Letter of Reference System

A complete playbook for requesting, supporting, following up on, and maintaining letters of recommendation — built for someone applying to AI safety fellowships (MATS, Anthropic Safety Fellows, Redwood, METR), research labs, and graduate programs.

---

Quick-Reference Checklist (Run This Every Time)

Use this every time you need a letter. Check each box before moving to the next phase.

Phase 0 — Pre-Ask (Before You Contact Anyone)

• Clarify the exact deadline and submission method (portal link, email, etc.)
• Know what the letter needs to cover (general character? technical skills? research ability? leadership?)
• Identify 3–4 candidate recommenders — more than you need (build in redundancy)
• Rank them: who knows your work best × who has the most credibility for this specific role?
• Check your [[relationship-data]] entries for notes on last contact, shared work, and preferred platform

Phase 1 — The Ask

• Ask in-person or via email — never via text/Slack first
• Ask at least 4–6 weeks before the deadline (8 weeks for fellowships; 6 months for academic LoRs if you want them stored)
• Explicitly ask: "Do you feel you know my work well enough to write a strong letter?" — give them an easy out
• Attach your CV/resume to the initial ask
• Confirm they said yes before sending the full support packet

Phase 2 — The Support Packet (Send Within 48 Hours of Yes)

• Current CV/resume (PDF)
• 1-page brag sheet (see template below)
• Personal statement draft or application essay (if applicable)
• The specific program/job description + why you're applying
• 3–5 talking points you'd love the letter to address
• Deadline, submission link/instructions, and any word limits
• A draft of the letter (optional but powerful — see note below)

Phase 3 — Follow-Up

• Set a calendar reminder 2 weeks before deadline for a check-in
• If no confirmation by 10 days before deadline, send a polite nudge
• If still no response by 5 days out, activate backup recommender

Phase 4 — After Submission

• Confirm they submitted (check application portal)
• Send a thank-you within 24–48 hours of submission (not of your decision)
• Update their entry in [[relationship-data]]
• Add an outcome note when you hear back from the program

---

Part 1 — Timing: How Far in Advance to Ask

By Application Type

Rules of Thumb

• Earlier is almost always better. Professors are juggling 20+ asks at a time. Being early means your letter gets more thought and time.
• Never ask within 2 weeks of a deadline unless it's an emergency and you explicitly acknowledge the short timeline.
• If you want letters stored (e.g., Interfolio, for grad school applications in the future), ask early in the semester and frame it as a long-term ask.
• Rolling deadlines: If a program has rolling admissions, submit early. Ask recommenders the moment you decide to apply.

The Academic Calendar Problem

Professors are at their most swamped in: mid-October (fall midterms), late November/December (end of semester), mid-March (spring midterms), late April/May (end of spring). If your deadline falls in one of these windows, ask earlier than normal to compensate.

---

Part 2 — Who to Ask: Choosing the Right Recommenders

The Core Criteria (rank recommenders on all four)

1. Depth of knowledge of your work — Can they speak to specific projects, contributions, problems you solved? A professor who gave you an A but doesn't know your name is useless. A researcher who supervised one project for 6 months is gold.
2. Credibility for this specific audience — A letter from a well-known AI safety researcher carries more weight at Redwood Research than a letter from an industry manager. Match the recommender to the reader.
3. Enthusiasm — A lukewarm letter from a famous person is worse than an enthusiastic letter from someone less prominent. Ask people who genuinely think highly of you.
4. Reliability — Have they submitted for you before? Do they respond to emails? A brilliant recommender who goes silent is a liability.

The Recommender Hierarchy for AI Safety / Fellowships

Tier 1 (best):

• Research advisors / supervisors from research projects in alignment, interpretability, evals
• Professors who co-authored with you or supervised your thesis
• Researchers at orgs like Anthropic, Redwood, METR, Apollo who know your work

Tier 2 (strong):

• Professors from graduate seminars where you were a standout participant
• Industry mentors / managers from relevant internships
• Fellowship alumni who mentored you through a program (if they can speak to your research)

Tier 3 (avoid if possible):

• Professors who only know you from large lecture classes
• Managers from unrelated jobs
• Peers (unless the program explicitly asks for peer references)
• Anyone who seems hesitant or says "I'll do my best"

The "Strong Letter" Signal

When you ask, listen for this: "Absolutely, I'd be happy to write you a strong letter" vs. "Sure, I can try." The word strong matters. If they hedge, gently give them the out: "Only if you feel you know my work well enough to write something compelling — no pressure at all."

