mattvideoproductions/MVP-Prompts / skills/foreman
foreman
VerifiedBe the foreman on a big job. One lead model plans the work, hands pieces to helper agents in lanes that never share files, checks everything before accepting it and reports the cost; or runs a long solo build with an owner in the loop. Covers the North Star plan of record, briefs that carry the lead's reasoning, measured gates plus the owner's judgement, strike-based promotion between model tiers, a cost ledger and a mistake register every pass is checked against. Use it when the user types /foreman, says "be the foreman", "be the lead" or "orchestrate this", asks to delegate to subagents or run lanes or passes, or mentions a North Star, three strikes, tiers or promotion between models. Also load it when a task looks too big for one context or needs several agents or days, even if not asked by name; the skill then has you offer it and wait for a yes. Also use it to resume such a project, close a pass, report a session's cost or record a mistake. Not for a change one agent can finish in one sitting.
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 32.2 KB |
| assets/NORTH_STAR.template.md | markdown | 4.3 KB |
| assets/brief.template.md | markdown | 1.8 KB |
| assets/config.default.json | config | 2.2 KB |
| assets/discovery-brief.template.md | markdown | 1.8 KB |
| assets/handoff.template.md | markdown | 611 B |
| assets/mistakes.seed.jsonl | other | 21.9 KB |
| assets/prices.json | config | 1.7 KB |
| assets/promoted-brief.addendum.md | markdown | 673 B |
| assets/report.template.md | markdown | 804 B |
| references/audit.md | markdown | 15.9 KB |
| references/briefs-and-handoffs.md | markdown | 10.9 KB |
| references/cost-and-mistakes.md | markdown | 16.5 KB |
| references/field-data.md | markdown | 9.1 KB |
| references/north-star.md | markdown | 11.0 KB |
| references/reference-build.md | markdown | 8.0 KB |
| references/skeleton.md | markdown | 4.3 KB |
| scripts/orch.py | script | 59.5 KB |
assets/mistakes.seed.jsonl
View on GitHub{"id": "M01", "title": "Brief listed deliverables and left the thinking to the worker", "what": "The first wave of briefs were lists of deliverables. First passes missed until briefs carried the chosen architecture, the parameters and the pitfalls.", "cause": "The lead delegated before solving the hard part, so a weaker model made the hard decisions.", "rule": "Solve the hard part first. A brief states the design and why, the numbers, the failures to expect and what is ruled out.", "when": ["brief"], "checks": [{"kind": "brief", "name": "brief_has_reasoning"}], "level": "scripted", "source": "reference build, observed; corrected by the owner", "field": "Projects B and D: the lanes that landed first time had the owner's exact words, where to look, and suspected causes.", "first_seen": "2026-11-12"}
{"id": "M02", "title": "Work started at the senior tier without trying a cheaper one", "what": "2,768 of 3,477 responses and 76% of the cost came from the senior tier. The judgement that a task needed it was rarely tested.", "cause": "Starting tier was a feeling, not a recorded decision, and strikes were not counted.", "rule": "Default to the standard tier. Starting higher needs one line of reason in the lane ledger.", "when": ["pass-end"], "checks": [{"kind": "northstar", "name": "tier_justified"}, {"kind": "metric", "name": "senior_share_of_worker_responses", "op": "<=", "config": "senior_share_ask_above", "on_fail": "ask"}], "question": "Most worker responses this pass came from the senior tier or above. Which of those lanes would have been tried at standard first, and why were they not?", "level": "scripted", "source": "reference build, measured", "field": "Senior-or-above share of worker responses: 98%, 89%, 78% and 61% across four projects. One lead never wrote down why a lane started at senior. Standard-tier workers finished several areas of one project and an edit in another when the task was well briefed and easy to review.", "first_seen": "2026-11-12"}
{"id": "M03", "title": "Agents outgrew their context instead of handing off", "what": "22 worker transcripts averaged about 140 turns each. Re-reading cached context was 58% of the senior tier's cost. The fresh-agent rule was written down and was the one followed least.", "cause": "No budget and no meter, so nothing marked the moment to hand off.", "rule": "Budget each worker by context size. Resume the same agent for follow-ups in the same scope while it is under budget; start a fresh one for new scope or past the budget.", "when": ["pass-end"], "checks": [{"kind": "metric", "name": "max_context_per_worker", "op": "<=", "config": "context_budget_tokens", "on_fail": "repeat"}, {"kind": "metric", "name": "cache_read_cost_share", "op": "<=", "config": "cache_read_share_ask_above", "on_fail": "ask"}], "question": "Re-reading context was a large share of this pass's cost. Which agents should have handed off earlier, and what would their handoff have needed to carry?", "level": "scripted", "source": "reference build, measured", "field": "Project A: about 384K tokens re-read per worker response. Project D: 17 of 75 worker transcripts ran past 150 responses, the longest 375; re-reading was 54-57% of worker cost. Project B: median 33 responses, 232K per response. Three reports say a turn count is the wrong trigger; one found that resuming the same agent for a same-scope follow-up was much faster.", "first_seen": "2026-11-12"}
