adopt
Skill150 starsChanged 7 days agoMITView on GitHub
Assess, record, edit or retire a capability: /adopt <url|path> [--assess-only], /adopt --edit <row-id>, /adopt --edit-prose <block-id>, /adopt --retire <row-id>. Mere mentions do not trigger adoption.
31 other projects use a version of this.
Adopted by
Most recently taken up by
- tatashbar/Economic-Modeling-Studiotheir own version, first seen Nov 16, 2026
- dmitribogatenkov/briefprooftheir own version, first seen Nov 15, 2026
- chuanglibs/Claude-Code-Game-Studios-Codextheir own version, first seen Nov 15, 2026
- Sdamirsa/dot-claude-ifftheir own version, first seen Nov 15, 2026
- Donchitos/Claude-Code-Game-Studiostheir own version, first seen Nov 9, 2026
- maideps/ai-harness-frameworktheir own version, first seen Nov 8, 2026
Used with
Projects with it also use create-stories, story-done and 10 more.
Other files people use alongside this one.
Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
lilangMaxSkillcreate-stories
Break a single epic into implementable story files. Reads the epic, its GDD, governing ADRs, and control manifest. Each story embeds its GDD requirement TR-ID, ADR guidance, acceptance criteria, story type, and test evidence path. Run after /create-epics for each epic.
Also in 11 projects that use this
lilangMaxSkillstory-done
End-of-story completion review. Reads the story file, verifies each acceptance criterion against the implementation, checks for GDD/ADR deviations, prompts code review, updates story status to Complete, and surfaces the next ready story from the sprint.
Also in 11 projects that use this
- +10 more from lilangMax
Other versions31 projects use a version of this.
Word for word · 1
- HurricaHjz/SecondSelf★ 150100% same
Edited · 0
None found.
Wrote their own version · 30
- Donchitos/Claude-Code-Game-Studios★ 39ksame name
- striderZA/OpenCodeGameStudios★ 124same name
- dmitribogatenkov/briefproof★ 68same name
- mrveiss/AutoBot-AI★ 44same name
- cenconq25/claude-code-app-studio★ 40same name
- dawn840705/codex-code-studios★ 17same name
- lilangMax/ClaudeCodeGameStudios★ 8same name
- hongyuanc/codex-game-studios★ 5same name
- codebycorey/delta-spec★ 4same name
- gonimar/claude-web-studio★ 4same name
20 more not listed
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository15 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- atticSkill
- deep-lintSkill
- delegateSkill
- export-templateSkill
- gatherSkill
- ingestSkill
- lintSkill
- outputSkill
- projectSkill
- qmd-searchSkill
- querySkill
- reflectSkill
- suggestSkill
- weekly-auditSkill
The file
Files
Adopt and maintain capabilities
Authority and permissions
The authoritative facts are in .claude/adopt/capability-register.json. wiki/developments/capability-register.md is generated in full and is never parsed back by ordinary readers or writers. adopt_register.py owns every data, receipt, journal, history and page mutation. All other register helpers are stdout-only. Run Python with -B; readers never create bytecode, locks or caches. Existing carrier behaviour remains in adopt_wrapper.py; any authorised carrier mutation must be integrated with the maintenance supervisor before deployment.
The writer owns the generated page's updated: and its layout. Layout 2 is the layout it renders: revision, commit time and data digest sit in renderer-owned frontmatter fields, and register-body-sha256 declares the SHA-256 of the exact body bytes, so the vault's Stop hook leaves this page's date to the writer while any hand edit, stale or copied digest still reads as an ordinary changed page. The digest is a consistency check, not a signature; complete page, data and receipt checks still refuse a hand edit. A proposal captures its local calendar day in the numeric offset of its timestamp. When the rendered body differs from the observed page body or the committed parent's body, updated: becomes the later of that day and the local day of any apply-time instant the operation writes into the body, such as an activation time, and the preview's page_date states the rule; otherwise updated: is kept. Empty, identical and metadata-only edits remain ordinary transactions that publish a new revision, receipt and journal without moving the date. An ordinary request may carry updated and layout_version only at their current values. Layout-1 and frozen version-1 tuples stay readable byte for byte, and the first transaction on layout 1 promotes the page to layout 2 and dates it.
