rsp-implement
Skill1 starsChanged 8 days agoMITView on GitHub
Diagnose unexplained symptoms read-only or implement authorized bounded tasks or ready RSP Changes with ordinary own checks, risk-selected test-first work and fixed-finding correction. Not fixed-scope review or whole-document writing.
In this repository10 other files it gives its agents.
- AGENTS.mdAGENTS.md
- author-rsp-skillsSkill
- distill-upstreamSkill
- release-acceptanceSkill
- rsp-commitSkill
- rsp-docSkill
- rsp-reviewSkill
- rsp-shapeSkill
- rspSkill
- rsp-verifySkill
The file
We split this skill into the rules it gives an agent
8 rules. About 0 are common, 7 are distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈7
Distinctive · estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 35
When work affects modeled structure, responsibilities, dependencies or bindings, use architecture knowledge for reading, synchronization …
- Distinctive: line 41
Test-first: after the cause and authorized behavior are clear, load test-first only when explicitly required or a concrete changed risk …
- Distinctive: line 60
Test stable public behavior with real internal collaborators by default.
- Very specific: line 76
Use rsp-doc for document/Skill writing and whole-Change convergence within existing authority.
- Distinctive: line 78
After final mutation, establish minimum sufficient evidence for tracked Required checks or direct-task acceptance:
- Distinctive: line 81
Run missing, failed or invalidated checks and risk-justified additions, including checks affected by integration.
- Distinctive: line 82
If no verification method is prescribed, choose the lowest-cost existing test, static check, build or other evidence that decisively …
- Distinctive: line 85
Before completion, apply the same admission criteria to retained tests.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
RSP Implement
For tracked work, read the shared work model when its identity and ownership definitions are missing from the current context. Use logical WorkRefs in commands and reports; cite file locations separately. Direct invocation needs no prior Core execution.
Implement one bounded authorized task, run minimum sufficient own checks and retain necessary tracked Change evidence. Select the mode before loading its method. Changing methods within the same authority needs no handoff. No mode grants Git, review, lifecycle, external-action or approval authority.
Responses follow explicit instruction, personal instructions, then conversation language. Preserve existing artifact language and canonical identifiers, commands and values. New artifacts follow explicit instruction, effective configuration, scoped project instructions, then conversation language.
Query RSP settings only for a decision depending on an unresolved value. Use the selected CLI's successful rsp config --json summary, including defaults and inheritance, and refresh on relevant drift or recovery. Raw YAML is diagnostic only. Projection failure blocks dependent choices, not independent authorized work; resolve material conflicts with the owner.
Select and inspect
Mutation requires one of:
- A bounded direct task with explicit scope and acceptance; no WorkRef or RSP writeback.
- A ready Change selected by explicit WorkRef or resolved from the FocusSet, with settled acceptance and exact product authority. Multiple markers are candidates; a Group Brief is context, not executable work.
Diagnosis-only investigates a bounded symptom read-only without a ready mutation owner or WorkRef. A suspected cause grants no mutation. For an authorized fix, diagnose an unexplained cause first; after confirmation continue the same request only with settled owner, scope and mutation authority.
Inspect applicable instructions, the selected Change/Brief when tracked, relevant Specs and decisions, and the owning code and tests. Use supplied paths and discover missing context as needed; a dispatch summary does not replace source inspection. Operate only in the execution location supplied by the host.
For optional context discovery, distinguish no matches from execution errors: no matches should not block independent subsequent reads, while missing required inputs and actual execution errors still need handling.
Within the confirmed checkout/worktree, prefer paths relative to the command or tool's explicit working directory for repository reads, searches and edits. Reuse the known execution root; do not switch worktrees to make a path resolve. Use absolute paths when the tool requires them, accessing files outside that root, or explicitly locating another authorized root; a path alone does not establish checkout identity or revision.
When work affects modeled structure, responsibilities, dependencies or bindings, use architecture knowledge for reading, synchronization and own checks.
Select the method from the request and evidence:
- Diagnosis: for an unexplained symptom or diagnosis-only request, load diagnosis before any fix. Diagnosis-only remains read-only and ends with cause evidence or an unresolved discriminating check.
- Ordinary implementation: for a clear authorized correction, implement directly; no branch reference is needed.
- Test-first: after the cause and authorized behavior are clear, load test-first only when explicitly required or a concrete changed risk makes pre-mutation RED safer. Testability or being a fix is insufficient.
- Fixed findings: for a known review report and bounded investigation authority, load finding-resolution. Supply dispositions and one authorized correction with checks; Core owns any required cross-capability re-review loop. A report or accepted finding alone grants no mutation.
These are conditional methods, not separate controllers.
Preserve authority
Identify outcome, owner, verification, and pre-existing work. For tracked work, update only authorized Change sections; a direct task creates no Change, WorkRef, or RSP lifecycle state. Preserve unrelated modified, staged, and untracked work. Stop when overlap would discard, guess, or rewrite pre-existing intent.
Git delivery, publication, deployment, approval, and out-of-scope deletion require separate explicit authority. For conflicts, inspect base/ours/theirs, preserve unrelated work, resolve only evidenced scope, rerun affected checks, and stop before staging, continuing, aborting, or committing. Verify named findings first.