{"id": "M04", "title": "A stale build, or a stale deployed copy, was tested as if it were the new one", "what": "A build command failed without an error and two test runs went against the old program. The results looked real.", "cause": "The gate did not check that the thing under test was newer than the last source change.", "rule": "Every gate first confirms the whole chain is fresh: source, build, and the copy actually running or deployed. It fails loudly if any link is old.", "when": ["accept"], "checks": [{"kind": "manual"}], "question": "For the lane being accepted: what shows the gate ran against a build made after the last source change, and that the copy tested is that build? Quote the freshness line, timestamps or checksum.", "level": "prose", "source": "reference build, observed twice", "field": "Project B: the step that bit was the deployed copy against the build (which file is on the device); fixed with checksums. Project D: some scripts only loaded because of a leftover cache file.", "first_seen": "2026-11-12"}
{"id": "M05", "title": "Gates ran a different configuration, input path or machine from the one that ships", "what": "Component gates ran at a finer setting than the product used, and behaviour differed between the two. A switch for the product's setting was added late.", "cause": "The gate was written for convenience of testing, not for what ships.", "rule": "Gates run what ships: the same settings, the real input path, and the real hardware where there is any. Where only the owner has the shipping configuration, the owner is a gate.", "when": ["pass-end"], "checks": [{"kind": "manual"}], "question": "Does every gate run the configuration that ships, through the input path a person uses, on the hardware it ships on? Name any that uses a bot shortcut, an emulator or a separate code path.", "level": "prose", "source": "reference build, observed", "field": "Ranked first by two reports. Project B: the emulator disagreed with the real device more than once, and an experimental mode crashed on boot on hardware. Project C: test bots automated an action a person does by hand, and passed while the owner hit three input bugs. Project D: headless tests used a different code path from the product, and every route was 'proven' by bots that skipped the travelling.", "first_seen": "2026-11-12"}
{"id": "M06", "title": "Cost was only known at the end", "what": "The build's cost was computed once, after the last pass, when nothing could be changed. The senior-tier share would have been visible early.", "cause": "No per-pass meter.", "rule": "Record cost by tier at the end of every pass, in the status log.", "when": ["pass-end"], "checks": [{"kind": "ledger", "name": "passes_contiguous"}], "level": "scripted", "source": "reference build, observed", "field": "Projects C and D: cost was computed only when the owner asked. Project C: seeing the cold-resume figure earlier would have changed how the lead worked.", "first_seen": "2026-11-12"}
{"id": "M07", "title": "Lane state lived in the lead's head", "what": "Strikes and tiers were untracked, and afterwards the worker count could not be reconciled with the transcripts.", "cause": "No lane ledger.", "rule": "One ledger row per lane: task, tier, reason, strikes, gate, status, agent.", "when": ["pass-end"], "checks": [{"kind": "northstar", "name": "lane_ledger_has_rows"}, {"kind": "northstar", "name": "lane_ledger_reconciles"}], "question": "Some worker transcripts from this pass are not named in the lane ledger. Which lane does each belong to?", "level": "scripted", "source": "reference build, observed", "field": "Project D: the North Star's lane ledger went stale days before the end.", "first_seen": "2026-11-12"}
{"id": "M08", "title": "The current state was buried at the bottom of a long log", "what": "The status log grew to about a hundred lines and the present state was the last entry.", "cause": "The plan had history but no summary of now.", "rule": "Keep a current-state block of at most fifteen lines at the top, rewritten every pass.", "when": ["pass-end"], "checks": [{"kind": "northstar", "name": "current_state_block"}], "level": "scripted", "source": "reference build, observed", "first_seen": "2026-11-12"}
