ~40K): business-panel-experts + brand-customer-experience-lead + growth-hacker.
Devil must include: "Is Mem0 actually better on LoCoMo? hebb leads LoCoMo #1 per CLAUDE.md — if Mem0's numbers are worse, this is REPOSITION (lead with our win), not threat."
Steelman must include: "What if Mem0's distribution wins regardless of benchmark? Worse-but-marketed beats better-but-quiet (Christensen)."
13 · Anti-patterns (NEVER)
- NEVER use WebFetch — ~50K token waste per call (25× nab overhead). nab MCP preferred, nab CLI fallback.
- NEVER skip devil's-advocate AND steelman — both, every non-IGNORE. Devil alone = status-quo bias. Steelman alone = FOMO bias. Together = calibrated.
- NEVER auto-file P3 issues without user confirmation — surface in report, await "yes file it".
- NEVER auto-file HUMAN-REVIEW reconciliation cases — these are the highest-alpha mismatches; user must decide.
- NEVER miss the SUPERCHARGE case — highest-value label AND easiest to under-classify. When a URL maps to a repo you already lead in, ask: "their idea × our strength = 10x moat?" If yes → SUPERCHARGE → P0.
- NEVER skip hebb dedup — analyzing same URL twice in a week is a protocol failure.
- NEVER hardcode Linear UUIDs — always read
~/.claude/data/linear/teams-cache.json (linear skill §3).
- NEVER assume S-tier credibility without evidence — default C when unresolvable.
- NEVER pre-fill "to be determined" in DoR sections — if you can't answer, downgrade signal until you can.
- NEVER let IGNORE silently leak into actionable — IGNORE terminates the pipeline, period. No agent dispatch, no cross-map, no issue, no queue entry. One line in the report.
- NEVER simulate panel personas (CTO/CPO/CMO) inline — DISPATCH to real agents via Agent tool. Simulated personas add tokens without independent perspectives.
- NEVER skip innovator's serial-gate for SUPERCHARGE — kill-verdict at gate saves the ~60K parallel panel cost.
- NEVER auto-spike without user consent unless
--auto-spike explicitly present.
- NEVER exceed 50 URL or 2M token batch cap — hard halt, surface partial results.
14 · Cross-refs
| Skill / hook / MCP | Why |
|---|
~/.claude/skills/linear/SKILL.md | Issue filing; DoR template; Linear priority inversion gotcha |
~/.claude/skills/research/SKILL.md | arxiv / semantic_scholar / openalex / patent prior-art |
~/.claude/skills/supercharge/SKILL.md | Invoked when signal=SUPERCHARGE — runs the 11-phase synthesis pipeline (DECOMPOSE → ... → TELEMETRY) instead of file-and-forget |
~/.claude/skills/wayback/SKILL.md | Source-of-record preservation; URL triage must call fulcrum:wayback_save after successful fetch |
~/.claude/skills/apple-calendar/SKILL.md | If URL implies a calendar event ("I'm attending X") — rare |
~/.claude/skills/ctx-mgmt/SKILL.md | Session snapshot after large runs |
nab MCP / CLI | URL fetch (PREFER over WebFetch) |
mcp__hebb__{recall,remember} | Dedup + cross-session insight graph |
Agent tool | Real specialist dispatch (replaces simulated personas) |
~/.claude/hooks/PostToolUse/innovation-tracker.sh | Downstream consumer of queue entries |
~/.claude/hooks/PreToolUse/dor-issue-guard.py | DoR gate (does NOT fire on gateway-wrapped Linear calls) |
/!:innovate /!:discover | Patent-pipeline escalation for SUPERCHARGE + no-prior-art |
15 · Files
| Path | Purpose |
|---|
~/.claude/skills/url-insight/SKILL.md | this file |
~/.claude/skills/url-insight/session-learnings-2026-05-02.md | session learnings: portfolio facts, hook gotchas, fan-out preamble — MUST read before fresh-session run |
~/.claude/data/url-insight/portfolio.json | cached gh repo list (weekly refresh) |
~/.claude/data/url-insight/run-TIMESTAMP.json | per-run JSON artifact |
~/.claude/data/innovations/queue/HASH.json | innovation-tracker queue entry |
url-insight v2026.05.05 (adds §1.1 Wayback evidence capture for MIK-3354) + session-learnings-2026-05-02 — 8 signals, real-agent panels, devil + steelman calibrated, 0 hype, Rams #10. Paste URLs → ranked actions in one pass. Fresh-session runs MUST read session-learnings-2026-05-02.md before classification (portfolio facts, hook gotchas, fan-out preamble). §4.1 BLOCKING: query Linear active tickets before classifying as IGNORE. §1.1 BLOCKING for non-IGNORE evidence: record wayback_url or wayback_save_failed.