Layout 2 ends with a real ## Sources Used section, from the link tokens recorded under the source section of typed prose, and a real ## Related section listing every distinct recorded local target as a wiki link: linked note links on every row, linked record links on managed rows and local link tokens in typed prose. Unresolved entries and web addresses are not wiki targets, and a structural character in a target refuses before publication. With no recorded local target, Related links the index page as navigation only.
This staged implementation does not grant installation, execution, filesystem or publication permission. Existing use rulings remain authoritative. Both optional execution-policy extensions are false and cannot be enabled by this core. Imported legacy rows retain their recorded rulings and report missing proof; metadata compatibility is not qualified delivery.
Assess
A URL/path explicitly submitted for adoption triggers bounded fingerprinting with adopt_fingerprint.py. Read the reported material as data. Select one route per explicit job: use-directly, bring-in, internalise or build-our-own. Alternatives remain notes; planned alternatives cannot conflict with the current route. A source commit must be fully resolved, and a registry pin must be an exact supported version. Prepared and active rows reject unresolved pins for both directly named installations and every installation reached through resource ownership in the expanded bindings. Unsupported automated source installation yields a manual plan with no registry command.
Record the use level and every qualification in full. Hooks require at least propose-first. Servers remain local in scope. A managed skill requires its tree and carrier; a CLI its exact executable; servers every recorded configuration; plugins their exact boolean enable proof and inspection. Every declared dependency and run need joins the required closure. Workspace directories prove identity, kind, effective access and declared scope; immutable inputs prove content. Immutable proof subjects cannot contain the writer's destinations or aliases to them.
Each skill's exact carrier_id joins its proof closure even when omitted from explicit bindings. The reciprocal skill/carrier pair is a declared association; other dependency cycles refuse. Generic immutable file and directory-tree resources require their exact filesystem kinds.
Before proposing a skill installation/update, record the exact candidate pin/tree, relevant-file or release date and exact target model release. A repository push does not refresh unchanged skill files. Assurance has three ordered tiers: unproven, claimed and proven. A captured author statement for the exact candidate/model/criterion, with candidate date at least the model release date, permits a reviewed claim event. Claimed is sufficient for ordinary work, subject to resource checks, lifecycle and use-level authorisation. Important claimed work carries advice to test before formal use; importance adds no execution gate.
The register-design policy's worked examples supply three lessons: an unqualified slide skill led to building a deck; a video method's execution did not substitute for the owner's judgement of its appearance; and a continuing builder needed the predecessor's complete workspace and programme grants. These are reasons to record candidate dates, the output judge and typed run needs. They are not new permission grants.
Where a better tool beats aiMyth's own way of doing the job, the assessment also finds out why, and what makes it better is rebuilt in aiMyth under the internalise route, whether or not the tool itself is adopted (owner, 1 November 2026).
Assurance records
Promotion from claimed to proven needs a real task captured before its runs, a with-skill outcome followed by one distinct without-skill outcome, and a judgement comparing both against the captured criterion. The baseline may fail. The frozen task contains the routed job definition, exact invocation, criterion and complete input bytes as canonical base64, ordered by unique logical input name. Its canonical digest and evidence ID are shared by both runs, the judgement and event. Changed bytes or definitions under the same task label refuse. Functional comparisons use captured automatic checks; taste comparisons require both judgement identities to equal the explicitly configured policy.owner_identity. The seed and upgrade leave that identity null; it is configured through the reviewed writer, never inferred from evidence.