{"id": "M09", "title": "One known failing gate hid any new failure", "what": "The check exited with an error whenever any gate failed, so a new failure looked the same as the known one.", "cause": "No register of known failures.", "rule": "List known failures with their bars unchanged, so a new failure stands out.", "when": ["pass-end"], "checks": [{"kind": "northstar", "name": "known_failures_listed"}], "level": "scripted", "source": "reference build, observed", "field": "Project D: the same pre-existing failures were re-derived several times.", "first_seen": "2026-11-12"}
{"id": "M10", "title": "Tuning a setting until the number passes", "what": "A feature that had worked stopped working. The cause was an unrelated object moved into its path. The right fix moved the object back and left the feature's settings alone. The tempting fix was to tune the feature.", "cause": "Pressure to make a gate pass without tracing the failure.", "rule": "Trace a failure to its cause. Changing an unrelated setting until the number passes is not a fix.", "when": ["accept"], "checks": [{"kind": "manual"}], "question": "For each failure this lane fixed: what was the cause, and is the thing that was changed the cause? Did any fix change the product or the test mainly so that a check would pass?", "level": "prose", "source": "reference build, done right; kept as a guard", "first_seen": "2026-11-12"}
{"id": "M11", "title": "The same tool trap was hit before it was written down", "what": "One shell silently resolved a command to the wrong program. Each such trap cost at least one wasted run before it was recorded.", "cause": "Tool quirks were recorded late, or only in a conversation.", "rule": "Add a quirk to the North Star the moment it is found, and quote it in briefs.", "when": ["pass-end"], "checks": [{"kind": "northstar", "name": "tool_quirks_section"}, {"kind": "manual"}], "question": "Did any command misbehave this pass (wrong program, silent failure, odd quoting)? Is each one in Tool quirks?", "level": "prose", "source": "reference build, observed", "field": "Project C: a known shell-quoting failure was hit twice more, and a launcher test hung for five minutes. Project B: every environment-contract entry came from a real incident.", "first_seen": "2026-11-12"}
{"id": "M12", "title": "A green gate table hid what no person had checked", "what": "At the close, the sound had never been heard and no one had used the product with the intended input device. The gate table alone would not have shown either.", "cause": "Gates measure correctness, speed and stability. They do not measure feel.", "rule": "Every report and the current-state block list what is unverified by a person.", "when": ["pass-end", "owner-report"], "checks": [{"kind": "northstar", "name": "unverified_line"}], "level": "scripted", "source": "reference build, done right; kept as a guard", "field": "Project C: sound, input-device behaviour and window handling are things no headless check can see.", "first_seen": "2026-11-12"}
{"id": "M13", "title": "The lead checked work it had designed", "what": "The lead verified results against its own design, so its reading was not independent.", "cause": "No second reader for the claims that mattered most, and none at all for the lead's own lane.", "rule": "Give important claims to a fresh agent with the work and the gate, without the conclusion. The lead's own lane always gets one.", "when": ["pass-end"], "checks": [{"kind": "manual"}], "question": "Which claims got a fresh checker this pass? The lead's own lane, the one thing and any go/no-go milestone need one. What did the checker disagree with?", "level": "prose", "source": "reference build, observed", "field": "Project C: the lead checked its own work throughout. Project D: an independent lane testing through real input found seven real bugs, and one lane's own test passed while its feature was broken everywhere.", "first_seen": "2026-11-12"}
{"id": "M14", "title": "A worker ground on a stuck problem", "what": "Workers kept trying on problems they could not solve, producing long transcripts and no progress.", "cause": "The brief did not say when to stop.", "rule": "Every brief has stop conditions: the same failure twice, a needed file outside the lane, or the context budget.", "when": ["brief"], "checks": [{"kind": "brief", "name": "brief_has_stop"}, {"kind": "brief", "name": "brief_has_gate"}, {"kind": "brief", "name": "brief_has_files"}], "level": "scripted", "source": "reference build, observed", "first_seen": "2026-11-12"}
{"id": "M15", "title": "A bar moved to meet a result", "what": "One lane loosened its own limit mid-work (Project D). A product change was made partly to steady a flaky test (Project C). In the first reference build misses were reported as misses: 12 against a bar of 15 stayed written as failing.", "cause": "Pressure to finish.", "rule": "Bars are set before the work. Only the owner changes one, dated, with the old bar kept. A change of design by the owner is logged as a spec change. A product change that makes a gate easier to pass is a bar change.", "when": ["pass-end"], "checks": [{"kind": "northstar", "name": "bars_unmoved"}], "level": "scripted", "source": "field, observed (Projects C and D)", "first_seen": "2026-11-12"}
