Agent Skills: Research-to-Essay Skill

Research-to-Essay Skill

Systematic workflow for producing publication-grade essays from research. Handles multi-source synthesis, narrative construction, voice calibration, and citation management.

Citation Fidelity Firewall (HARD GATE — never violate)

Citing is reporting what a real source actually said. Writing persuasively is not the same as inventing the evidence. Every citation, statistic, study name, author-year, institution, direct quote, finding, and URL in the essay MUST come from a web_search or web_fetch you actually executed in this session and whose result you saw with your own eyes. A fluent sentence resting on a fabricated source is a fabrication, not an essay — no matter how plausible it reads or how perfectly it fits the argument. This is the same failure class as fabricating an author's anecdote during prose polishing: persuasive surface, invented substance.

The hard rule: no retrieved source → no citation. If you did not run a search/fetch that returned a given source, you may not cite it, name it, link to it, or attribute a claim to it. Before writing any citation, you must be able to point to the specific tool call in this session whose output contains that exact source. If you cannot, the citation does not exist.

You must STOP — not improvise — when research is thin. If research was never run, or ran but returned too little to support the essay, you have exactly these honest moves, in order:

1. Run the searches now — execute the web_search/web_fetch calls and cite only what comes back.
2. Tell the user you cannot cite responsibly — say plainly that the available verified sources are insufficient, and ask whether to research further or proceed with explicitly uncited, clearly-hedged claims.
3. Reduce the claim to match the evidence — soften or cut any assertion you cannot ground in a retrieved source, rather than manufacturing support for it.

You may NEVER manufacture a citation, author name, author-year, institution, study title, statistic, finding, or link to satisfy a structural target. The "5-8 sources" / "8-12 sources" / "6-10 sources" minimums (see Step 2) are a floor on VERIFIED sources actually retrieved this session — not a quota to pad toward. If you have four real sources, you have four; invent a fifth and the whole essay is compromised. Falling short of the floor is a signal to research more or narrow the claim, never to fabricate.

Every URL must be one you actually retrieved. Do not construct, guess, autocomplete, or "reconstruct from memory" plausible-looking URLs (e.g. a stanford.edu/hai/... path that looks right). A URL you did not see returned by a tool is a fabricated URL even if the domain is real. When in doubt about whether you actually retrieved a source, treat it as not retrieved and remove the citation.

This gate outranks every stylistic, structural, and completeness goal in this skill. The "Evidence & citation check" in Step 5 and the "always cite empirical claims" rule in Step 4 presuppose this firewall: a citation only counts if it passed this gate first.

Final pass — provenance re-check: before delivering, walk every citation, statistic, quote, and URL in the essay and confirm each traces to a tool result you saw this session. Roll back any you cannot trace. This check outranks every other quality signal.

Core Workflow

1. Intake & Planning

Parse user request to determine:

• Format target: Substack (1500-3000w), LinkedIn (150-300w), Academic (3000-8000w), or Executive Brief (500-1000w)
• Topic & angle: What question/claim is central?
• Essay structure: Which arc fits? (Persuasive, Exploratory, Diagnostic, Narrative-Conceptual, Synthesis)
  • Consult references/essay-structures.md for detailed arc patterns
• Voice profile: Which register? (Poetic Rigor, Professional Signal, Scholarly Precision, Surgical Clarity)
  • Consult references/voice-profiles.md for characteristics and forbidden patterns

Output from this phase: Research plan with target structure and voice

2. Research Execution

Conduct systematic research following source credibility hierarchy:

Search strategy:

• Start with primary sources (research papers, official data, technical documentation)
• Layer in expert analysis (domain specialists, academic reviews, investigative journalism)
• Add informed commentary (practitioner Substacks, conference talks) for applied context
• Avoid weak sources (social media speculation, content marketing, AI-generated farms)

Source quality requirements: (these minimums are a floor on VERIFIED sources you actually retrieved this session via web_search/web_fetch — see the Citation Fidelity Firewall above. Never pad toward a number by inventing sources; falling short means research more or narrow the claim.)

• Minimum 5-8 sources for persuasive essays
• Minimum 8-12 sources for exploratory essays
• Minimum 6-10 sources for diagnostic essays
• Always include strongest counter-argument sources
• Prioritize recent sources for rapidly-changing topics, foundational sources for stable concepts

Citation extraction:

• Record: title, URL, author, date, credibility tier (1-4), key claims
• Use web_fetch to read full articles when web_search snippets insufficient
• For each source, extract 3-5 core claims explicitly
• Tag sources with themes for clustering

Consult references/research-patterns.md for:

• Source credibility hierarchy (Tiers 1-4)
• Research strategy by essay type
• Quality checks and anti-patterns

3. Synthesis

Organize research into thematic structure using one of two methods:

Method A: Manual thematic clustering (for simpler essays)

• Group claims by theme, not by source
• Identify convergent claims (multiple sources agree) → high confidence
• Identify divergent claims (sources disagree) → flag as tension
• Map claim dependencies (which claims require which others)

Method B: Script-assisted synthesis (for complex multi-source essays)

• Create JSON file with sources in required format (see script usage below)
• Run scripts/synthesize_sources.py <sources.json> <output.md>
• Review generated synthesis report showing themes, convergence, tensions

Script format:

[
  {
    "title": "Source Title",
    "url": "https://example.com",
    "source_type": "primary",
    "claims": ["Claim 1", "Claim 2"],
    "themes": ["theme1", "theme2"],
    "date": "2025-01-15",
    "credibility_tier": 1
  }
]

