drafting-with-grounded-sources

Drafting with grounded sources

When to invoke

• Defending a specific claim with paragraph-level evidence from the library
• Developing or extending a thesis statement with cited support from one or more sources
• Editorial pass on a draft section with flagged unsourced claims
• Engaging with a paper's argument by drawing out its specific commitments and supporting them with quotes
• Explaining a methodology or finding attributed to a specific paper, where the explanation is itself a defended claim
• Building a research note that synthesizes claims from 3 or more sources, where the user has supplied the claims to defend

When NOT to invoke (use using-local-library-mcp instead)

• "Give an overview of approaches to X"
• "Summarize what the literature says about Y"
• "Tour what these sources contain on Z"
• "Explain [topic] drawing on @A and @B" (when no specific claim is named)
• General exploration where the user wants to discover what's in the corpus, not defend a position
• Literature summaries whose goal is breadth-of-coverage rather than support-for-an-argument

The orientation skill produces well-quoted multi-source prose with inline @citekeys and chunk references. It does this without requiring claim-decomposition — appropriate when the user has questions, not theses.

Iron law

• Quote into scratch before paraphrasing. Cannot quote = not grounded = not citable.
• One proposition, one grounding pass. "A and B" needs A grounded and B grounded separately.

Procedure

Step 1 — Determine input shape and decompose

This SKILL drafts prose that defends specific claims with cited evidence. The audit-trail appendix earns its keep only when discrete claims anchor the prose. Three input shapes to handle differently:

Shape A — claims supplied (e.g., "defend the claim that X using @A and @B"; "develop the argument that Y, drawing on @A"; "extend this thesis: [thesis] with evidence from @A").

List the supplied propositions. Each becomes one row in the grounding-summary appendix. Proceed to Step 2.

Worked example: "Vaswani et al. 2017 introduced the Transformer architecture, which replaced recurrence with attention and reduced sequential compute" decomposes to:

1. @Vaswani2017 introduced the Transformer architecture.
2. @Vaswani2017's Transformer replaced recurrence with attention.
3. @Vaswani2017 claims this reduces sequential compute.

Shape B — claims implicit, structure is argument-shaped (e.g., "extend this argument with evidence from the library", "draft a section arguing [position] from these sources", "make the case for X using @A").

The prompt asks for argument-defense but doesn't fully specify the propositions. Surface 2–4 inferred propositions for user review BEFORE invoking the kernel:

Your prompt asks me to argue for [implied position] drawing on @A and @B but doesn't fully specify the propositions. Based on the argument structure, I'd decompose into:

1. [inferred proposition 1]
2. [inferred proposition 2]
3. [inferred proposition 3]

Confirm, refine, or replace before I ground them.

Wait for user response before proceeding. Inferred propositions reflect training-derived priors about the sources, not user intent — surfacing them keeps the user in the authorship loop.

Shape C — topic exploration, no argument structure (e.g., "give an overview of approaches to X", "summarize what these papers say about Y", "explain [topic] drawing on @A and @B" without a stated claim, "tour the literature on Z").

This SKILL is the wrong tool. Return control with:

This prompt looks like exploration ("[the trigger phrase]") rather than argument-defense. drafting-with-grounded-sources is for the latter; its audit-trail appendix earns its keep only when prose defends discrete claims. Two paths:

1. If you have specific claims in mind, supply them (1–3 propositions) and I'll ground them via this SKILL.
2. If you want a tour of what these sources contain on the topic, the using-local-library-mcp orientation skill produces well-quoted multi-source prose without requiring claim-decomposition.

Which do you want?

Wait for user response. Do not silently treat exploration as argument-defense — that produces a paragraph grounded against propositions you invented from training memory, with hidden authorship.

Step 2 — Ground each proposition

Per proposition: invoke grounding-against-library (six-step procedure). Capture per-source quoted evidence into a scratch buffer.

If the kernel returns not-found for a proposition: drop it OR rewrite to match what the corpus supports. Do not paraphrase past not-found.

Step 3 — Write the prose

Draft from the scratch buffer. Inline @citekeys at every claim. Paraphrase from quoted excerpts, never from memory.

Step 4 — Build the grounding-summary appendix

At the end of the draft:

## Grounding summary

- Proposition 1: <verbatim>
  - Source: @Citekey1 (Chunk N)
  - Quote: "..."
  - Status: support

- Proposition 2: <verbatim>
  - Source: @Citekey2 (Chunk M)
  - Quote: "..."
  - Status: partial — caveat noted in prose

The appendix scales with draft size. For long drafts, invoke section-by-section to keep the appendix manageable.

Output format

• Draft prose with inline @citekeys at every claim
• Grounding-summary appendix at end (one entry per proposition, with source citekey, chunk index, quoted excerpt, status)

Failure-mode table

Excuse	Reality
"I'll cite this from memory; the paper definitely says it"	Memory is where false grounding enters. Quote first; cite from quote.
"Top result was thematically related; that's enough to cite"	Citation-shopping. The chunk must support the specific claim.
"mentions / discusses / argues mean the same thing"	They don't. Match the verb to the quote.
"I'll add the citekey at the end and figure out the chunk later"	Citations decay. Capture the chunk index when you capture the quote.
"This proposition's a connector — it doesn't need grounding"	Connectors quietly assert relationships. Ground them or mark as synthesis.
"The user gave me a topic and citekeys, so I'll synthesize a paragraph from those sources"	If the prompt is exploration (overview / summary / tour) rather than argument-defense, this SKILL is the wrong tool. Redirect to using-local-library-mcp. The audit-trail appendix earns its keep only when prose defends discrete claims.
"The user gave me an argument-shaped prompt without claims; I'll just decompose into propositions I expect the papers to make"	Inferred propositions reflect training-derived priors about the sources, not user intent. Surface them for review before grounding so the user can confirm or refine.

Red flags — stop

• About to add inline citekey without a corresponding grounding-summary entry
• About to skip decomposition because the section "is short"
• About to keep a not-found proposition by fudging the wording
• About to consolidate two propositions into one citation
• About to draft prose before the scratch buffer is populated
• About to draft an "overview of [topic]" or "summary of approaches to [X]" using this SKILL — that's exploration, not argument-defense; redirect to using-local-library-mcp
• About to invoke the kernel on propositions you generated from memory of the paper, without first asking the user to confirm or refine them on a topic-only or implicit-claims prompt