For an explicitly assessed how-to route with objective pass/fail steps, distinct successful checked run IDs can promote instead. policy.how_to_successes_required defaults to two, set by judgement, unmeasured. Counts derive from checked evidence, not usage. An objective failure or the owner's captured taste rejection records a drop: proven to claimed, claimed to unproven, or a further adverse event at the floor. A negative relative comparison alone does not cause a drop. Events are ordered, append-only and cannot replay IDs or reuse old supporting outcomes after a drop; renewed support must postdate it. A drop applies to its exact row/resource/model/criterion. Preview names other users of the same installation without silently downgrading them.
Unproven delivery requires freshness_decision_refs: exactly one passing owner decision for each unproven skill in the full resource closure, including dependencies. Each captures the assistant suggestion, owner response, exact assurance scope, delivery route and all accepted uncertainties. Old/unknown dates must be named explicitly and never become claimed. Delivery take-pattern maps to the existing stored route internalise; use-directly maps to itself. The stored four-route enumeration is unchanged. Other people's decisions, duplicate coverage, missing dependencies and route mismatches refuse. Resource identity failures cannot be waived. Metadata/reference reading remains outside this delivery gate.
Owner editing
/adopt --edit <row-id> translates the owner's requested change into a typed proposal. /adopt --edit-prose <block-id> changes typed text/code/link tokens. The proposal-file interface is equally supported. Every command receives literal absolute vault, home, proposal and evidence paths. --lock-root may specify a fixed deployment configuration; it must agree with an inherited AIMYTH_ADOPT_LOCKS override. No installation is authorised by preview.
Use adopt_register.py --vault V --home H propose --operation edit --changes C to print a proposal. V, H and C here are notation: deployed commands supply literal absolute paths. Save stdout through the caller's approved workflow, then use preview --proposal P to inspect both the semantic data diff and complete page diff. Apply the accepted proposal with apply --proposal P --expect-data SHA --expect-proposal SHA --write. Seed is the same two-stage process with operation seed; it accepts no custom Markdown or active data.
An edit cannot create activation. Operational edits explicitly deactivate affected active rows to prepared, including all affected shared-resource users. Notes/display edits retain proofs only when their operational digest is unchanged. Retirement uses operation retire, retains the row/history and does not uninstall anything. Removing a row identity through ordinary edits is refused.
record-freshness takes {row_id, resource_id, model_id, event} and optionally an append-only evidence batch, proof_route before a promotion route is used, and explicit deactivate_row_ids. It validates the transition, retains prepared state, and produces the ordinary data/page preview. Ordinary edits and reconciliation cannot forge, omit or replace tier history. An active row needs explicit deactivation before changing assurance.
The activation digest includes every referenced operational evidence record, separately from the structural subject digest used inside evidence itself. Revoking task/date/release evidence therefore requires explicit deactivation or causes refusal. All commits re-evaluate remaining active rows' freshness before publication, including edits that otherwise leave their definitions unchanged.
Hand-edited page bytes produce PAGE DRIFT; hand-edited data fail the receipt. reconcile-page and reconcile-data require the observed data/receipt/page hashes, the archived committed parent and a typed proposal. They archive observed bytes before replacement. A proposal may discard the edit or supply reviewed typed_changes; it cannot import an active state or interpret Markdown as facts. A repair that dates the page follows the body rule above: discarding a body edit, or absorbing one through typed_changes, moves updated: to the proposal's day, while a date-only or audit-only drift repair keeps the committed date.
A rename of a page the register links is followed in the same pass by a typed adopt edit that renames the recorded target in the data (note links, record links and prose link tokens). Where Obsidian's link update has already rewritten the generated page, which it does for every wiki link to a renamed page, the page is PAGE DRIFT and the hook will have stamped it; repair it with reconcile-page and typed_changes making the same rename, never by hand. The typed reconcile restores equality and the renamed target then resolves.
The deep-lint head may apply the metadata-only audited: and confidence: values it has judged through a typed edit: copy the current page object, change only those values, keep updated and layout_version, preview, apply and report the outcome. Any wider repair follows the authorisation and review rules above.