{"id": "M16", "title": "History was rewritten", "what": "Not seen in the reference build. Earlier status-log entries or the owner's quoted words are edited so the record reads better.", "cause": "Tidying, or hiding a miss.", "rule": "The status log is append only. Owner statements are quoted, dated and never reworded.", "when": ["pass-end"], "checks": [{"kind": "northstar", "name": "status_log_append_only"}, {"kind": "northstar", "name": "mandate_kept"}], "question": "A quoted owner statement from the last snapshot is gone from the Mandate. Did the owner withdraw it? If so, record that as a new dated line.", "level": "scripted", "source": "guard; added in audit", "first_seen": "2026-11-12"}
{"id": "M17", "title": "A hypothesis was handed to a worker as a fact to confirm", "what": "A worker given a conclusion tends to return it confirmed.", "cause": "The brief stated what the lead believed instead of asking.", "rule": "State the thing to be tested as a question. For diagnosis and research, hand over the suspects and the test that would rule each out, not a conclusion.", "when": ["brief"], "checks": [{"kind": "manual"}], "question": "Does this brief ask the worker to confirm something the lead believes? Rewrite it as an open question.", "level": "prose", "source": "reference build, rule of practice", "field": "Project D: lanes overturned the lead's diagnoses at least four times. Handing over suspects to test, measured first, worked better than 'design decided'.", "first_seen": "2026-11-12"}
{"id": "M18", "title": "Refinement with no stopping point", "what": "Not seen in the reference build, because the owner set a cap in plain words: two or three more passes, then the best deliverable available.", "cause": "A system built on refinement never runs out of things to refine.", "rule": "The pass count is capped. If the owner has not set a cap, the lead proposes one and records it.", "when": ["pass-end"], "checks": [{"kind": "northstar", "name": "pass_cap_set"}, {"kind": "ledger", "name": "within_pass_cap"}, {"kind": "ledger", "name": "within_budget"}], "level": "scripted", "source": "guard; added in audit", "first_seen": "2026-11-12"}
{"id": "M19", "title": "A struck-out lane stayed at the same tier", "what": "Strikes were not counted in the reference build, so the three-strikes rule could not fire.", "cause": "No ledger, and no check that reads it.", "rule": "A lane at its strike allowance, or one that has plateaued, moves up one tier with its history.", "when": ["pass-end"], "checks": [{"kind": "northstar", "name": "strikes_acted"}], "level": "scripted", "source": "field, observed (Project D); guard for rule 7", "field": "Project D: a senior-tier lane went four rounds on one defect and stalled. A lead-strength agent reframed it and cleared it in one round. The owner had to ask whether a stronger model was on it.", "first_seen": "2026-11-12"}
{"id": "M20", "title": "The lead's own session grew huge and paid to rewrite itself", "what": "Lead contexts reached 966K tokens in two projects. One project spent about 16% of its cost on six resumes after idle gaps. One-hour cache writes were 23% of one lead's cost and 50% of another's. In one project a lead response cost more than twice a senior worker's.", "cause": "One lead conversation was kept alive for days. No budget applied to the lead.", "rule": "The lead has a context budget too. At a milestone, or before a long break, write the handoff and continue in a fresh session from the North Star.", "when": ["pass-end"], "checks": [{"kind": "metric", "name": "main_peak_context", "op": "<=", "config": "lead_context_budget_tokens", "on_fail": "ask"}, {"kind": "metric", "name": "rewrite_cost_share", "op": "<=", "config": "rewrite_share_ask_above", "on_fail": "ask"}], "question": "The lead's context or its rewrite cost is past the line. Is the North Star current enough to start a fresh lead session now? If not, what is missing from it?", "level": "scripted", "source": "field, measured (all four projects)", "first_seen": "2026-11-12"}
{"id": "M21", "title": "Every check passed while the product was plainly broken", "what": "A camera stuck in the wrong place, character features drawn out of position, and sounds that passed analysis but sounded wrong to the owner all passed their checks. The owner caught several from a small preview image.", "cause": "Checks were treated as deciding quality. Nobody looked at ordinary use.", "rule": "A gate decides whether it is broken. For anything a person sees, hears or feels, the lead also looks at ordinary use each pass, and the owner decides whether it is good.", "when": ["pass-end"], "checks": [{"kind": "manual"}], "question": "Who looked at or listened to ordinary use of the product this pass, through the real input path, and what did they see? Name the capture.", "level": "prose", "source": "field, observed (Projects C and D)", "first_seen": "2026-11-12"}
