recursive-growth
暂无简介
- License 许可
- MIT
- 最近更新
- 2026-11-15
- Stars
- ★ 0
- Forks
- 0
02 资源文件 2
- 参考文档LESSONS.md781 B
- 参考文档SKILL_CHANGELOG.md1007 B
03 SKILL.md 正文
recursive-growth — session-start recall, and a skill that keeps correcting itself
Three founder directives, one skill: (1) every session should be pointed straight at the
master plan, not left to re-derive "what's next"; (2) whatever fable-debugger learns from
a real bug should make the hive smarter next time, not evaporate at session's end; (3) skills
themselves should get corrected, in their own files, as their own use turns up something
stale — not just accumulate lessons in a side log nobody rereads.
Part 1 — Session-start recall (run this first, every session)
cat memory/planning/CLAUDE.md # confirms the current pointer
cat memory/planning/2026-07-19-unified-forward-plan.md # the actual plan
Report back plainly, before doing anything else: which phase is next (per the plan's own
"Phase 0" → "Phase 7" ordering), what's already closed, and what's explicitly founder-blocked
(Phase 5 — never scheduled as a task, only tracked). This is the same discipline
hive-conductor's Phase 0 RECALL already applies to sovereign D1/Vectorize memory
(scripts/recall_context.py) — this skill is that same idea pointed at the one committed
plan file instead of a network-backed index, so it works even with zero egress, and it names
the concrete next phase rather than a fuzzy "prior related work" brief. Run both; they answer
different questions (RECALL: "has the hive done something like this before anywhere in its
history?" — this skill: "where exactly are we in the one plan we already agreed to?").
If memory/planning/CLAUDE.md's pointer and the Founders Visionary Folder's own pointer
(Project_file/Founders Visonary Folder/ACTIVE/2026-07-19-current-master-plan-pointer.md)
ever disagree about which file is current, that disagreement is itself the thing to fix
first — the two lineages pointing at two different plans is exactly the reconciliation gap
this plan's own Phase 1c exists to close. Don't silently pick one; say so and fix the stale
pointer.
Part 2 — Growing from fable-debugger (recursive learning from bugs)
Every time fable-debugger's loop reaches step 6/7 (VERIFY-REAL, then CLOSE — a fix that
actually proved itself against a real check, not an escalation) do two things:
- Always: append one entry to
LESSONS.mdin this skill's own directory — symptom, root cause (one sentence, per fable-debugger's own step 3 discipline), the fix, and the verification command that proved it. This is the raw, unfiltered log — append-only, never pruned, never summarized away. Every entry should be real (a fix that actually shipped), never a hypothetical "here's a bug that could happen." - Only on genuine recurrence (the same root-cause shape appears a second time, possibly in a different file): promote it into this file's own Recurring Patterns section below, generalized past the specific file/symptom into the reusable lesson. This is the actual "recursively builds and grows" mechanic — the skill's own body of knowledge gets denser over time, but only with things that proved themselves twice, not every one-off. A skill that grew by appending every single bug would become noise nobody reads; a skill that only grows on real recurrence stays worth reading years in.
Recurring Patterns (empty until a lesson genuinely recurs — do not seed this with
guesses; the first entries here should be real, already-observed repeats)
(none yet — see LESSONS.md for the raw log this section promotes from)
Part 3 — Optimizing other skills' own files, after they run
After any skill in .claude/skills/ completes a real run (not a dry read), do one light
pass, bounded by hard rules:
- What counts as an optimization: a stale example that no longer matches the repo (a
renamed file, a moved endpoint), a broken script path, a described behavior the run just
proved wrong or incomplete, a missing cross-reference to a skill that's clearly relevant
(the way this section itself should have existed the first time
fable-debuggerandhive-conductorwere written — see the changelog entries this skill added to both on 2026-07-19, the first real proof of this mechanic, not just a description of it). - What never counts: rewriting a skill's core contract, loosening a documented hard
boundary (
threat-sandbox's "never absorb the payload,"session-harvest's ownership/license gate, anyPERMISSIONS.mdtier boundary,agent-harness's never-adjudicate-your-own-verification rule), or any edit that isn't traceable to something this run actually observed. When in doubt, don't edit — log it inLESSONS.mdas a candidate instead and let it recur before acting. - Every edit is small and logged: one paragraph or one line changed, a one-line entry
appended to
SKILL_CHANGELOG.mdin this directory naming which skill, what changed, and why (the observed evidence), and committed normally — never silent, uncommitted drift. This is Tier 1 underPERMISSIONS.md(internal, reversible via git, no external party, no money) — but the changelog is what keeps it honest under F-004 (explainability): anyone cangit log/git blamea skill file and find the reason a line changed. - This skill is not exempt from itself: when running
recursive-growthitself turns up something stale in this very file (a broken path, a Recurring Pattern that turns out wrong), the same rule applies — small edit, changelog entry, no silent rewrite.
