docs: prompt-packs plan — declarative extensions and a self-extending Chat - #625
Open
kcarnold wants to merge 3 commits into
Open
docs: prompt-packs plan — declarative extensions and a self-extending Chat#625kcarnold wants to merge 3 commits into
kcarnold wants to merge 3 commits into
Conversation
… Chat Design doc for the lowest tier of the extensibility ladder: extensions as prompt packs (prompt + tool grants, inert data) executed by the existing UI, with Chat as the conversational authoring environment via save_pack / test_pack tools. Includes the runtime tool vocabulary table and a coverage analysis showing ~80% of the sketched concepts (interface-concepts.md, interaction-concepts.md) reduce to packs given ~10 generic tools; the rest need own rendering surfaces and stay in the hosted-tool tier (tool-launcher-plan.md, now cross-referenced). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01EUSXFcRA7ox4u9x1QENyHw
…pace; add MCP analysis The memory/history tools (remember, recall, read_history, get_revisions, search_corpus) were reinventing a filesystem and git. Replaced with a per-(user, pack) virtual workspace: files, grep, and git history, whose idiom the model already knows and whose contents the writer can read, export and delete. Names just-bash / just-git as candidates, with the undici network dependency flagged as due diligence rather than assumed away, and states the covenant boundary explicitly: the document is not a writable file in the workspace, and the workspace has no network. Adds a build-vs-reuse section (W3C Web Annotation TextQuoteSelector for the phrase-anchoring issues, OAuth device flow as precedent, XMPP's model without its wire) and a section on the three roles MCP could play here: outbound client to third-party servers as the cheap unlock, MCP-shaped internal tool definitions, and an inbound server as the powerful but values-fraught case that tests whether the covenant belongs to our UI or to the platform. Also sharpens the ask_writer rationale to floor control, non-chat surfaces, and typed study data, rather than the ability to ask at all. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01EUSXFcRA7ox4u9x1QENyHw
…ost, not servers they run The doc's founding question is whether a domain expert who is not a software engineer can extend Thoughtful, and its own distribution story failed that test: 'a writing center publishes an MCP server' assumes a public HTTPS endpoint, multi-tenant auth and someone on call — more engineering than the monorepo PR the ladder exists to avoid. Adds a 'Who can actually publish?' section with the principle that LLMs made authoring cheap and operating no cheaper, a table of what skill each distribution mechanism actually demands, and the rule that anything asked of a domain expert must be something we host. Names collections (shared sets of packs on our infrastructure, Google-Docs-shaped) as the real distribution primitive, with author-hosted static URLs as an option rather than a requirement, and states what hosting distribution obliges us to take on — provenance, reporting, moderation — since that cost is what made self-hosting tempting to write down. Repositions MCP accordingly: the writing center is a consumer of servers others already run, never an operator. Collections added to the phasing. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01EUSXFcRA7ox4u9x1QENyHw
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Adds
docs/prompt-packs-plan.md, a design doc for the lowest tier of the extensibility ladder, grown out of a discussion of what it would take for Thoughtful to be extensible by people who aren't us — including domain experts who aren't software engineers.What's in it
button,persona,lens,ritual), executed by the manual tool-loop pattern already proven inmy-words/interaction/liveResponder.ts.offer/ask_writer/highlight/annotatecover most,remember/recalladd the stateful ones,read_history/get_revisions/search_corpusunlock the rest. The honest 20% (Loom, mindmap, voice, Skim View, Inkwell) needs its own rendering surface or transport and belongs in the hosted-tool tier.save_pack/test_pack/list_packs, so a writer can turn a conversational behavior into a persistent pack by asking ("make that a button"), with provenance and full-prompt legibility as the trust story.view/highlightinto Chat), Send hosts ranges, not needles: one text-matching implementation across editors #589/#591/#592`` (phrase matching thathighlight/`propose_edit` lean on), #615 (consent gate that `read_history` depends on).Also adds a short cross-reference at the top of
tool-launcher-plan.mdmarking where the pack tier ends and the hosted-tool tier begins.Docs only — no code changes.
🤖 Generated with Claude Code
https://claude.ai/code/session_01EUSXFcRA7ox4u9x1QENyHw
Generated by Claude Code