Recommendation Letters — Evidence-Anchored References Written About a Candidate

The faculty task this skill owns is writing about someone else: the letter a professor sends to an admissions committee, a hiring chair, a tenure case, or a fellowship panel. It is the mirror image of alterlab-academic-career, which helps the candidate write their own CV and statements. Here the author is the recommender, the subject is a third party, and the failure mode is not modesty but fabrication — the warm letter full of superlatives that an evaluator cannot act on because nothing is anchored to evidence.

This skill turns a structured dossier (what the recommender actually knows) into a letter whose every strong claim is earned by a specific, checkable example, with specificity and register calibrated to the letter type.

When to Use This Skill

Use this skill when the request is to draft, revise, or evidence-check a letter written about a candidate by a recommender:

• "Write a recommendation letter for my PhD student applying to postdocs."
• "I'm a reference for a colleague's faculty application — draft the letter."
• "I've been asked for an external review letter for X's tenure case."
• "Draft a fellowship / award nomination for my mentee."
• "Here's my draft letter — flag anything I claim without backing it up."
• "Make this reference letter more specific; it reads like boilerplate."

The skill always treats the recommender as the author of record: it produces a draft for them to verify, edit, sign, and submit, never a letter it claims to send or to vouch for on its own authority.

Does NOT Trigger

If the candidate is writing about themselves, stop and hand off to alterlab-academic-career. The two skills are deliberately disjoint.

The Five Letter Types (calibration table)

Each type has a different reader, decision, and register. The scaffold script keys off this taxonomy. Full per-type guidance, openings, and worked structure live in references/letter_types.md.

The external-review letter is the one most often miswritten as an advocacy letter. It is not: tenure reviewers are asked to evaluate, and a letter that only praises reads as uninformative. references/letter_types.md spells out the register shift and the standard "would this person get tenure here?" prompt.

The Core Discipline: Evidence-Anchored Claims

Every adjective must be cashed out by a specific example. The pattern:

Claim → Evidence → Significance. "She is an unusually independent researcher" (claim) "— in her second year she reframed the project after a null result, designed the follow-up assay herself, and that pivot became the paper's central finding" (evidence) "— work I would expect from a senior postdoc, not a second-year student" (significance / calibration).

A claim with no evidence in the dossier is the single most common defect. The no-fabrication guard (scripts/claim_guard.py) lints a draft for exactly this: unsupported superlatives ("the best student I have ever had"), unanchored rankings ("top 1%"), and evaluative adjectives that no nearby sentence backs with a concrete instance. It does not invent the missing evidence — it flags the claim and asks the recommender to supply the example or soften the claim.

Calibrated specificity is the companion rule: superlatives and rankings are allowed and valuable — but only when the recommender can ground them ("top 2 of the ~40 PhD students I have advised in 18 years"). An ungrounded "top 1%" is weaker than a grounded "the strongest of the six students in this year's lab." See references/evidence_and_specificity.md for the claim taxonomy, the ranking grammar, and the rhetoric of comparison.

Hard rule — zero fabrication

This skill writes only from the supplied dossier. It must never:

• invent achievements, metrics, rankings, dates, or anecdotes the recommender did

not provide;

• assert a comparison ("top 5%") the recommender did not make;
• manufacture a relationship detail (how long / in what capacity they were known).

When the dossier is thin, the correct output is a draft with bracketed gaps ([recommender: a specific example of independence here]) plus a short list of the facts the recommender must supply — not a fluent letter that papers over the holes with plausible fiction. A letter is a signed personal attestation; inventing its content is a research-integrity problem, not a style choice.

Workflow

1. Collect the dossier (structured intake)

Ask for, or extract from what was provided: letter type; candidate name + role; target (program/position/award + institution); relationship (capacity and duration the recommender knew the candidate); the candidate's accomplishments (with any metrics, artifacts, anecdotes); the specific qualities the target asks to be addressed; and deadline / length norms. references/intake_checklist.md is the full prompt list; missing fields become bracketed gaps, never invented content.

2. Scaffold the letter

uv run python skills/faculty-life/alterlab-recommendation-letters/scripts/letter_scaffold.py \
    --type tenure-external \
    --candidate "Dr. A. Yılmaz" \
    --relationship "external reviewer; known through the field since 2016" \
    --out letter_skeleton.md

The scaffold emits a type-appropriate section skeleton — opening that states the relationship and basis for assessment, body sections matched to the type's evaluation dimensions (e.g. independence/impact/standing for external review; potential/trajectory/fit for admission), and a calibrated closing. Each section carries an inline reminder of the Claim → Evidence → Significance pattern. Pass --list-types to see the five supported types. The script is stdlib-only.

3. Draft from the dossier

Fill each section using only dossier facts. Lead every evaluative claim with the specific example that earns it. Match the register to the type (advocacy for admission/fellowship; arms-length evaluation for external review).

4. Run the no-fabrication guard

uv run python skills/faculty-life/alterlab-recommendation-letters/scripts/claim_guard.py \
    letter_draft.md --json

It returns each flagged sentence with a reason (unsupported-superlative, unanchored-ranking, evidence-free-claim) and a fix prompt. Resolve every flag by adding the recommender's evidence or softening the claim — never by inventing support. Re-run until clean (exit 0).

5. Final review hand-off

Present the draft plus: the list of bracketed gaps the recommender must fill, the guard's residual flags, and a one-line reminder that the recommender must verify every factual claim and sign/submit it themselves.

Self-Check Before Returning a Letter

• Every superlative earned? Each "best/strongest/exceptional" is followed by a

concrete, dossier-sourced instance — or it is bracketed for the recommender.

• No invented facts. Nothing in the letter — metric, ranking, anecdote, date —

is absent from the supplied dossier.

• Register matches type. External-review reads as evaluation, not cheerleading;

admission/fellowship reads as grounded advocacy.

• Relationship stated up front. The opening makes the basis for the assessment

explicit (how, how long, in what capacity).

• Gaps surfaced, not filled. Thin areas appear as bracketed asks, not fiction.
• Right skill? A self-authored career document belongs to

alterlab-academic-career; hand it off.

References

• references/letter_types.md — per-type readers, openings, structure, and the

external-review register shift.

• references/evidence_and_specificity.md — Claim → Evidence → Significance,

the claim taxonomy, ranking grammar, and comparison rhetoric.

• references/intake_checklist.md — the structured dossier intake prompt and how

thin fields become bracketed gaps.

Part of the AlterLab Academic Skills suite.