How Many to Have On Deck

• Maintain a recommender bench of 4–6 people at all times (see Part 9)
• For any given application, choose 2–3; keep 1–2 as emergency backups
• Diversify: academic + industry + research if possible

---

Part 3 — How to Ask

In-Person vs. Email

• In-person first is ideal if you have regular meetings with the person (office hours, lab meetings, supervision calls). It signals importance and allows real-time conversation about fit.
• Email is fine — and sometimes better for busy people — if you draft it carefully. Follow up in person if they don't respond within 5–7 days.
• Never ask for the first time via Slack DM or text. Use those platforms only for logistics after they've said yes.

What to Say in the Ask

The ask should do four things:

1. Remind them of who you are and your shared work (briefly)
2. Describe what you're applying for and why it matters to you
3. Ask if they'd be comfortable writing a strong letter
4. Offer an easy exit

---

Templates

Template 1 — The Email Ask (Professor/Research Supervisor)

Subject: Letter of Recommendation Request — [Program Name]

Hi [Professor/Dr. Last Name],

I hope you're doing well. I'm reaching out because I'm applying to [Program Name] — [one sentence description, e.g., "a 6-week AI safety research fellowship at Anthropic"] — and I'd be honored to have your support.

I've been thinking about [brief, specific reference to your shared work — e.g., "the interpretability work we did on activation patching last semester" or "the CS598 seminar where you pushed my thinking on scalable oversight"] and I believe your perspective on my research ability would be particularly valuable for this application.

The deadline is [date]. Would you feel comfortable writing me a strong letter of recommendation? I completely understand if you don't have the bandwidth or if you'd prefer not to — I just wanted to ask given how much I valued working with you.

If you're able to, I'll send over everything you'd need (CV, draft personal statement, the program details, and some talking points) right away.

Thank you either way —
[Your name]

---

Template 2 — The Email Ask (Industry / Lab Contact)

Subject: Quick Ask — Reference Letter for [Program/Company]

Hi [Name],

Hope things are going well on your end. I'm applying to [Program/Company] and wanted to ask if you'd be willing to serve as a reference for me.

Given [specific shared work — e.g., "your mentorship during my MATS stream" or "the evals work we collaborated on"], I think you'd be able to speak to exactly what they're looking for: [1–2 sentence summary of what the program values, e.g., "research rigor and the ability to work independently on open-ended alignment problems"].

The deadline is [date]. Would you be open to this? Happy to send over all the details. No pressure if the timing doesn't work — I know you're busy.

Best,
[Your name]

---

Template 3 — The In-Person Ask Script

"Hey [Name], do you have two minutes? I'm applying to [Program] and I was hoping to ask if you'd be willing to write me a letter of recommendation. Given [your shared work], I think your perspective would be really valuable. The deadline is [date]. Would you feel comfortable writing a strong letter — and if not, no worries at all, I'd completely understand."

---

Part 4 — The Support Packet: What to Provide

A great support packet turns a mediocre letter into a great one. Your job is to make it easy for your recommender to write something compelling without having to remember everything from scratch.

What to Include

1. Current CV / Resume (1–2 pages, PDF) The version most relevant to the application. If it's a research fellowship, lead with research experience. If it's a job, lead with technical skills and projects.

2. The Brag Sheet (1 page) This is the most important thing you can provide. It's a direct-use cheat sheet for your recommender. See template below.

3. Personal Statement Draft Even if it's rough, include it. Recommenders write stronger letters when they understand your narrative and can reinforce it without contradicting it.

4. Program/Job Description + Why You're Applying (2–3 sentences) Tell them what the program is evaluating for. Include any specific traits or skills the application asks evaluators to assess.

