Citation Injector (LLM-first edits; budget-as-constraints)

Purpose: make the pipeline converge when the draft is:

• locally citation-dense but globally under-cited (too few unique keys), or
• overly reusing the same citations across many subsections.

This skill is intentionally LLM-first:

• you edit output/DRAFT.md using the budget report as constraints
• the helper script is validation-only (it never injects prose)

Inputs

• output/DRAFT.md
• output/CITATION_BUDGET_REPORT.md (from citation-diversifier)
• outline/outline.yml (H3 id/title mapping)
• citations/ref.bib (must contain every injected key)

Outputs

• output/DRAFT.md (updated in place)
• output/CITATION_INJECTION_REPORT.md (PASS/FAIL + what you changed)

Non-negotiables (NO NEW FACTS)

• Only inject keys listed for that H3 in the budget report.
• Do not introduce new numbers, new benchmarks, or superiority claims.
• Do not add narration templates (This subsection ..., Next, we ...).
• Do not produce cite dumps like [@a; @b; @c] as the only citations in a paragraph.

Paper-voice injection patterns (safe sentence shapes)

Use these as sentence intentions (paraphrase; do not copy verbatim).

1. Axis-anchored exemplars (preferred)

• Systems such as X [@a] and Y [@b] instantiate <axis/design point>, whereas Z [@c] explores a contrasting point under a different protocol.

2. Parenthetical grounding (short, low-risk)

• ... (e.g., X [@a], Y [@b], Z [@c]).

3. Cluster pointer + contrast hint

• Representative implementations span both <cluster A> (X [@a], Y [@b]) and <cluster B> (Z [@c]), suggesting that the trade-off hinges on <lens>.

4. Decision-lens pointer

• For builders choosing between <A> and <B>, prior systems provide concrete instantiations on both sides (X [@a]; Y [@b]; Z [@c]).

5. Evaluation-lens pointer (still evidence-neutral)

• Across commonly used agent evaluations, systems such as X [@a] and Y [@b] illustrate how <lens> is operationalized, while Z [@c] highlights a different constraint.

6. Contrast without list voice

• While many works operationalize <topic> via <mechanism> (X [@a]; Y [@b]), others treat it as <alternative> (Z [@c]), which changes the failure modes discussed later.

Anti-patterns (high-signal “budget dump” voice)

Avoid these stems (they read like automated injection):

• A few representative references include ...
• Notable lines of work include ...
• Concrete examples include ...

If your draft contains these, rewrite them immediately using the patterns above (keep citation keys unchanged).

Placement guidance

• Prefer inserting citations where the subsection already states a concrete contrast or decision lens.
• If you must add a new sentence/mini-paragraph, place it early (often after paragraph 1) so it reads as positioning, not as an afterthought.
• Keep injections subsection-specific: mention the subsection lens (H3 title / contrast_hook) so the same sentence cannot be copy-pasted into every H3.

Workflow

1. Read the budget report (output/CITATION_BUDGET_REPORT.md)

• Treat Global target (policy; blocking) as the PASS line for the pipeline gate (derived from queries.md:citation_target; A150++ default: recommended).
• If Gap: 0, do nothing: write a short PASS report and move on.
• Otherwise, for each H3 with suggested keys, pick enough keys to close the gap to target:
  • small gaps: 3-6 keys / H3
  • A150++ gaps: often 6-12 keys / H3 Prefer keys that are unused globally and avoid repeating the same new keys across many H3s.

2. Inject in the right subsection

• Use outline/outline.yml to confirm H3 ordering and ensure the injected sentence lands inside the correct ### subsection.

3. Inject with paper voice

• Prefer one short, axis-anchored sentence over a long enumerator sentence.
• Keep injections evidence-neutral (NO NEW FACTS) and avoid new numbers.
• Before you commit an injected key, confirm it exists in citations/ref.bib.

4. Write output/CITATION_INJECTION_REPORT.md

• Record which H3s you touched and which keys were added.
• Mark - Status: PASS only when the global target is met.

5. Verify

• Rerun the validator script (below) to recheck the global target.
• Then run draft-polisher to smooth any residual injection voice (citation keys must remain unchanged).

Done criteria

• output/CITATION_INJECTION_REPORT.md exists and is - Status: PASS.
• pipeline-auditor no longer FAILs on “unique citations too low”.

Script (optional; validation only)

You usually do not run this manually; it exists so a pipeline runner can deterministically validate the target.

Quick Start

• python .codex/skills/citation-injector/scripts/run.py --workspace workspaces/<ws>

All Options

• --workspace <dir>
• --unit-id <U###> (optional; for logs)
• --inputs <semicolon-separated> (rare override; prefer defaults)
• --outputs <semicolon-separated> (rare override; default validates output/CITATION_INJECTION_REPORT.md)
• --checkpoint <C#> (optional)

Examples

• After you manually inject citations and write the report:
  • python .codex/skills/citation-injector/scripts/run.py --workspace workspaces/<ws>