llm-tldr

llm-TLDR

Goal: lossless compression of your OWN output. Cut words, keep information. Dropping a caveat, number, decision, or risk to "make it shorter" = FAILURE, not a TLDR. This file is its own example — keep it this dense.

Decision rule (the one rule)

KEEP if it changes a decision; CUT if it only adds length.

• KEEP (decisive atoms): results, numbers, %, counts, pass/fail, error codes, IDs, paths, file:line, versions, flags, env vars, named entities, the why-it-matters, caveats, risks, assumptions, unknowns, open decisions.
• CUT: greeting, flattery, "Here is a summary", "I will now…", process narration, restated prompt, hedging filler, repeated context.
• Tie-breaker: when unsure, KEEP. Lossless beats short.

(This taxonomy is canonical — referenced, not re-listed, elsewhere.)

Algorithm (run in order)

1. BLUF. Line 1 = the single most decision-relevant takeaway. No preamble, no restating the question.
2. Extract atoms. List every decisive atom (see Decision rule).
3. Score KEEP/CUT. Apply the Decision rule to each atom.
4. Order by importance, not chronology. Sort KEEP atoms: bottom line → what changed/found → evidence → caveats/known issues → open decisions. Drop the timeline.
5. One bullet per atom. Each = bold label/claim + its one decisive detail. One line where possible. Define a complex term inline only if load-bearing.
6. Compress expression only. Strip articles, hedges, repetition. Keep exact values verbatim (random()<0.85, 31/31 PASS, txn_8f3a2c, env flags). Prefer the concrete fact over a vague gloss; state verification status (pass/fail + counts).
7. Calibrate honesty. Surface unknowns, assumptions, failures explicitly. Say "I don't know" / "unverified" / "cannot, because X" when true. No false confidence, no epistemic cowardice.
8. End with decisions if open. Numbered, actionable options. Omit if nothing is pending.
9. Scale to input. Trivial input → one-line TLDR. Never pad to fill a template (see micro-example below).

Format

TLDR:
- **<Label>:** <claim> — <decisive detail / data / status>.   (importance order)
- **Caveats / known issues:** <what is unverified, risky, or broken>.

Pick any: (1) <action> (2) <action> (3) <action>

TLDR: is an allowed single-token header, not preamble. Drop the bullets/next-steps sections when the content does not warrant them.

Worked example (the gold target)

Input: long agent session that fixed a payment service, then user types "TLDR?".

TLDR:
- Why it faked success: PaymentService was always a stub — "charged" via random()<0.85 + a fake txn ID. Never hit a real gateway.
- Fixed it: charge() now returns honest states (authorized / declined / needs-3DS), never fabricates. Real gateway wired behind PAYMENTS_API_KEY; stub off unless MOCK_PAYMENTS=true. Verified: build PASS, 31/31 tests PASS, sandbox PASS.
- Refunds: live gateway works, but I can't confirm settlement timing — the sandbox does not emit settlement webhooks.
- Known issue: webhook handler misclassifies refunds as chargebacks (status-string match bug), so refund events get the wrong ledger entry.
- Nothing committed (still on main).

Pick any: (1) fix the refund/chargeback misclassification (2) delete the dead stub (3) verify settlement against a live (non-sandbox) account (4) commit.

Why it works: BLUF; importance-ordered bold-labeled bullets (claim + decisive detail); data kept verbatim; honest about limits and why; known issues surfaced; numbered decisions; zero filler.

Before -> after (the failure mode)

BAD (naive shortening — drops decisive facts):

Looked into the payment service and fixed it. It used stub data before but now does real charges. Tested it too. A couple small things left.

GOOD (lossless — similar length, real information):

Root cause: stub charged via random()<0.85 + fake txn ID, never hit a real gateway. Fix: real gateway behind PAYMENTS_API_KEY, stub off unless MOCK_PAYMENTS=true; build + 31/31 tests + sandbox PASS. Open: refund/chargeback misclassification bug, 1 dead stub, nothing committed.

BAD lost the cause, the proof (31/31), and what's still broken. GOOD kept all three in fewer words.

Micro-example (scale to content)

Input: "Did the build pass?" → TLDR: Build PASS, 31/31 tests PASS. One line in, one line out — no bullets, no header, no padding.

Grounding (Anthropic)

• Direct, no preamble (steps 1, 7): "Respond directly without preamble. Do not start with phrases like 'Here is...', 'Based on...', etc." (prompt engineering)
• Fact-based, not self-celebratory (step 6): "Provides fact-based progress reports rather than self-celebratory updates" (prompt engineering)
• Less verbose by default (steps 5, 9): "Less verbose: May skip detailed summaries for efficiency unless prompted otherwise" (prompt engineering)
• Calibrated honesty (step 7): "It acknowledges its own uncertainty or lack of knowledge when relevant, and avoids conveying beliefs with more or less confidence than it actually has." (constitution)
• No vagueness to placate (step 7): "Claude should be diplomatically honest rather than dishonestly diplomatic. Epistemic cowardice—giving deliberately vague or noncommittal answers to avoid controversy or to placate people—violates honesty norms." (constitution)

All five quotes verified verbatim against the linked sources (June 2026).

Checklist (run before sending)

• Line 1 = bottom line (BLUF); no preamble.
• Bullets importance-ordered, not chronological; each = bold label + decisive detail.
• All decisive atoms preserved verbatim (see Decision rule).
• Verification status stated (pass/fail + counts).
• Caveats, risks, unknowns surfaced; confidence calibrated.
• Open decisions numbered (or omitted if none).
• Length scaled to input; zero filler, flattery, or restated question.