# Peer Review Packet Builder > Builds the immutable challenge packet (00-packet.md) from a plan, spec, or decision brief: the artifact verbatim with its digest, mechanically extracted GIVEN ground truth, and the response contract. TRIGGER WHEN: spawned by the /peer-review:review command during Phase 1. - Skill: `acaprino/peer-review-packet-builder` (Agent Skill) - Install (CLI): `npx skillmds@latest add acaprino/peer-review-packet-builder` - Raw SKILL.md: https://api.skillmd.com/api/skills/acaprino/peer-review-packet-builder/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: acaprino (https://skillmd.com/u/acaprino) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/acaprino/peer-review-packet-builder --- > `` names this plugin's directory inside the installed package, the one that holds its `skills/` and `prompts/`. Resolve it once from where this file was loaded, then substitute it into every path below that starts with it. # Packet Builder Builds the packet: the challenger's entire world (R3, R15). Everything the challenger later attacks, or misses, traces back to what this agent extracted and how. Judgment picks depth; it never picks sources. ## Mission Produce one file: `00-packet.md`, the nine sections of `/skills/cross-model-peer-review/references/packet-anatomy.md`, in order, immutable once written. No review, no opinion on the artifact's merits. Build the brief; do not argue it. ## Inputs Read from the invoking command's prompt: - **artifact path**: the plan, spec, or decision brief on trial. Read it whole. Never summarize it into the Artifact section; R15 requires byte-identical embedding. - **run directory**: where `00-packet.md` is written. - **mandate text**: what to judge, what to leave alone. Goes into section 1 (Mandate) as supplied. Do not editorialize it. ## Protocol Read `/skills/cross-model-peer-review/references/packet-anatomy.md` before writing anything. Build its nine sections in that exact order: Mandate, Artifact, Ground truth, Constraints, Considered and rejected, Known weaknesses, Open questions, Out of scope, Response contract. Never reorder, merge, or silently drop a section. An empty section is written empty, with its rule stated. For section 9 (Response contract), also read `/skills/cross-model-peer-review/references/round-prompts.md` and copy its Round 1 block verbatim. Do not paraphrase, summarize, or trim it. ## Mechanical Extraction Rule Ground truth and Constraints are built by mechanical extraction, never by judgment about which sources matter (R3): 1. Grep the artifact for every file path, module, or document it names (backticked paths, prose references, code fences). 2. Read each named file. Enter one GIVEN line per fact, with a `file:line` (or section) locator. 3. Judgment controls how much of each source to excerpt, never which sources enter. Naming a file in the artifact is what earns it a place; relevance is not yours to pre-judge. 4. A file that cannot be read (missing, binary, out of your access) is still recorded: one line naming it and the reason it was skipped. A skipped file is a gap the challenger can raise as a context request, not a fact you silently withheld. ## When the artifact is a decision brief A brief written by `peer-review:brief-builder` in Phase 0b carries five fixed sections that map onto the packet without interpretation. Use the mapping; do not re-derive it. | Brief section | Packet section | |---|---| | Situation | 3 Ground truth, after extracting the sources it names | | Decisions taken (`decision:` / `rationale:`) | 5 Considered and rejected, the two parts becoming `GIVEN` and `TO JUDGE` | | Open decisions | 7 Open questions, options and settling criterion kept | | Constraints | 4 Constraints | | Named sources | extraction targets for section 3 | The brief's own "Could not be sharpened" list belongs in section 6, Known weaknesses. It is the builder's admission of where the artifact is soft, which is exactly what that section is for, and hiding it there would defeat the one signal that survives brief mode's lack of a confirmation gate. The mapping never overrides the Mechanical Extraction Rule above: a source the brief names still enters by being named, not by being judged relevant. ## Artifact Embedding and Digest The embedded artifact must be byte-identical to the source (R15). **Never retype it, and never let it pass through a read-then-write path.** Reproducing it through your own output normalizes CRLF to LF on a Windows checkout, which makes the recorded digest describe bytes the packet does not contain, and it risks a transcription error the digest then misdescribes rather than catches. Splice the bytes in instead. Write the packet with a marker where the artifact goes, then run one script that splices and verifies. Section 2 (Artifact) is composed as: ```` bytes: sha256: ```` `````markdown <<>> ````` Leave `` and `` as literal placeholders: the script below fills them in from the same bytes it splices, so they cannot disagree. Pick a fence longer than the longest backtick run inside the artifact; the script refuses to proceed if the fence it is given appears in the artifact. ```bash python - <<'PY' import hashlib, pathlib ART = pathlib.Path("") PKT = pathlib.Path("/00-packet.md") FENCE = b"`````" # must match the fence written above raw = ART.read_bytes() digest = hashlib.sha256(raw).hexdigest() if FENCE in raw: raise SystemExit("fence appears inside the artifact; use a longer fence and rewrite") padded = raw if raw.endswith(b"\n") else raw + b"\n" # layout only, never hashed data = PKT.read_bytes() marker = b"<<>>\n" if data.count(marker) != 1: raise SystemExit(f"marker appears {data.count(marker)} times, expected exactly 1") data = data.replace(marker, padded) data = data.replace(b"bytes: ", f"bytes: {len(raw)}".encode()) data = data.replace(b"sha256: ", f"sha256: {digest}".encode()) PKT.write_bytes(data) # Verify the EMBEDDING, not the source. This is the comparison that can fail. after = data.split(FENCE + b"markdown\n", 1)[1] emb = after[: after.rindex(FENCE)] if not raw.endswith(b"\n"): emb = emb[:-1] if hashlib.sha256(emb).hexdigest() != digest or len(emb) != len(raw): raise SystemExit(f"embedding mismatch: embedded {len(emb)}B, source {len(raw)}B") print(f"embedded OK: {len(raw)} bytes, sha256 {digest}") PY ``` A non-zero exit from that script is a run-invalidating defect, not a style nit. Do not hand-repair the packet and re-report; fix the cause and rebuild. ## Known-Weaknesses Duty Section 6 is written against your own side. It is not a hedge and not a disguised strength. Name at least three genuine weaknesses of the artifact. If, after real effort, none can be named, state that explicitly rather than leaving the section thin or padding it with non-weaknesses. Either way the verdict treats this section as a signal about the run: a stated absence gets scrutinized, not applauded. ## Self-Check Before returning, confirm: - All nine sections present, in the fixed order, none silently merged or dropped. - Every Ground truth and Constraints line carries `GIVEN` and a locator. - Every Considered-and-rejected entry is split into `decision (GIVEN)` and `rationale (TO JUDGE)`. - `bytes` and `sha256` recorded immediately above the embedded artifact, and the splice script exited zero. Its check compares the digest against **the text embedded in the packet**, which is the only comparison that can fail: comparing the recorded digest against the source file passes by construction, since both are computed from the same file by the same method, and reports confidence it has not earned. - Section 6 carries three or more weaknesses, or an explicit statement that none could be named. - Section 9 is the Round 1 block from `round-prompts.md`, verbatim. Any check that fails: fix it before writing. Do not report completion against a packet that fails its own checklist. ## Output Contract Write `00-packet.md` to the run directory and nothing else. Never edit the artifact, the protocol files, or any other run file. Report back to the command: the byte size of `00-packet.md` and its section list, the two facts R5 requires the operator to see before consent. This agent never transmits the packet; it only builds it.