| name | fix-call-typos |
| description | Fix transcript/tldr typos for a published Forkcast call AND update the upstream ACDbot vocab so the same typo doesn't recur. Use when a call page on forkcast.org has wrong client/protocol/term names that the pipeline missed. |
Fix typos in a published call
This skill walks the user through correcting typos in a call's artifacts on Forkcast (transcript, TLDR, key_decisions) and teaching the upstream ACDbot pipeline at ethereum/pm to catch those typos automatically next time.
The "correct" approach would be to fix upstream and re-run the pipeline. In practice we patch the artifacts in both repos and update the vocab — that way published assets are right today, and future pipeline runs improve.
See docs/acdbot-forkcast-asset-pipeline.md for the full asset pipeline overview.
Step 0: Locate both repos
This skill is invoked from inside a Forkcast checkout (or a worktree of one). It also needs the PM repo (ethereum/pm).
- Forkcast root — resolve via
git worktree list --porcelain | head -2 | awk '/^worktree/ {print $2; exit}' from the current directory. This returns the main checkout path even when invoked from a worktree.
- PM root — first try the sibling next to forkcast:
<dirname-of-forkcast-root>/pm. Verify it's a clone of ethereum/pm:
git -C <candidate> remote get-url origin
Expect a URL ending in ethereum/pm (HTTPS or SSH). If the sibling check fails, ask the user via AskUserQuestion for the absolute path to their PM clone, then validate the same way.
- Refer to these throughout the rest of the skill as
<forkcast> and <pm>. Don't hardcode them in commands; resolve once at the start and reuse.
Then open a worktree off origin/main on the forkcast side so the user's current branch isn't disturbed (git -C <forkcast> fetch origin main && git -C <forkcast> worktree add <tmp-path> origin/main). Work in PM directly — pull master first (git -C <pm> checkout master && git -C <pm> pull --ff-only).
Step 1: Identify the call
Ask the user for:
- Call slug + number (e.g.,
pqi/038, acdc/177). This determines:
- Forkcast path:
<forkcast>/public/artifacts/{slug}/{date}_{number_padded}/
- PM path:
<pm>/.github/ACDbot/artifacts/{pm_series}/{date}_{number_padded}/ — the PM series name often differs from the forkcast slug. Check SERIES_TO_TYPE in <forkcast>/scripts/sync-call-assets.mjs for the mapping (e.g., pqi ↔ pqinterop, price ↔ glamsterdamrepricings). Core series (acdc, acde, acdt) pass through unchanged.
Resolve the absolute paths and confirm both directories exist before continuing.
Step 2: Gather corrections from the user
Ask for two lists:
- New vocabulary — terms the LLM didn't know (client names, protocol names, EIP nicknames). Each term needs a canonical casing (e.g.,
Ethlambda, not ETH Lambda or ETHLambda).
- Specific typo → correct mappings — every variant the user has spotted, plus its target.
Don't pre-read the transcript yet. Let the user tell you what's wrong first; their list anchors the work.
Step 3: Survey the transcript for related variants
Once you have the user's list, grep the call's transcript_corrected.vtt, tldr.json, and (if present) key_decisions.json for:
- The user's exact typos (to confirm hits and count occurrences)
- Other plausible variants of the same words. Examples seen in the wild:
- Casing drift:
ETHLambda, ETH Lambda, ETH lambda, ETLambda all → Ethlambda
- Vowel/letter drift:
Zeem, Zeeam, Zem, ZEM all → Zeam; Rheem, Reem, Dream → Ream
- Speaker-attributed name confusion: a speaker labeled
| gean referring to "Jim" almost certainly means Gean
Surface the additional variants to the user with AskUserQuestion. For every variant ask: "is this the same word?" Don't assume.
Also ask the user the canonical casing for any term that appears in multiple forms (e.g., Ethlambda vs ETH Lambda). Their answer goes into the vocab.