5. Talking Points (3–5 bullet points) Specific things you'd like the letter to address. Give concrete examples. Write these as prompts, not commands:

• "One thing you might mention is our work on [X], specifically my contribution of [Y]"
• "You might speak to how I approached [problem] — I'd love for the letter to address intellectual initiative"

6. Submission Logistics

• Exact deadline (with timezone)
• Submission link or instructions (portal URL, email address, upload system)
• Word limit or page limit if any
• Whether the letter is confidential / waived

7. A Draft Letter (Optional but Powerful) For busy recommenders, offering to draft a letter they can edit and sign dramatically increases the quality and submission rate. Frame it as doing them a favor, not putting words in their mouth:

"I've put together a rough draft that you're completely welcome to change or ignore entirely — I just wanted to give you a starting point so this doesn't feel like a blank page."

This is normal and widely accepted. Use the draft to highlight the things you want emphasized.

---

The Brag Sheet Template

## Brag Sheet for [Program Name] — [Your Name]
Prepared for: [Recommender Name]
Deadline: [Date]

---

### What I'm Applying For
[1–2 sentences on the program and what they're selecting for]

### Why I'm a Strong Fit (from my perspective)
[3–4 bullet points — the case you'd make for yourself]

### Experiences I'd Love the Letter to Reference
1. [Project/course/situation] — [What you did, what it demonstrated]
2. [Project/course/situation] — [What you did, what it demonstrated]
3. [Project/course/situation] — [What you did, what it demonstrated]

### Traits I'd Like the Letter to Speak To
- [e.g., Research initiative: the time I independently extended the project scope to include X]
- [e.g., Intellectual rigor: how I pushed back on Y assumption during our meetings]
- [e.g., Collaboration: how I coordinated with the team on Z]

### Logistics
- Deadline: [Date and time, with timezone]
- Submission: [URL / email / upload portal]
- Word limit: [If applicable]
- Confidentiality: [Waived / not waived]

### My Personal Statement (attached)
[Notes on your narrative and how you hope this letter reinforces it]

---

Part 5 — Follow-Up Cadence

The Three-Touch Rule

1. +1 week after support packet sent: A brief confirmation check — "Just wanted to make sure you received everything okay — happy to answer any questions."
2. 2 weeks before deadline: A gentle reminder with deadline confirmed — "Just a heads up that the deadline for [Program] is [date]. Please let me know if you need anything from me!"
3. 5–7 days before deadline: A more direct check-in — "Wanted to loop back since the deadline is coming up on [date]. Please let me know if you've had a chance to submit or if there's anything blocking you."

Tone Guidelines

• Always thank them first before asking for a status update
• Never make them feel guilty — they're doing you a favor
• Give them a direct action: "Let me know if you've submitted" (not "have you done it yet?")
• If the portal has a tracker, check it yourself before sending a nudge

What Not to Do

• Don't follow up more than once per week
• Don't send multiple follow-ups on the same day via different platforms
• Don't express anxiety or panic — keep every message calm and grateful

---

Template 4 — The Two-Week Reminder

Hi [Name],

I just wanted to send a friendly reminder that the deadline for [Program] is coming up on [date]. I hope things have been manageable on your end!

Please let me know if you need anything from me — additional context, a revised draft, or any other support. And thank you again for doing this — I really appreciate it.

[Your name]

---

Template 5 — The Final Nudge (5–7 Days Out)

Hi [Name],

I hope you're doing well. The deadline for [Program] is [date] — just [X] days away. I wanted to loop back in case anything came up or in case there's anything I can do to make the submission easier.

[Optional: if they're using a portal] The submission portal is at [link] and your instructions are in the earlier email I sent.

Thank you again — this truly means a lot to me.

[Your name]

---

Part 6 — When They Say Yes / No

When They Say Yes

• Respond within 24 hours with gratitude and the support packet
• Confirm the deadline one more time in the response
• Ask if they have a preferred format for receiving materials
• Add a calendar event to check in 2 weeks before deadline

When They Say No

A no is not a failure — it's information, and often a kindness. People who know they can't write you a strong letter are doing you a favor by declining.

What to say:

"I completely understand — and I really appreciate you being upfront. No worries at all. [Optional: Is there someone else you'd suggest I reach out to for this type of application?]"

What to do:

• Move to your next candidate immediately
• Do not express disappointment — keep the relationship intact
• Don't burn any bridges; you may need this person for a future ask

---

Part 7 — When Things Go Wrong

Scenario 1: The Recommender Goes Silent

This is the most common problem. Professors and researchers are busy. Radio silence doesn't always mean no — sometimes it means forgotten.