Admit boundaries and permanent tests
Inspect existing owners and callers. Prefer, in order, no addition when the outcome is already met, repository reuse, platform or standard-library facilities, existing dependencies, then the smallest complete implementation. Keep rules with their invariants; do not force unrelated responsibilities through flags or speculative interfaces.
Hide internal representation and coordination behind the necessary caller contract, including ordering and failures. Judge decomposition by what callers and maintainers must understand, not file size or layer count.
Add defensive logic or a new boundary only for an explicit requirement or an evidenced producer, production consumer, trust/lifecycle transition and material consequence. Typed same-process values and imagined or test-only consumers do not establish that need. Backups, snapshots and digests likewise need a real recovery or independent identity/integrity consumer; do not create them as routine preflight. Preserve required recovery safeguards and independently trusted integrity checks.
Test stable public behavior with real internal collaborators by default. Substitute external or nondeterministic boundaries as needed. Focused unit tests remain useful for meaningful algorithms, parsers and critical boundaries; choose evidence by risk, not a per-function/file/branch allocation or unit/integration ratio.
Reuse sufficient existing evidence before adding a permanent test. A new test needs a distinct plausible regression, observable consequence and evidence gap worth its maintenance. Exercise the consumer outcome through the relevant production path; an already-covered forwarding hop adds no confidence.
Derive expectations from the contract or independent examples. Assert outcomes and effects, including rejection; pin details, counts or order only when contract-relevant. Private structure, source wording and constants compared with themselves are not contracts; public defaults can be. When setup leaves the target condition uncertain, inspect it or use a targeted probe rather than adding ritual preconditions. Keep scenario-defining inputs and assertions visible and share incidental setup when clearer.
Continue within the evidence boundary
Investigate new failures and correct in scope. For a tool-only obstacle, inspect cause/effects before a permitted repair or replacement. Continue only with safe replay and unchanged goal, owner, scope, baseline, authority and required evidence. Named checks and command-owned RSP operations retain their methods.
Stop for a repeated failure without new discriminating evidence, non-convergence, changed owner/scope/acceptance/authority, unknown mutation effects, unsafe or one-shot replay, or unavailable mandatory evidence or environment. An independent reviewer or verifier required by acceptance remains a separate read-only owner and cannot be replaced by the implementer's checks.
Implement and verify
Implement the smallest complete slice. For authorized Change updates, load the current work record rules when absent from context. Update affected work and evidence as facts change; reopen invalidated results rather than append fix/retry tasks or check history. Plan or contract changes retain their decision authority. Direct work reports outcomes without RSP artifact mutation.
Use rsp-doc for document/Skill writing and whole-Change convergence within existing authority. Local wording fixes and mechanical status/evidence updates needing no whole-record reconciliation remain direct. Without Doc, write grounded guidance and check affected examples/links locally. Author self-checks never certify independent Review.
After final mutation, establish minimum sufficient evidence for tracked Required checks or direct-task acceptance:
- Reuse traceable passing results when the tested boundary and relevant code, dependencies, configuration, tools, inputs and environment still apply. Phase changes and unrelated edits do not invalidate evidence.
- Run missing, failed or invalidated checks and risk-justified additions, including checks affected by integration. Named checks retain their method and coverage; explicit reruns and independent execution remain binding.
- If no verification method is prescribed, choose the lowest-cost existing test, static check, build or other evidence that decisively establishes acceptance. A new test is not required.
- Match the actual target, inputs, coverage and completion to the declared check. Invocation errors or incomplete observations do not by themselves prove a product defect. Record command, tested scope, result, source and omissions, distinguishing reused evidence from new execution. Failed, unavailable or untraceable evidence cannot establish completion.
Before completion, apply the same admission criteria to retained tests. Remove disposable probes, fixtures and helpers that add no distinct future confidence unless retention is required by user, Change or project authority; establish sufficient evidence for the final retained state.
Check changed comments, test names, documents and handoff prose against the accepted result and authoritative baseline. For comments, retain accurate information that nearby code leaves unclear, including constraints, rationale or usage; remove restatement. Omit repeated Tasks/logs and discarded session attempts, while keeping negative facts needed for removals, safety, compatibility, migration, audit, failures, unresolved risks or requested comparisons. This prose check changes no behavior, public contract, test, snapshot or diagnostic.
Return ownership
Report the outcome, actual paths, verification, material limits and any external effects or pre-existing user work. State whether the selected Change or direct task is completed, partial, blocked before implementation, verification-failed, verification-unavailable, or verification-blocked:
- Failed means an exercised defect.
- Unavailable means a missing tool, dependency, service, credential or environment prevented execution.
- Blocked applies only when scoped checks pass but a required gate fails solely from a confirmed pre-existing or out-of-scope baseline defect. Affected gates are never waived.
When work remains, report owner (WorkRef for tracked work, direct request otherwise), authority, changed artifacts, reused and new checks, blockers and next action as needed for safe continuation. A non-trivial handoff may use Core's fuller continuation; ordinary same-owner corrections need no template or new request. The response is not durable truth.
Claim completion only when tracked required Tasks (if any) and applicable checks pass and no blocker remains. Do not claim review, archive, Git delivery, or release unless separately performed with explicit authority.
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.

