Compressing Knowledge
Compression is trustworthy only if something that never saw the original can still pass a test on it. Treat every draft as unverified until a source-blind agent proves otherwise.
When Not To Use
- Some information loss is acceptable — just summarize; the harness is overkill.
- The task is prose style or code tightening, not fact preservation — use a style or simplification pass instead.
- No subagent can run with a genuinely empty context (see Verification) — say so; offer only an unverified summary if still wanted.
Workflow
Baseline. Record the source's token count with a consistent proxy (wc -w is enough — exact tokenizer accuracy doesn't matter, only consistency across iterations).
Extract test questions. Before drafting, pull out question/answer pairs, each grounded in one distinct fact and traceable to a specific passage. Scale count to length: roughly one per 40 source tokens, minimum 5, or the user's stated count. Done when no two questions share a fact and every answer is written down.
Verify the harness. Spawn a fresh, non-fork subagent (Agent, e.g. general-purpose) with only the original source and the bare questions (no answers). Grade its answers against the recorded ones yourself. A failure means the question is ambiguous or ungrounded — fix or drop it and re-run until the source passes every question. This is the ground truth later drafts are measured against.
Cut, then test. Cut redundancy, restatement, and filler into a candidate draft — cutting content, not whitespace or formatting, which doesn't reduce real footprint. Spawn a fresh subagent with only the candidate draft (never the original) and the bare questions; grade its answers.
- All pass → new safe anchor. Return here and cut further.
- Any fail → floor found; go to step 5.
Repair. Patch the failing draft with the minimum addition that restores the missing fact(s) — don't revert wholesale to the last anchor. Re-test with another fresh subagent until it passes every question, then retry step 4 from the repaired draft.
Stop. Finish when two consecutive repair-then-shrink attempts fail to beat the last safe anchor's size, or a user-given budget is exhausted. Never hand back a draft that hasn't itself passed a fresh, source-blind test at 100%.
Report. Deliver the final text plus: baseline and final token counts and reduction %, the full question/answer set (for re-verifying future edits), and an iteration log (draft tokens → pass/fail → what failed → repair made).
Verification
The grading subagent must never have seen the original and must never be a fork — forks inherit your context, including the source, which defeats the test. Use a fresh, non-fork Agent call every time. If the environment can't spawn one, verification is unavailable: say so plainly and never claim the result is lossless without it.
Grade strictly: a hedge ("not stated"), a partial answer, or a wrong fact all count as failure — the method depends on treating "probably fine" as failure.
1---2name: compressing-knowledge3description: Compresses documents, prompts, or knowledge sources into fewer tokens while verifying that every answerable fact survives, using a test-question harness graded by a fresh, source-blind agent. Use when asked to shrink a doc, README, or prompt losslessly, cut token or context-window footprint without losing information, or confirm a compression keeps every fact intact.4license: MIT5---67# Compressing Knowledge89Compression is trustworthy only if something that never saw the original can still pass a test on it. Treat every draft as unverified until a source-blind agent proves otherwise.1011## When Not To Use1213- Some information loss is acceptable — just summarize; the harness is overkill.14- The task is prose style or code tightening, not fact preservation — use a style or simplification pass instead.15- No subagent can run with a genuinely empty context (see Verification) — say so; offer only an unverified summary if still wanted.1617## Workflow18191. **Baseline.** Record the source's token count with a consistent proxy (`wc -w` is enough — exact tokenizer accuracy doesn't matter, only consistency across iterations).20212. **Extract test questions.** Before drafting, pull out question/answer pairs, each grounded in one distinct fact and traceable to a specific passage. Scale count to length: roughly one per 40 source tokens, minimum 5, or the user's stated count. Done when no two questions share a fact and every answer is written down.22233. **Verify the harness.** Spawn a fresh, non-fork subagent (`Agent`, e.g. `general-purpose`) with *only* the original source and the bare questions (no answers). Grade its answers against the recorded ones yourself. A failure means the question is ambiguous or ungrounded — fix or drop it and re-run until the source passes every question. This is the ground truth later drafts are measured against.24254. **Cut, then test.** Cut redundancy, restatement, and filler into a candidate draft — cutting content, not whitespace or formatting, which doesn't reduce real footprint. Spawn a fresh subagent with *only the candidate draft* (never the original) and the bare questions; grade its answers.26 - All pass → new safe anchor. Return here and cut further.27 - Any fail → floor found; go to step 5.28295. **Repair.** Patch the failing draft with the minimum addition that restores the missing fact(s) — don't revert wholesale to the last anchor. Re-test with another fresh subagent until it passes every question, then retry step 4 from the repaired draft.30316. **Stop.** Finish when two consecutive repair-then-shrink attempts fail to beat the last safe anchor's size, or a user-given budget is exhausted. Never hand back a draft that hasn't itself passed a fresh, source-blind test at 100%.32337. **Report.** Deliver the final text plus: baseline and final token counts and reduction %, the full question/answer set (for re-verifying future edits), and an iteration log (draft tokens → pass/fail → what failed → repair made).3435## Verification3637The grading subagent must never have seen the original and must never be a fork — forks inherit your context, including the source, which defeats the test. Use a fresh, non-fork `Agent` call every time. If the environment can't spawn one, verification is unavailable: say so plainly and never claim the result is lossless without it.3839Grade strictly: a hedge ("not stated"), a partial answer, or a wrong fact all count as failure — the method depends on treating "probably fine" as failure.
h3y6e