delegated-build-stint
Skill2 starsChanged 4 months agoMITView on GitHub
Contract for delegating a build task to a subagent (or running one yourself as the delegate) against a frozen plan or spec, with machine-checkable acceptance and honest handoff reporting. Invoke when spawning an agent to build a harness, kit, runner, or feature from a plan it must not reinterpret.
In this repository6 other files it gives its agents.
- deneuralyzerSkill
- experiment-freeze-gateSkill
- groundSkill
- heartbeatSkill
- literature-scope-disciplineSkill
- seam-proof-buildSkill
The file
We split this skill into the rules it gives an agent
3 rules. About 0 are common, 2 are distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈2
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 13
Every rule below was compiled from a real incident.
- Very specific: line 22
S2 - Import-check EVERY CLI path: Every flag combination the deliverable exposes (--all, --check-only, --smoke, ...) gets at least an …
- Distinctive: line 30
S3 - Config-integrity test: Every frozen constant in the plan (grids, thresholds, anchors, magic numbers, seeds) is duplicated verbatim in …
≈ “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
/delegated-build-stint - build against a frozen plan, hand back evidence
A build stint executes a frozen plan with zero design discretion. It never changes the plan's constants, interfaces, thresholds, or scope; a plan problem is an escalation, not a judgment call. Its output is the built artifact plus a report the delegator can audit in minutes.
Every rule below was compiled from a real incident.
Build rules
S1 - Machine-checkable acceptance ONLY. Acceptance = tests green on constructed inputs with stated tolerances, integrity checks green, every entry point exercised. "Looks right" and "should work" are not acceptance states.
S2 - Import-check EVERY CLI path. Every flag combination the deliverable
exposes (--all, --check-only, --smoke, ...) gets at least an
import-and-parse exercise before the stint reports done. Modules imported by
only one path are exactly where this bites.
Incident: a runner's --all mode crashed in production on a wrong import in
a module that only --all loaded - the stint had exercised only the
--check-only and --smoke paths, both green.
S3 - Config-integrity test. Every frozen constant in the plan (grids, thresholds, anchors, magic numbers, seeds) is duplicated verbatim in a test that fails if the built config drifts from the plan.
S4 - Declared-deviations report. Any place the implementation differs in FORM from the plan's stated formula or structure - while being algebraically or behaviorally equivalent - is listed explicitly, with the equivalence argument and the test that certifies it. Silent equivalences are audit debt. Worked example: a stint implemented a mathematical quantity in a rearranged but equivalent form, declared it, and attached a 1e-12 numerical-equality test. Audit took one read instead of a derivation.
S5 - NEVER take custody of a long run. The stint's acceptance ends at "machine certified": self-tests plus a fast, declared smoke. The long or unattended run is launched by the delegator (or the human), or - only if short enough to block on - run as a blocking call whose completion is inside acceptance. A stint never spawns a detached run and reports complete. The mechanic, three real incidents: a subagent that spawns a detached child process has no live children of its own, so the harness sees it idle and reports the stint finished - while the real run is un-run or half-run. Each instance cost the delegator a disk-verification and relaunch pass. The build was fine every time; the custody handoff was the failure.
S6 - Prior deliverables are READ-ONLY. Frozen artifacts from earlier
increments are never edited; reuse is by import. git status at stint end must
show changes confined to the new deliverable's own directory - and the report
quotes it.
S7 - House conventions still bind. File-size caps, typing discipline, dependency policy, formatting. Delegation is not an exemption.
Reporting (the handoff)
The stint ends by emitting three artifacts:
- Compact report - claims WITH evidence pointers (
file:line, command + captured output). Never a narrative summary. Why: the report is written by the agent that did the work, and the residual failure mode of self-reporting is mis-attribution rather than invention - a claim can be verbatim-true and still credited to the wrong source or scope. Evidence pointers make spot-audits nearly free; narrative makes them archaeology. - Trace index - one line per step: action, files touched, outcome. The delegator greps it on any discrepancy.
- Full transcript / logs - untouched on disk for deep dips.
The report states plainly what was NOT done. Incomplete work is reported as "incomplete, X remains" - never dressed up as a scope decision.
Escalation (the ONLY pings allowed)
- Plan ambiguity that blocks execution (not "could be nicer").
- A sanity check named in the plan fails (abort and say which).
- Infrastructure failure after 2 retries.
- Runtime exceeds the plan's budget by more than 3×.
Nothing else. No mid-run status pings; design suggestions go in the report's observations section, not into the code.
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.

