| name | guide-ingest |
| description | Ingests offline deep-research exports into README guides for kokoro-coreml: cleans markdown, corrects library/API claims with Context7 when docs contradict the draft, annotates firsthand repo contradictions without stripping useful “worth trying” external synthesis, and adds bidirectional cross-links in README/Guides and README/Notes. Use for guide-ingest, research-to-guide, or corpus linking of a new guide from Deep Research / browser export. |
Guide ingest
Purpose
Turn raw external research into a durable repo guide: a synthetic library
entry that captures what the internet says about a domain (the “afternoon of
Googling” artifact from create-guide).
kokoro-coreml runs Kokoro TTS on Apple Silicon via Core ML / ANE. Guides stock
external mechanism knowledge (Core ML conversion, ANE scheduling, Metal/MPS,
iSTFT, Swift/Core ML glue). They are not campaign go/no-go decisions.
Ingest must:
- Match this repository’s markdown conventions
(
README/Guides/content/markdown-authoring-guide.md).
- Correct documented library/API facts when Context7 (or other primary
docs) contradict the draft.
- Annotate contradictions from firsthand local knowledge — do not silently
rewrite the guide into a campaign memo, and do not delete useful external
synthesis just because it is unproven.
- Sit correctly in the bidirectional web of
README/Guides/ and
README/Notes/.
A guide remains an external-knowledge artifact, not a dial-a-friend verdict
for the current plan. Campaign go/no-go, measured latency/quality tables, and “do not launch
arm X” decisions belong in README/Notes/ (link them); keep the guide reusable
as domain synthesis.
Use When
- Importing deep-research (or similar) output into
README/Guides/....
- The user invokes guide-ingest, research-to-guide, or the same workflow
in natural language.
- A new guide needs outbound links to notes/guides and inbound links from existing docs.
Do Not Use When
- The task is only a small markdown typo fix (use
markdown).
- The task is only where to put a note (use
write-notes).
- The work is inline code documentation (use
documentation).
- There is no external raw artifact — stop and write a note instead of inventing
a guide from local reasoning alone.
Prerequisites
- Context7 MCP available for
resolve-library-id / query-docs (or the
repo’s equivalent Context7 tools). Read the tool schema in the MCP
descriptors before calling tools.
- Raw source: path to
raw-report.md (or equivalent browser-downloaded export).
Procedure
0. Provenance gate
- Confirm an external raw source path exists (usually
llm-workflows/outputs/create-guide/**/raw-report.md).
- Record that path (and SHA-256 when practical) near the top of the ingested
guide.
- If there is no raw external artifact, stop — use
write-notes, do
not mint a guide from agent synthesis alone.
1. Normalize to proper markdown
- Read the
markdown skill and the guide index
README/Guides/content/markdown-authoring-guide.md
(and README/Templates/guide-template.md when adding a new file)
for repo tone and structure.
- Produce guide-shaped markdown:
- real markdown links, not bare URLs
- blank lines around headings and lists
- language-tagged fences where applicable
- single trailing newline; no unnecessary HTML
- for messy exports: unescape punctuation, strip data-URI images / broken
image blocks, and repair Swift/code fences without inventing missing facts
- Add a short top-of-file purpose blurb if the template or sibling guides use
one; match the tone of nearby guides in the same folder.
- Strip or refuse campaign-advice framing from the draft (“you should kill this
arm”, “no-go for your bakeoff”) when it is clearly advice to this repo’s
plan. Preserve the underlying external mechanism, ratios, failure modes, and
recipes. Put the campaign decision in a note and link it.
2. Correct vs annotate (do not over-prune)
Guides are allowed to include plausible, unproven, “worth trying” material
from the public internet. Ingest is not a demand for 100% validation before a
claim may appear.
| Situation | Action |
|---|
| Context7 / official docs contradict a library API, flag, default, or deprecation | Correct the guide to match current docs |
| Obvious hallucination (invented API, fake flag, fabricated citation) | Remove or rewrite; note the rejection briefly if useful |
| Community heuristic / paper recipe / “people report X” with no local proof | Keep, labeled as community / secondary / worth-trying |
| Firsthand local measurement or pinned-stack run contradicts the draft | Annotate in-place (correction callout + link to the note/receipt); do not delete the external claim without saying why |
| Claim outside Context7 (Core ML / ANE scheduling, Metal, Swift glue) | Prefer primary docs / man pages / source pins when checking; else keep labeled, do not fake “verified” |
Procedure:
- Extract verifiable library/API claims and check them with Context7
(resolve library ID, then query docs). Prefer current official behavior for
those factual corrections.
- Scan
README/Notes/ and related guides for firsthand contradictions
(measured runs, pinned-stack corrections). Add explicit callouts such as
“Local evidence (link): … contradicts the draft’s claim that …”.
- Do not gut sections solely because they lack a Context7 hit or a local
A/B test. Unverified ≠ worthless for a synthetic external library.
- Do not rewrite the guide into a qualification-campaign status document
during ingest. Link status notes; keep the guide domain-scoped.
3. Add outbound links (guides and notes)
- While reading sections, identify concepts that already have a home in
README/Guides/ or README/Notes/.
- Add inline markdown links on the first strong mention in each section (or
a compact “Related Documentation” / “See also” list if inline would clutter).
Use repo-relative paths from the new guide file.
- Prefer linking to one canonical guide per topic rather than duplicating
long explanations.
- When local campaign decisions exist, link them as related notes, not as
the guide’s thesis.
4. Add inbound links (corpus updates)
Goal: related docs point back to the new guide so agents discover it from
both directions.
- Discover candidates: search
README/Guides/ and README/Notes/ for
overlapping keywords, library names, and headings (e.g. ripgrep). Include files that already link to adjacent topics.
- Edit sparingly: add a link in “Related”, “See also”, or the most
relevant paragraph—minimal diff, no drive-by rewrites.
- If a high-level notes file covers the same subsystem, add a short bullet or
sentence there with a link to the new guide (see
write-notes for
consolidation habits).
- Update
README/Wiki/canonical-source-coverage.md
(and README/Wiki/README.md when needed) when adding a new guide file.
5. Close the loop
- Rely on the
markdown skill for structure; run pytest only if the change overlaps Python/runtime code.
- Give the user a short summary:
- provenance path recorded
- API/docs corrections (Context7 or primary)
- firsthand contradiction annotations (and note links)
- what was kept as unlabeled or labeled “worth trying”
- inbound link files touched
Handoff Rules
| Situation | Hand off to |
|---|
| Repo markdown rules, lint, structure | markdown |
| Campaign decision / measurement write-up | write-notes |
| Docstrings / inline code documentation | documentation |
| Need a new external synthesis run | create-guide |
| Advice / second opinion on what to do next | second-opinion |
References