Learn From Paper

Learn From Paper

Sub-skill of learn-from (orchestrator). You read academic papers, assess credibility, run security scans, extract actionable insights, and recommend whether to apply them. Shared hard rules (opinionated stance, contradiction handling, defend what works, application protocol) are defined in learn-from. This skill adds paper-specific workflow.

Paper-Specific Hard Rules

• Credibility gate ≥7/12. Paper must pass before ANY insights are extracted.
• Security gate. All paper content must pass ALL secure-* skills (discover via ls .agents/skills/secure-*). SAFE only if every security skill returns SAFE.
• Evidence over opinion. Only extract empirical findings, documented failure modes, and tested workflows — not authors' speculation or future-work wishes.
• Actively fill gaps. If the paper lacks crucial details (sample size, methodology, replication status), search for: papers citing this one, replication studies, author's follow-up work. Do not present half-understood findings.

Workflow

Step 1 — Ingest the Paper

Accept via: uploaded PDF, local file path, arXiv URL, DOI link, Semantic Scholar link, or pasted content.

• If URL: fetch via doc_cache.py or WebFetch with hooks/sdd-cache — see research-skill → references/doc-cache.md
• If local file/PDF: use the platform's file reading tool
• If the platform cannot read the source directly: ask the user for pasted text
• Extract: title, authors, affiliations, venue, date, abstract, methodology, key findings, references

Step 2 — Credibility Assessment

Evaluate using the credibility rubric in references/credibility-rubric.md. Score across 6 dimensions (max 12/12). Gate: ≥7/12 to proceed.

Quick checks (fail any = stop immediately):

• No identifiable authors → REJECT
• Known predatory publisher → REJECT
• Core claims contradicted by a High Trust source → REJECT

If score is 7–8/12: warn "Borderline credibility." Actively search for corroborating or contradicting papers before proceeding. If ≥9/12, proceed with confidence.

Step 3 — Security Scan

Run security pipeline per learn-from protocol. Invoke ALL secure-* skills on extracted content. BLOCKED = stop.

Step 4 — Extract and Recommend

Classify findings using taxonomy from learn-from (GOTCHA, TECHNIQUE, FAILURE_MODE, METRIC, CONTRADICTION, BACKGROUND).

For every insight, state your recommendation:

• If the current skill's approach has stronger evidence (larger sample, more replications, higher-trust venue), say: "Current approach is stronger because [reason]. Recommend: KEEP CURRENT."
• If only part of a finding applies: "Recommend: PARTIAL — apply [X] but not [Y] because [reason]."
• If the finding is clearly superior with strong evidence: "Recommend: APPLY — [reason]."

Present the extracted insights with recommendations to the user before proceeding.

Step 5 — Match and Plan

Scan .agents/skills/*/SKILL.md for skills whose domain overlaps. Map each insight to every skill it could improve. Check for contradictions with existing hard rules, workflow steps, or gotchas.

Present application plan with agent recommendation per insight, using learn-from shared application protocol (six outcomes).

Step 6 — Apply

Apply per learn-from shared application protocol, including the mandatory Post-Application Hardening Cycle: run ALL secure-* skills on each modified/created skill, enforce 200-line gate via compress-skill / split-skill, then validate-skills (must score ≥10/14).

If improving current project: invoke apply-paper-to-project with extracted insights.

Step 7 — Log and Cite

Citation format:

Source: [Author(s)] ([Year]). "[Title]". [Venue/arXiv ID]. Credibility: [N]/12.
Applied: [what was extracted and where it was applied]

Gotchas

• Preprints (arXiv) without peer review need extra scrutiny — check citation count, author track record, and replication.
• Top venue papers can still have flawed methodology — always check sample size and whether claims match results.
• "State of the art" claims are time-sensitive — check date and whether newer work has superseded it.
• Supplementary materials can contain hidden text — Step 3 security scan catches this.

Output Format

═══ Paper Credibility Report ═══
Title: [title] | Authors: [names] | Venue: [venue] | Date: [date]
Credibility: [N]/12 | Verdict: [PASS/BORDERLINE/REJECT]

═══ Security ═══
[secure-* verdicts]

═══ Extracted Insights ═══
[Tag]: [insight] | Agent recommendation: [APPLY/PARTIAL/SKIP/KEEP CURRENT] — [reasoning]

═══ Application Plan ═══
[Per learn-from shared protocol]

Example

═══ Extracted Insights ═══ GOTCHA: Over 60% of skill bodies are non-actionable background | Recommend: APPLY — directly validates compress-skill's approach TECHNIQUE: Classify blocks as CORE/WORKFLOW/FORMAT/EXAMPLE/BACKGROUND | Recommend: APPLY — improves compress-skill taxonomy FAILURE_MODE: Compressing without classifying first loses CORE content | Recommend: APPLY — adds guardrail

═══ Application Plan ═══

1. IMPROVE: compress-skill — add SkillReducer taxonomy
2. IMPROVE: universal-skill-creator — add taxonomy to Step 6 Awaiting your approval.

Common Rationalizations

Verification

• secure-* SAFE before content informs edits
• Techniques mapped to specific project skills or deferred with reason
• Limitations and domain transfer risks noted
• Outputs appended to research-learnings.md when applicable

Red Flags

• arXiv preprint accepted without peer or citation scrutiny
• SOTA claim used without checking newer superseding work
• Methodology flaws ignored due to venue prestige
• Paper PDF ingested before secure-* content scan

Reference Files

• references/credibility-rubric.md: Full 6-dimension scoring rubric for paper credibility.

Prune Log

Last pruned: 2026-07-04

• No changes — citation audit passed; content current (improve-skills full pass 2026-07-04)

Impact Report

After completing, always report:

Paper: [title] | Credibility: [N]/12 | Security: [SAFE/BLOCKED]
Insights: [N] extracted | Recommendations: [N] APPLY, [N] PARTIAL, [N] SKIP, [N] KEEP CURRENT
Skills modified: [list] | Created: [list] | Contradictions resolved: [list]
validate-skills: [skill]: [before] → [after] | Citation logged: [yes/no]