natejswenson Codex runtime
When running in Codex, invoke this skill as $brandreport. Resolve scripts, assets,
and references from the directory containing this SKILL.md, regardless of the
current working directory. Existing ~/.claude/ personal-data paths remain valid
and are still used by the bundled scripts; they do not require Claude to run.
Map Read/Write/Edit/Bash to the available file and shell tools, and
WebSearch/WebFetch to available web tools. For AskUserQuestion, use an
available question tool or a concise chat question; wait for answers that gate
action. Use Codex's delegation tools for required subagents when available;
otherwise disclose that independent execution is unavailable. Discover connected
apps by capability rather than assuming Claude MCP tool names exist.
/brandreport — Give it just a name; it blind-searches the open web for that person's presence, keeps only what it can prove is them, and renders what it found as a press-styled brand report
You are running the brandreport skill.
Announce at start: "I'm using the brandreport skill — Give it just a name; it blind-searches the open web for that person's presence, keeps only what it can prove is them, and renders what it found as a press-styled brand report."
Commands below run from the directory containing this
SKILL.md($SKILL_DIR). Resolve it once. Pass--repo <path>to work against the user's repo.
The one rule
Never attribute unverified content: nothing reaches the report unless it is tied to the actual person by a recorded corroborating signal (cross-links, shared handles, bio matches) — same-name findings it cannot tie are listed as unconfirmed, never silently included and never silently dropped.
What is code and what is judgment
The split is declared in skill-invariants.json and checked — a deterministic
step whose command does not exist fails skillfactory verify.
| Deterministic — the machine decides | Command |
|---|---|
| create the run layout for a subject | node scripts/brandreport.js init |
| emit the per-platform handle-sweep checklist | node scripts/brandreport.js sweep |
| file each fetched artifact with provenance and identity status | node scripts/brandreport.js add |
| table the corpus and its confirmed/unconfirmed split | node scripts/brandreport.js status |
| enforce the attribution gate before anything renders | node scripts/brandreport.js gate |
| render the press-styled HTML report offline from the snapshot | node scripts/brandreport.js report |
| Model judgment — nothing on disk answers it | Why |
|---|---|
| the blind discovery itself — which searches to run, which hits to follow, when coverage is enough | no file records where a person exists online; the search plan is invented per subject and adapts to what each round surfaces |
| identity corroboration — deciding a hit is the same person | cross-links, shared handles and bio overlap are read and weighed, not pattern-matched; the gate only checks the decision was recorded, the model makes it |
| the brand analysis — themes, consistency across platforms, gaps, how a stranger would read this person | judgment about impression and coherence that no snapshot states |
The flow
1. Start the run
brandreport init --subject "<name>" — the run directory, the snapshot
corpus, the findings file, the report path. One table back. Runs live in
~/.claude/brandreport/<slug> by default; a run is fetched personal data
and never belongs in version control, so inside a repo point --out at a
gitignored directory (brandreport-runs/<slug> in this repo).
2. Discover — search rounds, then the sweep
The discovery itself is judgment; references/discovery.md is its shape:
seed from the bare name, anchor on the first artifact the person plainly
controls, widen from everything it links. Then the mandatory part:
brandreport sweep --handle <every handle the anchor uses> prints the
per-platform probe checklist. Probe every row before discovery may stop —
search indexes walled platforms badly, and the first real run of this skill
missed the subject's own LinkedIn and X by trusting search alone. An account
proven to exist but unreadable logged-out files with --existence-only;
an HTTP 200 with an empty body proves nothing either way.
File every artifact as it is fetched — brandreport add with a one-sentence
--corroboration for confirmed, --why for unconfirmed. The corpus and the
sweep table already answer most questions — never ask about anything in
it. Ask only when identity is genuinely ambiguous after probing (a
same-handle account whose display name matches no anchor), at most two
questions, one at a time. The subject may direct that such an artifact be
left out entirely before it is filed — honour it and say so in the
conversation; once filed, exclusion happens only through the report's
residue section, never by deletion.
3. Judge — write findings.json
Claims, themes, gaps, summary — the contract is references/anatomy.md.
Every claim and theme cites snapshot ids; gaps assert absence and cite
nothing.
4. Gate, then render, then show
brandreport gate, then brandreport report. Open the rendered report on
the user's own screen and put a screenshot in the transcript — a report the
user has not seen is not a result.
5. Re-runs refresh, never duplicate
Re-fetch the anchor artifacts first — profiles change, and a changed anchor
changes what downstream corroborations can say (add --id sN replaces a
snapshot in place, same citation key). Then re-run the sweep: platforms the
subject joined since, or newly cross-linked, are exactly what a stale report
misses.
Commands
| Command | Returns |
|---|---|
brandreport init |
creates a run directory for a subject name and returns its layout as a table — snapshot dir, findings file, report path |
brandreport sweep |
prints the per-platform probe checklist for one or more handles, with the walled-platform workaround for each — run before discovery may stop |
brandreport add |
files one fetched artifact into the snapshot with provenance — URL, fetched-at, kind, identity status (confirmed/unconfirmed) and the corroboration note — and returns the updated corpus row |
brandreport status |
tables the whole corpus: every snapshot with its source, kind, identity status and corroboration, plus counts of confirmed vs unconfirmed |
brandreport gate |
enforces the one rule as code: exits non-zero if any confirmed item lacks a recorded corroboration, any findings claim cites a snapshot that does not exist, or any unconfirmed item is cited by a confirmed-section claim |
brandreport report |
renders findings + snapshot into the press-styled HTML brand report, fully offline — refuses to render if gate fails |
Rules that are not negotiable
- Never attribute unverified content: nothing reaches the report unless it is tied to the actual person by a recorded corroborating signal (cross-links, shared handles, bio matches) — same-name findings it cannot tie are listed as unconfirmed, never silently included and never silently dropped.
- Never claim a result you did not observe. Say what you verified and what you did not.
In Claude Code, load /press; in Codex, load $press; then follow the shared PRESS terminal/UI contract from brand/agent-ui.md. Do not copy or override that contract here.