菜单
Prevents lossy compression from erasing conditions, exceptions, and uncertainties during summarization.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Designs API contracts for consumer experience first — making correct usage obvious and incorrect usage impossible.
Replaces adjectives with evidence, claims with demonstrations, and promises with proof for conversion-oriented writing.
Forces the model to build a genuine prosecution case against its own answer before delivering it.
Detects and corrects objective drift during extended tasks by periodically checking work against the original request.
Decides whether to ask clarifying questions or proceed with an answer, optimizing for information value vs. delay cost.
Self-selects and sequences the right depth-skills proportional to task consequence, preventing both under- and over-analysis.
| name | fidelity |
| codename | FIDELITY |
| internal | Compression Integrity Verifier |
| version | 1.1 |
| tier | integrity |
| trigger | ["summarize","give me the short version","TLDR","bottom line","any time complex analysis is condensed into a final answer"] |
| description | Prevents lossy compression from erasing conditions, exceptions, and uncertainties during summarization. |
| author | Kshitijpalsinghtomar |
| tags | ["compression","summary","fidelity","truth-preservation","delivery"] |
| artifacts | ["critical-information-tags","compressed-version","fidelity-diff","fidelity-verdict"] |
| composable_with | ["provenance","contradict","anchor"] |
You just did deep thinking. You explored approaches, found edge cases, identified caveats, surfaced conditions. Now you're compressing it into a clean final answer.
This is the moment truth disappears.
Two pressures compete: be thorough and be concise. The result is lossy compression:
The summary is cleaner, shorter, more confident — and less true than the analysis that produced it. The user makes decisions based on a simplified reality that you know is incomplete.
Before writing the compressed version, read the full analysis and tag every item that, if dropped, makes the summary misleading. Write each tag:
CRITICAL INFORMATION TAGS
────────────────────────────────────────
TAG 1 — CONDITION:
Full: "[recommendation] IF [condition]"
If dropped: user tries it where [condition] is false → [consequence]
TAG 2 — EXCEPTION:
Full: "True EXCEPT when [scenario]"
If dropped: user applies universally → hits [scenario] unprepared
TAG 3 — UNCERTAINTY:
Full: "[confidence level] because [evidence state]"
If dropped: user treats as certain → no contingency when wrong
TAG 4 — ALTERNATIVE:
Full: "Option B was close — better if [condition changes]"
If dropped: user can't adapt when conditions change
TAG 5 — DEPENDENCY:
Full: "Depends on [X] being true/available/stable"
If dropped: user doesn't verify [X] → failure when [X] is absent
────────────────────────────────────────
Not every answer has all five types. Tag what exists. The types to scan for:
Artifact: The tagged list. Step 3 verifies each tag survives compression.
Write the clear, concise answer you want to deliver. Do not consult the tags. Write naturally — as concise as the content allows.
Artifact: The compressed version. Step 3 diffs this against Step 1.
For each tag from Step 1:
FIDELITY DIFF
────────────────────────────────────────
TAG 1 — CONDITION:
In compressed version?: [yes — preserved / no — dropped]
If dropped: could the user make a wrong decision? [yes / no]
If yes: RESTORE
TAG 2 — EXCEPTION:
In compressed version?: [yes / no]
If dropped: could the user be surprised by a failure? [yes / no]
If yes: RESTORE
...
────────────────────────────────────────
Restore rule: Any tag that is dropped AND could lead to a wrong decision or surprise failure MUST be restored.
Artifact: The fidelity diff showing what was preserved and what was restored.
Restoration does not mean making the summary as long as the full analysis. Use minimum-length preservation techniques:
The goal: the compressed version is as TRUE as the full analysis, not as LONG.
FIDELITY VERDICT
────────────────────────────────────────
Critical items tagged: [count]
Preserved in first draft: [count]
Restored after diff: [count]
Intentionally omitted: [count] — [justification per item]
Status: [LOSSLESS / ACCEPTABLE / UNACCEPTABLE]
Lossless: All critical items preserved. Summary is as true as the analysis.
Acceptable: Minor items omitted with justification. No decision risk.
Unacceptable: Critical items missing. Revise before delivery.
────────────────────────────────────────
The model's best thinking happens during exploration. Its worst habit is discarding that thinking during delivery. If the final answer is a lossy compression of the truth, the user acts on an incomplete version of what the model itself knows is more complex. This skill ensures the distance between what-the-model-knows and what-the-user-receives is minimized — not by being verbose, but by preserving the specific pieces of truth that change decisions.