Synthesis output: Thematic map showing:

• Core themes with supporting sources
• Convergent evidence (agreement across sources)
• Divergent claims (tensions or debates)
• Gaps or under-supported areas

4. Drafting

Build essay iteratively using chosen structure template:

Template selection:

• Use assets/essay-template.md for Substack/long-form
• Use assets/linkedin-template.md for LinkedIn posts
• Adapt templates based on selected essay structure from Step 1

Drafting principles:

• Lead with strongest material: Hook in first paragraph, no throat-clearing
• Integrate sources naturally: Embed citations in argument flow, don't list separately
• Section logic: Each section should build necessarily on the previous
• Evidence before abstraction: Concrete examples, then pattern extraction
• Tension acknowledgment: Include counter-arguments and complications honestly
• Progressive depth: Can write full essay in one pass OR build iteratively:
  • Pass 1: Outline with section headers
  • Pass 2: Fill core argument sections
  • Pass 3: Add evidence and citations
  • Pass 4: Write intro/conclusion last

Voice application:

• Apply selected voice profile consistently (from Step 1)
• Check against forbidden patterns in references/voice-profiles.md
• Calibrate tone dimensions: warmth, certainty, abstraction, humor

Citation style:

• Substack/LinkedIn: Inline hyperlinks on key phrases, footnotes for tangential details
• Academic: Numbered footnotes/endnotes with full bibliography
• Executive: Minimal citation, only for key data points
• Always cite: empirical claims, direct quotes, novel frameworks, counter-intuitive findings
• Never cite: common knowledge, your own synthesis, widely-known facts

5. Refinement

Quality assurance checks before delivery:

Structural review:

• [ ] Hook is genuinely compelling (test: would you click "read more"?)
• [ ] Stakes are established early (why should reader care?)
• [ ] Each section advances the argument necessarily
• [ ] Conclusion reframes rather than summarizes
• [ ] Length appropriate to format (Substack: 1500-3000w, LinkedIn: 150-300w)

Voice & style check:

• [ ] Run prose-polish skill on draft
• [ ] Check for forbidden patterns in selected voice profile
• [ ] Verify tone consistency throughout
• [ ] Confirm readability for target audience

Evidence & citation check: (presupposes the Citation Fidelity Firewall — a citation only counts if every author, statistic, quote, and URL traces to a web_search/web_fetch result you saw this session)

• [ ] Provenance verified: every citation, statistic, study name, quote, and URL traces to a tool call you actually executed this session — none manufactured, none reconstructed from memory
• [ ] Every major claim has warrant (evidence or citation)
• [ ] Primary sources used for factual claims
• [ ] Counter-arguments acknowledged with credible sources
• [ ] No citation decay (secondary sources when primary available)
• [ ] Links functional and actually retrieved (not plausible-looking URLs you constructed), citations complete

Platform-specific polish:

• LinkedIn: Paragraph breaks every 2-3 sentences, key phrases bolded, CTA included
• Substack: Section transitions smooth, footnotes formatted, metadata complete
• Academic: All citations complete, methodology transparent, limitations noted

6. Delivery

Present final essay as artifact with metadata:

Include:

• Complete essay in appropriate markdown format
• Word count and target audience notation
• Source list with tiers noted
• Key frameworks or concepts referenced
• Research date and any time-sensitivity notes

Optional additions based on context:

• Alternative versions for different platforms (e.g., Substack long-form + LinkedIn teaser)
• "Further Reading" section organized by theme
• Open questions or research gaps identified
• Suggested images or visual elements

When to Use References

Load these files as needed:

• references/voice-profiles.md — When clarifying voice characteristics or checking against forbidden patterns
• references/essay-structures.md — When uncertain about narrative arc or need structure template
• references/research-patterns.md — When evaluating source quality, planning research strategy, or checking synthesis methodology

Load scripts when:

• scripts/synthesize_sources.py — When dealing with 8+ sources requiring systematic thematic clustering

Quality Signals

High-quality output:

• Opens with genuine insight, not preamble
• Every paragraph necessary, no filler
• Sources integrated into argument, not appended
• Counter-arguments acknowledged, not buried
• Conclusion offers new lens, not recap
• Voice consistent and appropriate to format
• Citations complete and properly tiered
• Length justified by complexity, not padding

Red flags:

• Generic opening ("In today's world...")
• List structure when narrative needed
• No acknowledgment of complexity or tradeoffs
• All sources from same perspective
• Summary conclusion
• Inconsistent tone or register shifts
• Weak or missing citations for key claims
• Excessive length without proportional depth

Iteration Protocol

After delivering draft, typical refinement requests:

• "Make this more [voice]" → Reload references/voice-profiles.md and adjust tone calibration
• "Add more evidence for X" → Return to research phase for specific claim
• "This section feels weak" → Restructure using references/essay-structures.md patterns
• "Too long / too short" → Audit for filler vs. density, adjust scope
• "Challenge this argument" → Load strongest counter-sources, revise tensions section

Anti-Patterns to Avoid

• Don't search once and write—iterate research based on draft gaps
• Don't list sources separately from argument—integrate naturally
• Don't write intro first—write it last after you know what you said
• Don't ignore voice profile constraints—they prevent AI slop
• Don't cite weak sources when primary available—tier matters
• Don't EVER manufacture a citation, author, statistic, or URL to hit a source count—an invented Tier-1-looking source is the worst failure this skill can produce (see Citation Fidelity Firewall)
• Don't pad length artificially—every paragraph must earn its keep
• Don't summarize in conclusion—reframe or extrapolate instead