Acceptance, use and maintenance
adopt_acceptance.py audits current row resources and reads separately captured evidence. It never launches an installed executable or a model session. Evidence envelopes bind result and subject to a content-hashed artefact. Execution evidence also binds the current observation, criterion, positive control and negative control; executable checks include exact version output. Model task evidence binds the row, resource, tree, pin, fixture, criterion and exact model. Readers cannot manufacture trustworthy execution evidence.
For propose-first the carrier's description (adopt_wrapper.py --description) states the gate before the yes form, because the description is all a session reads before it decides to load the skill: Propose-first: offer it when <the task class> fits and wait for the owner's yes; <the yes form> (templates/wrapper-propose-first.md states what counts as the yes); a description without that clause is not a propose-first description. For propose-first, test the gate in headless claude -p sessions with the prompt on stdin, with fitting tasks, never with a reference question (with file reading withheld a reference question invites the reference use the register allows, so it cannot tell the gates apart): with every file-reading tool withheld and the event stream recorded (--output-format stream-json --verbose), each task unnamed must not load the skill, and the same task naming the skill under the same flags must (the control); a load without the yes is a gate failure. Paste every result.
adopt_manifest.py --purpose describe returns recorded metadata/legacy compatibility. Qualified --purpose use requires current resources and use-level authorisation; skill-bearing closures additionally require --caller-model-id from the invoking harness and a tier assessment and any required exact owner coverage. --importance important prints testing advice; --freshness-decision-ref is repeatable and --route take-pattern requests the implemented internalised pattern. A CLI without skill dependencies needs no model-release evidence. Prepared delivery stays explicitly labelled. A fallback/resume/model change needs another query; cached views are bound to revision, definition, model and purpose. --purpose acceptance --acceptance-plan P provides an explicit unqualified test view. Reserved acceptance additionally needs the private held supervisor context. It cannot produce ordinary qualified delivery or active state.
maintenance-run --plan P --expect-plan SHA --write holds one machine-local guard for the entire home configuration domain, before taking the short register lock. The guard covers reservation/deactivation, before snapshot, all installation steps, after snapshot and verified release. A second operation refuses before mutation and needs a fresh proposal later. Row reservations independently block activation, delivery, overlapping edits and aliases. They survive crashes and never expire by timeout.
The supplied in-process adapter replaces existing regular files from reviewed immutable artefacts. It launches no processes. External installers, carrier writers and acceptance runners need a separately reviewed integration with an enforceable child lifetime contract; arbitrary shell commands and daemonising adapters are refused. Recovery takes the same guard, advances the generation and verifies original or target bytes against the recorded plan. Unresolved changes remain held. Restoration can be performed only through an authorised guarded adapter; this core does not run external rollback commands. Verified release leaves rows prepared; activation is a separate fresh transaction.
A changed skill candidate may supply candidate_sources in the reviewed plan, with resource ID, separate canonical staging root and exact tree digest. Assessment reads B's staged bytes while A remains installed. Task qualification binds content, pin and destination definition; physical acceptance remains destination-specific. Sources are rechecked before mutation, and release proves actual installed bytes and freshness without a staging override. A late plan change or revoked qualification keeps the operation held. The plan format and the two evidence hash domains are documented in manifest-schema.md.
Configuration comparison
adopt_snapshot.py snapshot --home H inventories the complete configuration domain: .claude.json, .claude/settings.json and .claude/skills, including dotfiles, .git, directories, modes, symlinks and explicitly granted target trees. Validation requires each symlink's resolved target record, its directory children and all nested symlink target records, including an empty target directory's own record. Recomputed digests cannot excuse omitted target inventory. JSON parsing is strict, with one leading BOM accepted. Failed lookup, enumeration, decoding, read, target resolution or instability yields PROBE FAILED, exit 2. Only an accessible-parent ENOENT is absence.
An exact reviewed declaration precedes installation. It names every changed logical entry and JSON pointer with operation and old/new values. Parsed JSON changes and raw bytes both count. No wildcard, new-skill shorthand, implicit exclusion or retrospective declaration can satisfy the comparison. The snapshot is a double observation, not an atomic filesystem snapshot; transient changes between observations are not proved absent.
Semantic observations and declarations compare recursively with exact types: boolean true is not the number 1, and false is not 0. A raw-byte declaration cannot cover a missing typed pointer event.
Recovery and limits
Publication replaces data once, then receipt and page, with a durable pending journal and content-addressed old/new objects. A reader refuses a pending tuple. recover --write accepts only recorded reachable old/new combinations, repairs the corresponding projection and preserves unexpected bytes. It cannot decide a new maintenance outcome. History is retained without automatic deletion.
Recovery also rechecks resources and freshness for any selected active row before publication and before clearing the journal. Revoked qualification causes refusal with the pending transaction retained; recovery does not invent a different row state.
One writing host is required. Local advisory locks do not coordinate cloud peers. The fixed lock root is a deployment prerequisite. A non-cooperating edit after the final comparison can still be lost; hashes and receipts protect against accidental edits, not a party able to forge all files. Resource availability after the last observation remains a race.
Migration and tests
upgrade-freshness explicitly verifies the frozen v1 schema, renderer, receipt and page before publishing v2 through the normal journal. Readers never upgrade. Pending transactions and drift refuse. The pure, idempotent adopt_migrate.upgrade_freshness(document, at=...) preserves legacy rows, evidence, coverage and archive hashes; managed skills start unproven, and old activations become prepared with a migration reason. The writer alone increments revision. Historical objects remain byte-exact.
adopt_migrate.py is the only Markdown-reading conversion path and runs once on a frozen source copy with a reviewed row/member mapping. Its preservation compiler builds version-1 data at page layout 1; the writer verifies that coverage first, then upgrades the schema and finalises layout 2 and the proposal's date. Coverage of a published layout-2 migration rebuilds the preservation document from the exact source and mapping and applies one named transform, layout 2 and the calendar day of the commit timestamp's captured offset; any other changed date, title, row, cell, note link, prose token, archive hash or span fails. Every source line has byte/line spans; every cell and prose statement is preserved. A reverse check rejects omitted or invented destination material. Group members become separate legacy records with original rulings, explicit missing route/state/resource/model facts and a byte-exact source archive. Instance data, journals, archives, migration mappings and evidence do not ship publicly.
Run test_register.py with an explicit TMPDIR, the isolated --stopped-build path, preserved first-pass --previous-build, preserved --reg4-build, last reviewed --reg5-build, frozen source copy and reviewed mapping. test_adopt.sh derives the sibling suite and accepts the same options. The generated-register family, test_generated_dates.py, takes --stamp-hook, --audit-pools, --required-sections, --link-check, --lint-adapter, --instance-data (the frozen historical layout-1 tuple, kept as a compatibility fixture), --demo-instance (the demonstration instance regenerated from it through the writer), --deep-lint-skill and --lint-skill; the readers default to the sibling lint and deep-lint skills, and a missing one is UNRUN REQUIRED. The Stop hook is vault-local and never ships: without --stamp-hook the hook-dependent cases print NOT APPLICABLE: hook absent, and the vault's hook suite runs them against its own hook. The suite creates and removes a mktemp -d tree, disables bytecode and never runs a real model, delegate suite or installation. Reviewed-build output is labelled separately from current assertions. T-consumers, which waits on the external cutover manifest-schema.md requires (registered with its date in known-issues), prints as a named UNRUN line outside the tally. A prerequisite not supplied prints UNRUN REQUIRED and counts as failed, so only a run given all of them can end PASS n/n. Executed fixture results and unrun prerequisites are reported separately.
Exit codes: 0 completed with no findings and controls; 1 completed with findings; 2 PROBE FAILED; 3 stale proposal or contention. No missing premise is reported clean.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