The Protocol:

1. First silence (no response to initial ask within 5–7 days): Follow up via the same channel with a brief re-send of the original message.
2. Second silence (no response after follow-up): Try a different channel — if you emailed, try an in-person check-in or Slack if appropriate.
3. Third silence (10+ days, no acknowledgment at all): Activate a backup recommender immediately. Proceed in parallel — do not wait.
4. Approaching deadline with no submission confirmed: On day 5–7 before deadline, send a polite but direct message (Template 5 above) and simultaneously confirm your backup is ready to submit.

If you have to replace them: Move fast. Contact your backup with the full support packet immediately. You don't need to explain the situation in detail.

---

Template 6 — Re-Engage After Silence

Hi [Name],

I wanted to follow up on my earlier email about a letter of recommendation for [Program]. I know this time of year is incredibly busy, so no pressure — I just wanted to make sure my message hadn't gotten buried.

The deadline is [date]. If you're not able to do this right now, I completely understand. Just let me know either way so I can plan accordingly.

Thank you,
[Your name]

---

Template 7 — Activating a Backup Recommender

Hi [Name],

I hope you're doing well. I'm applying to [Program] with a deadline of [date] and I'm reaching out to ask if you'd be willing to write a letter of recommendation for me.

I know this is relatively short notice — [X days/weeks] — and I completely understand if that doesn't work. But given [shared work/context], I think your perspective would be really valuable, and I wanted to ask.

I can send over everything you'd need right away: my CV, a brag sheet, the program description, and a draft letter if helpful. The submission is [portal/email].

Please let me know if this is something you can do. Thank you so much.

[Your name]

---

Scenario 2: The Letter is Late

If the deadline has passed and the letter wasn't submitted:

1. Email the program directly and explain: "One of my recommenders has had an unexpected conflict. Is it possible to have a brief extension for this letter? I can have it submitted within [X days]." Many programs will accommodate this, especially for fellowships.
2. Simultaneously push your recommender one final time with the extended deadline.
3. If the program is inflexible, submit what you have and follow up afterward.

Scenario 3: A Weak Letter Warning Sign

If a recommender seems uncertain, hedging, or says things like "I'll try my best" or "I'm not sure I remember all the details of your work" — treat this as a red flag. Gently give them an out:

"I want to make sure you feel comfortable with this — if there's any uncertainty about whether you can write something strong, I'd completely understand stepping back. I don't want to put you in an awkward position."

A reluctant recommender will almost always take this out, which is better for both of you.

---

Part 8 — How to Thank Your Recommender

Timing

• Send a brief thank-you within 24–48 hours of confirmed submission — not after you get your result
• Send a second note when you get an outcome (accepted, rejected, waitlisted) — close the loop

Medium

• Email is standard
• Handwritten note is memorable and rare — if you have a close relationship with the person and they went above and beyond, this is worth doing
• A small gift (coffee card, book) is appropriate for a very close mentor who wrote multiple letters for you in one cycle

What to Say

Be specific. Don't just say "thank you for writing the letter." Reference the effort they put in or the relationship behind it.

---

Template 8 — Post-Submission Thank You

Hi [Name],

I just saw that the letter has been submitted — thank you so much. I know how much time and thought goes into writing one of these, and it genuinely means a lot to me that you took the time.

I'll let you know how it goes. Thank you again.

[Your name]

---

Template 9 — Outcome Follow-Up

Hi [Name],

I wanted to close the loop and let you know: I [was accepted to / didn't get into] [Program].

[If accepted:] I'm really excited about this opportunity, and your letter was a meaningful part of my application. Thank you again.

[If rejected:] I'm disappointed, but I'm already thinking about what's next. Thank you for your support — it meant a lot to have you in my corner, and I hope I can make you proud in the next round.

[Your name]

---

Part 9 — Maintaining Your Recommender Bench

The best time to strengthen a relationship with a potential recommender is before you need a letter. Build and maintain a bench of 4–6 people year-round.

