harness-garden
Skill49 starsChanged 14 days agoApache-2.0View on GitHub
Use when recent trajectories, notes, or refusals should become one evidence-backed harness improvement ticket and at most one guarded worker.
In this repository151 other files it gives its agents.
- AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- .github/copilot-instructions.mdCopilot instructions
- llms-full.txtllms.txt
- llms.txtllms.txt
- agent-readinessSkill
- appeal-scoreSkill
- bottleneck-mapSkill
- ci-repairSkill
- claim-repro-scoreSkill
- clear-out-wipSkill
- commit-cleanSkill
- conflation-scoreSkill
- cross-validateSkill
- curate-clusterSkill
- debt-cleanSkill
- debt-orchestratorSkill
- disambiguation-scoreSkill
- dojo-rsi-scoreSkill
- dos-dispatch-loopSkill
- dos-dispatchSkill
- dos-next-upSkill
- dos-replanSkill
- dos-witness-claimSkill
- field-borrowSkill
- fleet-waveSkill
- git-subagent-syncSkill
- goalSkill
- guard-rsi-scoreSkill
- harness-creatorSkill
- industry-scoreSkill
- issue-orchestratorSkill
- issue-queueSkill
- issue-triageSkill
- lightgap-scoreSkill
- memory-compactSkill
- milestone-scoreSkill
- modularizeSkill
- negframe-scoreSkill
- opencode-waveSkill
- operator-heaviness-scoreSkill
- persona-scoreSkill
- phased-planSkill
- plan-auditSkill
- quality-scoreSkill
- question-loopSkill
- refresh-cachedoc-numbersSkill
- refresh-readmeSkill
- releaseSkill
- resume-watchdog-auditSkill
- run-it-all-nightSkill
- score-2xSkill
- scorecardSkill
- scout-loopSkill
- signal-firstSkill
- skill-lifecycleSkill
- skill-overlapSkill
- skill-scoreSkill
- slop-scoreSkill
- sota-checkSkill
- The first 60 shown.
The file
We split this skill into the rules it gives an agent
5 rules. About 0 are common, 4 are distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈4
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 19
Run the supported cross-harness audit into allocated scratch:
- Very specific: line 26
Read the report, recent operator notes, relevant guard/refusal evidence, and open/closed issues.
- Distinctive: line 41
Prefer, in order: no change, deletion, configuration, an existing project primitive, standard library, then new machinery.
- Distinctive: line 45
Search open and closed GitHub issues with the observed error class, affected primitive, and desired outcome.
- Distinctive: line 62
Only after the durable issue exists, define a distinct file tree and ask DOS before launch:
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 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
Harness garden
Close one inward-looking improvement loop: observe -> classify -> subtract or ticket -> admit one worker -> witness. Compose existing project primitives; do not create another analytics stack, scheduler, or dispatch mechanism.
Fixed bounds
One pass may update or file at most one issue and launch at most one worker. It may conclude drop when fak already supports the capability or evidence is insufficient. Keep private trajectory content out of public artifacts.
1. Observe recent work
Run the supported cross-harness audit into allocated scratch:
fak tree-doctor --scratch-dir harness-garden --json
fak trajectory audit --since 7d --jsonl _scratch/harness-garden/recent.jsonl --md _scratch/harness-garden/recent.mdRead the report, recent operator notes, relevant guard/refusal evidence, and open/closed issues. Follow ../trajectory-audit/SKILL.md for sampling, privacy, baseline, and behavior classification.
A nonzero audit is evidence, not an empty report. Preserve its typed refusal and partial counts; never infer that missing output means no bottleneck. If refusal prevents honest diagnosis, make that refusal the candidate defect or stop as insufficient-evidence.
2. Classify before proposing
Classify the highest-leverage repeated friction as exactly one:
subtract: stale, overlapping, or unnecessary instruction/machinery;config: an existing primitive is correct but poorly defaulted or connected;skill: behavior can improve through concise, reusable guidance;tool: deterministic code or schema support is missing;inherent: no safe harness change can remove the constraint;insufficient-evidence: current evidence cannot support a change.
Prefer, in order: no change, deletion, configuration, an existing project primitive, standard library, then new machinery. Name the product defect separately from any guidance-side mitigation. Rank by repeated operator cost and breadth, not novelty.
3. Dedupe and choose one durable action
Search open and closed GitHub issues with the observed error class, affected primitive, and desired outcome. Choose exactly one:
drop- capability is present, evidence is weak, or cost exceeds value;update- add the new witness or narrowed done condition to the matching issue;file- create one issue only when no duplicate exists.
For update or file, include:
- For / Problem / Today / Better because / Witness against the real next-best alternative;
- centrality and P1-P4 from
docs/problems-we-solve.md; - smallest working spine, gold-plating boundary, done condition, operating envelope, privacy boundary, and follow-ons;
- audit command, bounded counts, typed refusals, and artifact paths rather than private transcript content.
Apply ../ticket-scope/SKILL.md. Validate through the supported issue-contract surface when available. If unavailable, preserve the full contract instead of inventing a substitute tool.
4. Admit one bounded worker
Only after the durable issue exists, define a distinct file tree and ask DOS before launch:
dos arbitrate --lane <leaf> --kind keyword --mode exclusive --tree <path>... --output json --explain
pwsh tools/launch_goal_detached.ps1 -PointerFile <allocated-prompt> -Workspace C:\work\fak -WorkKind gardening -PlanOnlyThe packet names the issue, owned paths, non-goals, required test/witness, and stop condition. Use fak worktree worker prepare|land|reap for implementation isolation as required by AGENTS.md.
Launch at most one worker after admission and detached preview are green. Keep the spawn preflight enabled. Never use a bypass, raw background process, alternate account, or broader tree to route around an account refusal, schema refusal, lease collision, or another typed gate. On refusal, record the ready-to-run command and typed outcome; that is a prepared packet, not a launch or shipped work.
5. Witness and close
A worker self-report is not evidence. Follow ../dos-witness-claim/SKILL.md, independently read the diff/artifact, run the scope-correct gate, and land only witnessed green work. Reconcile the issue with discoveries and follow-ons.
Report one claim per line:
- Decision:
drop | update #N | file #Nwith the evidence artifact. - Bottleneck: classification plus bounded count/refusal.
- Worker:
not planned | prepared | refused <TOKEN> | launched | witnessed <module@rev>. - Next check: one command or issue URL.
Proof scenarios
A pass is not dogfooded until evidence covers:
- existing capability ->
dropand no issue; - duplicate -> update one issue, no second ticket;
- absent capability -> one contract-complete issue;
- trajectory schema refusal -> partial evidence preserved, no false clean claim;
- account refusal -> worker remains prepared, not launched;
- lease refusal -> colliding worker does not launch;
- green worker -> independent witness before landing or done claim.
When not to use
- External repository or paper scouting ->
../scout-loop/SKILL.md. - Broad issue execution ->
../super-loop/SKILL.md. - Questions without implementation ->
../question-loop/SKILL.md. - One known ticket -> ticket-specific dispatch, not another garden pass.
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.