Relationship to the hive's other memory/growth mechanisms (so nothing here duplicates them)
hive-conductor's Phase 0 RECALL — queries sovereign D1/Vectorize memory for "has the hive done something like this before, anywhere." Network-dependent, degrades gracefully. This skill is its local, always-on sibling, scoped to one committed plan file.fable-debugger— the method that produces the lessons this skill grows from. This skill doesn't debug; it remembers what debugging taught, and corrects the skills involved.pr-retrospective/PR_LESSONS.md— PR-outcome-specific forward checks (what to watch for on the next PR touching a risky surface). This skill'sLESSONS.mdis broader (any verified fable-debugger fix, not just PR-shaped ones) and feeds skill-file corrections, not just a pre-flight checklist.session-harvest— distills a whole session's work intoHIVE_UPDATES/at close, under an ownership/license gate. This skill runs continuously within a session (every fable-debugger close, every skill run), and writes to skill files and the lessons log, not to the founder-facing updates channel.skill-census— audits whether skills are wired/orphaned, periodically. This skill is what keeps a skill's content accurate between census runs, not what discovers whether a skill is used at all.
Quick start
# 0. Session start — read the plan, report the next phase, before anything else:
cat memory/planning/CLAUDE.md
cat memory/planning/2026-07-19-unified-forward-plan.md
# 1. After a fable-debugger close — append the raw lesson:
# (edit .claude/skills/recursive-growth/LESSONS.md — symptom / root cause / fix / verification)
# 2. If that lesson's shape has appeared before — promote it into this file's
# "Recurring Patterns" section, generalized past the one instance.
# 3. After any skill runs — if something in ITS OWN files is now stale, make one small,
# logged edit and append one line to SKILL_CHANGELOG.md naming what/why.
Origin: founder directive, 2026-07-19 — "refer directly to this plan at the beginning of each session," "recursively builds and grows... any time fable 5 debugger is r[u]n," "learn recursively from any and all the bugs," "fully optimize any other skill... directly into their file[s]." Built the same way every other skill here was: a real, repeatable method, written down so any session can run it — not a promise, a discipline.
04 同仓库更多 Skills
preamble-gate
Use BEFORE doing any work in this repo — before the first file edit, commit, push, PR, branch operation, workflow dispatch, scheduled trigger or timer, agent/subagent spawn, dependency install, or message sent to an outside service. Forces a plain-language preamble to the founder first: WHAT you are about to do, WHY, HOW, why it is NECESSARY, the PROJECTED IMPROVEMENT, and the COST and RISK — then, for consequential work (governance, security, money, CI/triggers, rewriting history, multi-agent fan-out), waits for an explicit go-ahead. Read-only investigation (reading files, grep, git log/diff, answering questions) is exempt. Applies to every AI that works in this repo, not just Claude. Founder rule, 2026-11-10, written after an unattended automation ran up cost and froze the hive. Not a substitute for wired-or-not (which governs how completion is CLAIMED); this governs what happens BEFORE work starts.
visionary-recommender
暂无简介
syllabus
Generates a curated supplementary reading list from any course syllabus using Consensus academic search. Grill-me intake (syllabus input format + course audience + year range) plus a grouping forcing-options checkpoint before any search runs — so the reading list matches the course's level and recency need. Parses the syllabus to extract topics and learning outcomes, searches Consensus for recent peer-reviewed papers per topic, and produces a professionally formatted .docx with clickable Consensus links, plain-language summaries calibrated to audience level, and Bloom-higher-order discussion questions tied to course learning goals. Use when the user uploads a syllabus, course outline, or curriculum document and wants supplementary readings (e.g., 'create a reading list from this syllabus', 'find recent papers for my course') — even casual mentions with a syllabus attached should trigger this skill.
research
Default entry point for any research request — a hybrid router that classifies the question deterministically and either delegates to a specialist research skill (pulse for trends/sentiment, grants for NIH funding, litreview for academic literature, syllabus for course reading, patent for prior-art + IP landscape, dossier for entity research) or runs its own plan-decompose-multi-source-search-synthesize-cite fallback workflow when no specialist matches. Always surfaces the routing decision so users can override. Use when the user makes any research request that doesn't obviously match a more-specific specialist skill (e.g., "research [topic]", "look into [topic]", "what do we know about [topic]", "investigate [topic]", "find me information on [topic]", "do some research on [topic]", "I need to understand [topic]"). Output is a markdown briefing (default) or .docx document (on request) with full citations and an audit log.
pulse
Multi-source recency research skill that takes the pulse of any topic across Reddit, Hacker News, the open web, and optionally X/Twitter within a configurable recent window (default 30 days). Forcing intake clarifies topic specificity, angle (trend/sentiment/problems/opportunities/comparison), time window, and platform scope before searching. Returns a synthesized briefing with citations, engagement metrics, and cross-platform pattern analysis. Use when the user requests multi-source recency intelligence on a topic (e.g., 'pulse on [topic]', 'what's happening with [topic]', 'what are people saying about [topic]', 'current conversation about [topic]', 'take the pulse of [topic]', 'trending: [topic]', 'find me info on [topic]'), and for competitor research, trend discovery, tool comparisons, and audience sentiment analysis.
patent
Patent prior-art and landscape intelligence skill — not generic patent help. Commits to one of five sub-use-cases via forcing intake (novelty search / freedom-to-operate / competitive landscape / acquisition diligence / litigation prior-art) before any search runs. Searches Google Patents, Espacenet, USPTO, and optionally Lens.org for citation-graph signals. Output is an editable Word document (.docx) with verdict, ranked closest art (claim-text extracted), CPC-class-aware landscape, family-resolved hits, geographic coverage, FTO flags where applicable, strategy recommendations, and full audit log. Use when the user asks for patent searching or analysis (e.g., 'prior art search for [invention]', 'freedom to operate analysis for [product]'). Produces search signal, not legal advice — always recommends consulting a patent attorney before filing or licensing decisions. Trademark, copyright, and trade-secret questions are out of scope.