Skill library
Skills describe repeatable work: review a design, check a release or hand work to another agent. Keep your own procedures in the account library and assign them to the projects that need them. The free starter templates provide editable starting points; they do not run automatically.
Create and assign
Section titled “Create and assign”Open Skills under Shared libraries in the signed-in workspace. Create a skill with a title, short summary and complete procedure, or expand the starter templates and copy one into your library. Saving a skill does not assign it automatically.
Choose a project in the assignment panel, select the skills it should receive and save the assignments. Existing selections stay visible, including archived skills. The same skill can serve several projects; another account cannot access your library.
Each skill shows its assigned projects. The project’s own page also lists its active skills alongside standards. Choose View & use to inspect a skill and get the prompt for that project.
Edits create preserved revisions. Assigned projects see the current active version on their next read. Archiving removes the skill from agent discovery; restoring makes it available again where still assigned. Review changes before saving them: this release does not provide a separate approval stage or pinned per-project versions for skills.
Limits: 50 active and 200 retained skills per account, up to 10 assignments per project. Titles allow 120 characters, summaries 400 and procedures 16,000.
Discover, then load
Section titled “Discover, then load”The agent’s normal memory_read response contains skill summaries and exact revisions, without the full procedures. memory_list_skills refreshes that catalog for the verified project. Immediately before using a selected skill, memory_get_skill retrieves its current active, assigned procedure using project_id and skill_id. Use the revision returned by that call, even if the earlier catalog showed an older one.
Standards remain the project rulebook and are included in full. Skills are selected procedures for a relevant task. Neither overrides the user’s instructions, grants additional access or authorizes an action. Clients still need the tools required to carry out a procedure.
An agent can save an authorized execution log as a separate project note, including skill ID and revision, checks actually performed, findings, changes and the remaining next step. The log is reported evidence, not independent proof that a check ran. Memford does not execute workflows or verify an agent’s work itself.
Use from your connected agent
Section titled “Use from your connected agent”Connect the agent to the intended Memford project through MCP. No local SKILL.md, skill installer or additional project instruction is required for this workflow. Your procedures and revisions remain in Memford.
In the skill detail, choose an assigned project and copy the usage prompt into its connected agent. This asks the agent to fetch the current procedure; copying alone does not execute it. Archived or unassigned skills cannot provide an active-project usage prompt.
You can also ask in ordinary language, for example:
- “Which Memford skills are assigned to this project?”
- “Use the Memford skill Release readiness to review this change.”
- “Apply that Memford skill again using its current version.”
Use the title of a skill actually assigned to the project. The MCP connection supplies the authorized project and a metadata-only catalog during initialization. The agent can refresh it with memory_list_skills; both skill tools use the connection’s project when project_id is omitted. Explicit project IDs remain supported and cannot expand the connection’s access.
Immediately before applying the procedure, the agent retrieves it with memory_get_skill, uses the returned current revision and reports the work actually completed. A central edit is available on the next fetch, without changing local files or reconnecting first. Refresh the catalog after reconnecting, a supported change hint or an explicit request. Do not poll full procedures on every answer.
The server provides discovery guidance and the procedure; the agent performs the work with its own tools and permissions. A relevant task may prompt the agent to choose an assigned skill, but automatic selection is client/model dependent. Naming Memford and the requested skill makes the intended workflow explicit. No action is authorized merely by finding a procedure, and no execution result may be inferred from retrieval alone.
Missing MCP tools, an unavailable project or an archived/unassigned skill must be reported rather than replaced by an invented or stale procedure. An empty catalog means no active skills are assigned. Each client or execution environment still needs its own authenticated MCP connection and the tools needed for the work. Shared knowledge does not transfer local files or runtimes.
Native skill selectors and slash commands are client features, not a universal MCP interface. This workflow uses natural language and MCP tools, without requiring $memford or /memford. A client that ignores MCP instructions can still discover the catalog through the tools; unsupported or disconnected clients cannot use it. The initial catalog reaches the agent only if the client exposes the server’s initialization instructions. Verify actual client behavior instead of assuming automatic context insertion. See MCP server capabilities and Claude Code MCP.
If a Codex header helper can read the project but native MCP says “Project not found,” check the effective connection before changing the project. Stored OAuth credentials take precedence over a helper’s Authorization header. A distinct project-scoped server name can keep the helper separate from an unrelated saved OAuth grant; verify it in a fresh session and preserve other connections. See Codex MCP authentication.
Example: improve a design
Section titled “Example: improve a design”The design-review template asks for independent AI assessments of usability, visual design and accessibility. It specifies at most three improvement rounds, a target of eight out of ten in each area and no critical unresolved issue. Reviewers must support findings with concrete observations. Scores describe AI assessments, not human jury votes, formal accessibility certification or permission to deploy.
Keep connected agents informed
Section titled “Keep connected agents informed”The MCP initialization guidance and tool descriptions explain the discovery workflow. Every memory_read also carries the current capability revision and a compact refresh reminder. Compare it with the last guide revision successfully loaded in this session; read memory_features, or the public feature and working guide, when unknown or changed. Confirm the matching revision before treating the guide as current and reuse it while unchanged. The guide and MCP manifest share one source of truth, including Memford’s current working method. A failed fetch or mismatched revision must be reported, not treated as a successful update.
Supporting clients can receive existing project-resource change hints after assigned skills change, then read the catalog again. Product capability changes are discovered on the next read or reconnect, not through a guaranteed background push. Tool availability and inserting retrieved context into a running conversation depend on the client.
An existing MCP session can discover the updated working guide on its next project read; repeating setup or changing local skill files is unnecessary. Initialization guidance itself arrives at the next fresh connection. Use the client’s supported tool refresh or reconnect if a newly documented tool is missing. Optional lifecycle adapters remain separate and do not update themselves. The guide does not overwrite local instructions, activate paid services, install code or expand permissions.
GitHub import and synchronization, public community publishing and automatic workflow execution are not part of this release.