agentskills.codes
AU

audit-pitch

by karnrajsinghchauhanSource

Cross-check every factual claim in output/<name>/pitch.html against the sourced lines in research/companies/<name>.md, write AUDIT-<name>.md classifying each claim (BACKED / UNVERIFIED / NO BACKING), and end with a SEND-READY or NEEDS-MY-REVIEW verdict. Use when the user runs /audit-pitch <company n

Install

mkdir -p .claude/skills/audit-pitch && curl -L -o skill.zip "https://agentskills.codes/api/skills/download/16757" && unzip -o skill.zip -d .claude/skills/audit-pitch && rm skill.zip

Installs to .claude/skills/audit-pitch

Activation

This is the description your AI agent reads to decide when to run this skill — the better it matches your request, the more reliably it fires.

Cross-check every factual claim in output/<name>/pitch.html against the sourced lines in research/companies/<name>.md, write AUDIT-<name>.md classifying each claim (BACKED / UNVERIFIED / NO BACKING), and end with a SEND-READY or NEEDS-MY-REVIEW verdict. Use when the user runs /audit-pitch <company name> or asks to audit/verify a pitch artifact.
346 chars✓ has a “when” triggerlonger than Claude Code's old 250-char listing cap (fine on current versions)

About this skill

audit-pitch

Usage: /audit-pitch <company name>

Audit the pitch artifact against its dossier. You are checking the pipeline's honesty, not improving the pitch. Do not edit pitch.html. Do not silently fix anything. If a file is missing (no pitch, no dossier), stop and report which one — that is the failed stage.

Procedure

  1. Read output/<kebab-name>/pitch.html and extract every factual claim — any statement about the company (metrics, customers, funding, tech, quotes) or about the market. Skip: candidate self-descriptions from PROFILE.md/placeholders, rhetorical framing, and the proposal itself ("I'd build X" is an offer, not a claim — but any company-fact embedded in it counts).
  2. Read research/companies/<kebab-name>.md. For each extracted claim, find the dossier line that backs it and classify:
    • BACKED — matches a dossier line that carries a source URL. Record the URL.
    • UNVERIFIED — matches a dossier line explicitly marked UNVERIFIED. (This is a rule violation: UNVERIFIED claims must not appear in pitches — flag it.)
    • NO BACKING — no dossier line supports it. Hallucination risk. Flag loudly. Include the exact quoted sentence from the pitch.
  3. Also check numbers: a claim counts as BACKED only if the figure in the pitch matches the dossier figure exactly (no rounding up, no "over X" inflation of "X").

Output — AUDIT-<kebab-name>.md in the repo root

# AUDIT: <Company>

Audited: <date> | Pitch: output/<kebab-name>/pitch.html | Dossier: research/companies/<kebab-name>.md

## Claims fully backed (N)
| # | Claim (as written in pitch) | Source URL |

## Claims marked UNVERIFIED in dossier (N)   ← must be 0 for SEND-READY
| # | Claim | Dossier line |

## Claims with NO BACKING (N)   ← hallucination risks, must be 0 for SEND-READY
| # | Exact pitch quote | Why nothing in the dossier covers it |

## Placeholder status
PROFILE.md present: yes/no. {{PLACEHOLDER}} slots remaining in pitch: N.

## VERDICT: SEND-READY | NEEDS-MY-REVIEW
<one line of reasoning>

Verdict rules

  • SEND-READY only if: zero UNVERIFIED claims in the pitch, zero NO-BACKING claims, and zero {{PLACEHOLDER}} slots remaining.
  • Otherwise NEEDS-MY-REVIEW, with the blocking items named in the verdict line.
  • Placeholders alone (all claims clean) still force NEEDS-MY-REVIEW — an artifact with placeholders is not sendable.

Search skills

Search the agent skills registry