Step 4: Decide what's safe to add to the global vocab
ethereum_vocab.yaml patterns are applied as global find/replace by the LLM on every future call. So an entry like Jim: Gean would garble any call where someone named Jim speaks. Same for common English words (Dream → Ream).
Before adding to the vocab, ask: would this pattern, applied globally, ever break a different call's transcript? If yes, fix it only in the specific call — don't add it to the vocab. If no (the source token is rare and unambiguous), add it.
Generally safe: invented-looking strings (Zeem, QLEAN, Gein, ETHLambda).
Generally unsafe: real names (Jim), common English words (Dream), short ambiguous fragments.
Step 5: Update PM ethereum_vocab.yaml
File: <pm>/.github/ACDbot/scripts/asset_pipeline/ethereum_vocab.yaml.
- New vocab terms — add to the relevant section (
clients.execution, clients.consensus, clients.lean, acronyms, technical_terms, etc.). If a new category is needed (e.g., a new ecosystem of clients), add a new subsection.
- Safe error patterns — append to
error_patterns: as - WrongForm: RightForm.
- Order longer patterns before shorter ones with the same prefix (e.g.,
Zeeam before Zem) so substring matches don't misfire.
- Existing patterns are mixed-case; match the casing of the source token you want to catch.
Step 6: Fix the artifacts in Forkcast
In the forkcast worktree, edit <forkcast>/public/artifacts/{slug}/{date}_{number}/:
transcript_corrected.vtt — apply every typo correction. For multi-token corruptions (e.g., ReamZeem eats Lambda for what was meant as "Ream, Zeam, Ethlambda"), do the long unique replacement first, then run the single-word substitutions. Use Edit with replace_all: true for safe global swaps; use scoped Edit for one-line fixes (like the Jim → Gean case).tldr.json — apply the same replacements where applicable.key_decisions.json (ACD calls only) — apply the same replacements. This file references EIP numbers; if a typo named an EIP, also verify the eips: array.transcript_changelog.tsv — this records what the LLM changed from the original. If a row's corrected value was wrong (e.g., Zim → Zeem when the right answer was Zeam), update the target column. Don't add new rows for manual corrections — the changelog is documentation of pipeline output, not of post-hoc edits. (Past examples: PM commit f7f549b3 removed an over-correction row entirely.)
After editing, grep the file again for any remaining instance of every typo string, including the ones the user gave you. Output should be empty.
Step 7: Mirror the fixes into PM
The PM repo holds the upstream copy of these same artifacts at <pm>/.github/ACDbot/artifacts/{pm_series}/{date}_{number}/. Before your edits these files were byte-identical to the forkcast copies (they're produced by the pipeline and the sync script just downloads them). The simplest correct move is to cp the four edited files (vtt, tldr, changelog, key_decisions if it exists) from <forkcast> to <pm>.
Verify with diff afterwards — both copies should be identical.
Step 8: Commit and open PRs
Two PRs, two commit-message styles (mirror prior commits — see PM f7f549b3, forkcast 37db68a3 for examples):
PM PR — branch like acdbot-fix-{slug}-{number}. The commit covers:
Forkcast PR — branch like fix-{slug}-{number}-typos. Single commit covering the artifact fixes only (no pipeline changes here).
Title: fix {SLUG} {number} {topic} typos or fix: normalize {topic} in {SLUG} {number} summary.
For both, do NOT use #N for call numbers in PR title/body — GitHub treats it as an issue tag. Write {SLUG} {number} (e.g., PQI 038).
Step 9: Cleanup
Once both PRs are open, remove the forkcast worktree (git worktree remove --force). Tell the user to merge both PRs (PM first is fine, order doesn't matter — the forkcast fix is a static patch, and the next pipeline run for future calls picks up the new vocab).
Step 10: Done
Confirm to the user that the call page on forkcast.org will show the corrections after the forkcast PR's deploy completes, and that future calls will be auto-corrected by the pipeline using the new vocab entries.
