Readme

readme — the per-topic navigation index

Every report topic carries a navigation README.md at reports/<topic>/README.md, modeled on a research corpus's per-directory READMEs. Its structure:

# <Title>

**Research ID:** <slug>
**Created:** <date> | **Updated:** <date>
**Findings:** <N> (survived S, weakened W) | **Sources:** <N> unique URLs
**Falsification:** <date> — survived S, weakened W[, quarantined Q] ([report](…))   ← when a falsification report exists
**Status:** <status>

---

![<title>](_assets/<topic>-readme-hero.*)   ← optional, only when the hero asset exists

## Purpose            — 1–2 sentences, what the session decides/answers
## Dimensions         — bulleted; each dimension with its harness.config description
## Key Findings       — 4–10 SYNTHESIS-GRADE bullets (see below)
## Reports            — Type → Title table of the topic's constituents, in reader-consumption order
## Findings by Dimension — table of per-dimension counts
## Artifacts          — (only if channel-pack assets exist) File / Type / Size table
## Tags               — backtick-quoted tag tokens

The README is the served topic index: the site's content loader derives its Starlight title from the H1, so it renders as the topic's landing page at /reports/<topic>/, with the constituent report tree reached from its Reports table — the sidebar links only to this index, not the per-report tree (ADR-0009). It remains a navigation projection, not a MIF Level-3 report: like the blog/book channels it carries no MIF frontmatter and is exempt from the output-conformance gate. Its source of truth is the MIF substrate (reports/<topic>/findings/*.json, goal.json, harness.config.json). The Reports table is ordered by a single canonical Type map in scripts/build-topic-readme.sh (executive summary → briefing → synthesis → genre reports → falsification report → research progress).

When this runs

This skill wraps scripts/build-topic-readme.sh — the deterministic structural engine — and is responsible for the one thing the script cannot do: author synthesis-grade Key Findings.

• At synthesis: the report-synthesizer agent (which holds the surviving findings) writes the synthesis-grade Key Findings as part of Phase 4. That is the primary path — the README ships synthesis-grade, not skeletal.
• In the orchestration flow: /start invokes this skill after the orchestrator returns, to (re)synthesize prose and validate.
• On demand / safety net: run it any time; the check-readme-reindex PostToolUse hook flags out-of-band edits.

Pipeline

1. Resolve the topic set. --topic <id> (one), --all (every harness.config.json topic), or the active session's topic when no flag.

2. Build the deterministic backbone. Computes every count, date, the verdict breakdown, the source total, dimension rollup, and the report/artifact tables — never author these by hand:

   bash scripts/build-topic-readme.sh <topic>
   

3. Author synthesis-grade Key Findings and Purpose (the core of this skill). First Read reports/<topic>/README.md — the file step 2 just generated; the Edit tool refuses to write a file it has not read this session. Then read the actual findings (reports/<topic>/findings/*.json — title, summary, content, citations, verdict) and rewrite the ## Key Findings section so it reads like the corpus exemplar, then tighten ## Purpose to 1–2 sentences. Edit only those two sections; the script preserves them on later rebuilds.

   Key Findings bar — match this, do NOT regress to a list of finding titles:

   • 4–10 bullets that synthesize ACROSS findings — each bullet states an insight, a tension, or a converging consensus, not a single finding restated.
   • Carry the specifics that make it credible: named standards/tools/products, numbers, dates, versions, thresholds — drawn from the findings' content.
   • Lead with the decision-relevant signal (what the evidence establishes, where it is thin, what the call turns on), ordered by the goal's priorities.
   • Respect verdict nuance: weakened findings carry caveats; inconclusive ones are reported as open, never asserted; falsified ones are excluded.
   • Every bullet must trace to surviving finding content — invent nothing.

   Example of the bar (one bullet from the corpus):

   A credible dual-format consensus is converging: RFC 9457 Problem Details as the transport envelope, MCP-style structuredContent + isError as the machine channel, miette/rustc-style human rendering with primary span and cause chain… — shipped-and-proven, not aspirational.

   Contrast — the skeletal output this skill must REPLACE:

   • Copier propagates upstream template changes to instantiated projects.

4. Validate (the gate). Must pass before reporting done:

   bash scripts/build-topic-readme.sh <topic> --check
   

   Non-zero exit means the README is wrong — fix it, do not ship it. Markdown lint is enforced separately by the bundled md_guard hook on write.

5. --all: loop steps 2–4 over every topic.

Non-negotiables

• Key Findings are synthesized, not listed. A README whose Key Findings are one-finding-per-bullet restatements has failed this skill — re-synthesize.
• The README never carries MIF frontmatter and never claims a verdict (that is publish-report).
• Counts, dates, verdict breakdown, tables, and tags are deterministic — from the script, never invented.
• The --check gate must pass before the README is reported reconciled.
• --topic / --all only — never hardcode a topic id; this is a template.