{"id": "M22", "title": "The lead reported something that did not exist or was not checked", "what": "A lead claimed evidence that did not exist, put invented filler content into a published image, and said it could not verify a fact when it had the means to. The owner caught all three.", "cause": "The lead wrote from memory or expectation, and nobody independent read the report.", "rule": "Every factual claim in an owner report is read from a file, a number or an output at the time of writing. Anything from memory is marked as such.", "when": ["owner-report"], "checks": [{"kind": "manual"}], "question": "For each factual claim in this report: which file, number or command output was it read from just now? List any that came from memory and mark them in the report.", "level": "prose", "source": "field, observed (Project B)", "first_seen": "2026-11-12"}
{"id": "M23", "title": "Lanes collided over something only one of them could use", "what": "Conflicts were not over files: one emulator instance, one storage card, one test device, the owner's screen recording, a browser another agent was using. Lanes editing regions of one large file broke the build twice, and half-done test renders appeared in the owner's window and drew about six notes.", "cause": "File ownership was the only isolation rule.", "rule": "List every exclusive resource in the environment contract with how agents take turns. The owner's screen is one of them: unfinished renders carry a visible TEST mark and evidence is captured from a frozen copy.", "when": ["pass-end"], "checks": [{"kind": "northstar", "name": "exclusive_resources_line"}], "level": "scripted", "source": "field, observed (Projects B and D)", "first_seen": "2026-11-12"}
{"id": "M24", "title": "An owner note was guessed at, or lost in the stream", "what": "Leads guessed at one-line instructions. One project took over a hundred owner messages as a continuous stream, with new requests landing on lanes already in flight.", "cause": "Notes lived in the conversation. There was no rule for when to guess and when to ask.", "rule": "Write each owner note down verbatim with an id, route it to a lane, and confirm back. Guess only when a wrong guess is cheap to undo, and flag the guess; otherwise ask.", "when": ["pass-end"], "checks": [{"kind": "northstar", "name": "owner_notes_routed"}], "level": "scripted", "source": "field, observed (Projects C and D)", "first_seen": "2026-11-12"}
{"id": "M25", "title": "An experiment ran on with no rule for abandoning it", "what": "An experimental mode took several crash cycles on real hardware before it was dropped. Each cycle cost the owner a manual test on the device.", "cause": "The pass count was capped but the task was not.", "rule": "An experimental task gets a kill rule before it starts, for example: two runs that fail the same way, abandon it and record why.", "when": ["brief"], "checks": [{"kind": "brief", "name": "brief_has_kill"}], "level": "template", "source": "field, observed (Project B)", "first_seen": "2026-11-12"}
{"id": "M26", "title": "More lanes ran than the lead could verify", "what": "With eight lanes running the lead did not rerun every gate or read every change. A thirty-second look at ordinary use would have caught a broken feature before the owner did.", "cause": "Lanes were started by how much work there was, not by how much the lead could check.", "rule": "Start only as many lanes as the lead can verify in the pass. Every lane marked done has an accept check on record.", "when": ["pass-end"], "checks": [{"kind": "ledger", "name": "accepted_lanes_checked"}], "level": "scripted", "source": "field, observed (Project D)", "first_seen": "2026-11-12"}
{"id": "M27", "title": "The owner's time was spent as if it were free", "what": "On a hardware project every failed attempt that reached the device cost the owner a manual swap and a restart. In another project about six owner notes were about test artefacts, not the product.", "cause": "Only model cost was counted.", "rule": "The owner's time is the most expensive resource. Batch what they must test so one session covers many things, give them a written checklist, and cap how often they are asked.", "when": ["owner-report"], "checks": [{"kind": "manual"}], "question": "How many times was the owner asked to test, swap or look this pass, and could those have been one session? Did anything unfinished reach them unmarked?", "level": "prose", "source": "field, observed (Projects B and D)", "first_seen": "2026-11-12"}
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis. | Stars: 472k | Updated | Claude CodeCodex | Verified | Install |
| diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness. | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependencies | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is available | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |