Skills found in the code archaeology¶
Not every missing-looking reference needs a new skill. This register records what the September 19 audit found, what can be reused, and what still depends on another system. It accompanies Code archaeology: agents and skills and the Missing agents catalogue.
The register is documentation, not an installation plan. Public originals are linked directly. Proposed replacements from the research remain drafts; none were installed or tested here.
Berd: personal content, renames and private dependencies¶
| Trace | Finding | What it means for reuse |
|---|---|---|
write-like-me |
Copycat's prompt describes a personal style-guide skill created from samples. No universal finished guide was recovered. | The mechanism can be built; the person's voice cannot be reverse engineered from the persona alone. Keep separate named profiles and review learned habits. |
bb-tools |
Builderbot requires it; the authoritative external skill was not recovered. | A discovery-only draft is not a replacement for its service contract, CLI configuration or authentication. |
berdctl |
The help-migration commit explicitly calls the standalone skill a stale reference that never existed. | Use berd-help's CLI guidance, not an invented recovered original. |
goose-help |
Renamed and split in the same migration. | An old Berd help name, not today's separate Goose documentation-guide skill. |
goosectl-new-command |
Renamed to berdctl-new-command. |
Contributor guidance for building commands, not end-user CLI operation. |
buzz-handoff |
Complete separately distributed Berd skill. | Retrieve bounded Buzz context, work privately, return a message only with approval of exact content and destination. A compatible configured CLI is required. |
berd-orchestrator, berd-monitor |
Complete orchestration and monitoring skills. | Operational guidance, not proof of a recovered Conductor persona. Verify supported commands in the installed app. |
me-md, linting-skills |
Named in historical discussion; no standalone originals recovered in scope. | Insufficient evidence for faithful reconstruction. |
shared-voice.md |
An unresolved authoring dependency in early prompts, later made self-contained. | Not an established Agent Skill, and no complete shared-voice original was recovered. |
The deliberately removed feedback-triage skill¶
The historical goose-feedback-triage original is recoverable. It classifies and prioritises feedback, groups related reports, preserves evidence and reserves duplicate marking for genuine duplicates. It separates recommendations from an authorised execution pass and checks the results of changes.
Its public removal was deliberate. The workflow uses an internal Linear project, fixed team and label identifiers, known parent tickets and sq agent-tools linear commands.
This is a procedure tied to a live issue tracker, not a downloadable collection of all feedback. A portable version needs your tracker, project, conventions, connector and mutation authority. The audit did not access internal tickets or run those commands.
Buzz: complete examples and host-specific operations¶
| Trace | Finding | Reuse boundary |
|---|---|---|
github-research |
Complete Meadow Core skill: search code, issues and PRs; read evidence; report findings and gaps. | Useful method for Explorer or Reviewer. Check tools, access and rate-limit assumptions; research does not authorise publication. |
sprout-cli / buzz-cli |
The former discovery path links to the complete Buzz CLI skill. | A renamed/path-linked resource, not a lost skill. Some operations assume a current channel and managed identity. |
desktop-screenshot |
Complete repository-specific skill. | Depends on Buzz's mock bridge, builds and scripts. Its stale normal-build advice conflicts with the contributor guide's E2E requirement. PR publication is a separate external action. |
security-review, shared |
Example names in the persona-pack specification. | Not enough evidence to publish complete originals. A valid pack is not proof of deployed runtime behaviour. |
The Meadow README explicitly says direct persona-pack runtime integration is not implemented. Do not turn example subscriptions, hooks or triggers into Berd capability claims.
Goose: removed workflows, current skills and recipes¶
| Resource | Status | Reuse boundary |
|---|---|---|
edge-case-finder |
Complete historical skill with a supporting checklist. | Useful UI/flow review method. Finishing a checklist does not prove something defect-free; fixes need scope and approval. |
create-app-e2e-test |
Complete historical contributor skill. | Requires the old Tauri test driver, live dev app and matching helpers. Not generic desktop automation. |
ui-refactor-review |
Recovered standalone original; its focused pass also survives in later historical code-review guidance. | File removal does not establish that the method disappeared. Adapt to the target repository. |
goose-doc-guide |
Current embedded skill at the inspected Goose revision. | Goose documentation lookup, not Berd's former goose-help. |
web-search |
Current embedded skill at the inspected Goose revision. | Instructions need executable dependencies and, for some providers, configuration or credentials. |
compare-tasks |
Complete evaluation-specific skill. | Compare real instructions, traces and verifier results. Not a missing general reviewer. |
| RPI research recipe and Github Manager recipe | Reusable task definitions, a different resource type. | A recipe can include operational setup. It is not automatically a persona, skill or authorisation to run. |
Current skill-loading source and Summon documentation separate skills from delegated agents and recipes. Older migration wording should not be read as “skills no longer exist”.
Knowledge is not the same resource as a skill¶
For the proposed local brain, the recoveries suggest four different things to keep distinct:
| Resource | Example from the audit | What to preserve |
|---|---|---|
| Authoritative knowledge | The operating-model document named by block.md | Source, ownership, date, accepted version and audience |
| Raw evidence and working state | Research notes, conversations, Ralph's progress log | Provenance and context; not automatically accepted truth |
| Reusable procedure | Feedback triage or GitHub research | Trigger, steps, dependencies and approval boundaries |
| Live external state | Linear issues, Buzz channel membership | Retrieve through authorised tools; do not pretend a copied document is current |
Goose discovers shared agent files and skill directories. That is an instruction-sharing seam, not automatic knowledge synchronisation.
Its Memory extension stores category text files and preloads global memory at initialisation. The prompt asks for confirmation before saving, but the storage operation is not a separate pending-approval system. Chat Recall filters session types; it is not a search of every application's conversations.
Berd's inspected memory branch contains review and recall machinery but was unmerged into the inspected public main. The Berd backend lock pins an older Goose revision than this upstream audit. Buzz's Goose ACP connection likewise does not establish shared memory, permissions or session history.
Keep one deliberately maintained knowledge authority if that is the design you choose. Do not begin with uncontrolled two-way copying between app memory stores.
Audit scope¶
Source checked on 2026-09-19 against the pinned checkpoints and search limits. No private repositories or live trackers were accessed, no skills were installed, and no cross-app memory integration was tested.