Who Belongs on Your Bench

• Current and past research advisors/supervisors
• 2–3 professors from seminars or courses where you stood out
• Mentors from fellowships, lab rotations, or internships
• Senior researchers in the AI safety community who know your work directly

How to Maintain the Relationships

Bi-monthly (every 2 months minimum):

• Share a relevant paper, article, or piece of work they'd find interesting
• Update them briefly on what you're working on
• Ask for a 30-minute check-in if the relationship warrants it

After every milestone:

• Send them a note when you publish, present, or complete something significant
• Give them visibility into your trajectory — they're writing for a future version of you

Keep your [[relationship-data]] entries current:

• What did you work on together?
• What were their specific comments on your work?
• What do they think your strengths are?
• When did you last speak?

The Relationship Investment Rule

Don't only contact people when you need something. Each "ask" should be preceded by at least two meaningful "gives" (sharing work, engaging with their ideas, updating them on outcomes). This is especially important in a small field like AI safety where reputation compounds.

Before Application Season

• 1 month before you anticipate applying: reach out for a check-in / coffee chat
• Remind them of your recent work and share where you're applying
• Don't make the ask yet — just warm the relationship
• Then make the formal ask after this touchpoint

---

Part 10 — Fellowship and Program-Specific Nuances

MATS (Machine Learning Alignment Theory Scholars)

• Requires 2 references contacted in a later stage (not upfront letters)
• References are typically contacted by MATS directly after initial screening
• Ask references in advance and let them know they may be contacted
• Ideal references: researchers who can speak to your ability to do independent alignment research

Anthropic Safety Fellows / Policy Fellows

• Specifics vary by cohort — always check the current application page
• Letters are often not required upfront; references may be contacted if you advance
• Prioritize references from the AI safety ecosystem (Anthropic, Redwood, METR, ARC, etc.)

Academic PhD Programs

• Typically require 3 letters submitted through an application portal
• Letters should be from people who can speak to research ability specifically
• At least 2 should be faculty; one from industry is acceptable but not preferred at top programs
• The letter from your direct research advisor is the most important one

Professional Jobs at AI Labs (Anthropic, OpenAI, DeepMind, etc.)

• Reference checks typically happen post-offer, not during application
• When you receive an offer, you'll be asked for 3–5 references
• These are phone call-based, not written letters — but the same principles apply
• Give references advance notice that they may receive a call

External Fellowships (NSF GRFP, Hertz, etc.)

• Usually require letters at the time of application, often with a strict portal upload
• Letters should specifically address the fellowship's evaluation criteria (for NSF: intellectual merit + broader impacts)
• Give recommenders the evaluation rubric so they can align their language

Key Differences: Academic vs. Professional vs. Fellowship Letters

---

Part 11 — What Makes a Strong vs. Weak Letter

Understanding this helps you select the right recommenders and prepare them well.

Anatomy of a Strong Letter

1. Specificity over generality Strong: "Matt identified a critical flaw in our evaluation design that three senior researchers had missed, then independently developed a fix overnight." Weak: "Matt is a hardworking and dedicated student who always puts in maximum effort."

2. Comparative language Strong: "In 15 years of supervising graduate students, Matt is in the top 5% for independent research instinct." Weak: "Matt would be an asset to any team."

3. Story-driven evidence The best letters tell a short story about a specific moment or project. Not a list of adjectives.

4. Addressing the selection criteria A letter for an AI safety fellowship that never mentions safety thinking, interpretability, or research rigor is wasted.

5. Personal voice A letter that sounds like the recommender — not like a template they barely edited — signals that they care about the outcome.

Red Flags in a Letter (That You Should Prevent)

• Generic praise with no specifics
• Vague language about "potential" without evidence of actual performance
• Letters that focus entirely on grades and GPA
• Letters from people who were clearly asked at the last minute (they often sound rushed)
• Letters that are too short (< 400 words) without explanation

How to Help Your Recommender Write a Better Letter

• The brag sheet and talking points are your primary lever here
• If they're open to it, offer to draft a version they can revise
• In your support packet, say: "I'd especially love if the letter could speak to [specific thing] — I think it's the most relevant part of my background for this program"
• Don't be shy about this — a recommender who knows what you need will deliver a better letter

---

Related Notes

• [[relationship-system]]
• [[job-achieving-system]]
• [[cold-emailing-system]]
• [[important-meeting-system]]
• [[relationship-data]]